- Replace AnyModelLoader with ModelLoaderRegistry
- Fix type check errors in multiple files
- Remove apparently unneeded `get_model_config_enum()` method from model manager
- Remove last vestiges of old model manager
- Updated tests and documentation
resolve conflict with seamless.py
This substantially reduces the time spent encoding PNGs. In workflows with many image outputs, this is a drastic improvement.
For a tiled upscaling workflow going from 512x512 to a scale factor of 4, this can provide over 15% speed increase.
* add basic functionality for model metadata fetching from hf and civitai
* add storage
* start unit tests
* add unit tests and documentation
* add missing dependency for pytests
* remove redundant fetch; add modified/published dates; updated docs
* add code to select diffusers files based on the variant type
* implement Civitai installs
* make huggingface parallel downloading work
* add unit tests for model installation manager
- Fixed race condition on selection of download destination path
- Add fixtures common to several model_manager_2 unit tests
- Added dummy model files for testing diffusers and safetensors downloading/probing
- Refactored code for selecting proper variant from list of huggingface repo files
- Regrouped ordering of methods in model_install_default.py
* improve Civitai model downloading
- Provide a better error message when Civitai requires an access token (doesn't give a 403 forbidden, but redirects
to the HTML of an authorization page -- arrgh)
- Handle case of Civitai providing a primary download link plus additional links for VAEs, config files, etc
* add routes for retrieving metadata and tags
* code tidying and documentation
* fix ruff errors
* add file needed to maintain test root diretory in repo for unit tests
* fix self->cls in classmethod
* add pydantic plugin for mypy
* use TestSession instead of requests.Session to prevent any internet activity
improve logging
fix error message formatting
fix logging again
fix forward vs reverse slash issue in Windows install tests
* Several fixes of problems detected during PR review:
- Implement cancel_model_install_job and get_model_install_job routes
to allow for better control of model download and install.
- Fix thread deadlock that occurred after cancelling an install.
- Remove unneeded pytest_plugins section from tests/conftest.py
- Remove unused _in_terminal_state() from model_install_default.
- Remove outdated documentation from several spots.
- Add workaround for Civitai API results which don't return correct
URL for the default model.
* fix docs and tests to match get_job_by_source() rather than get_job()
* Update invokeai/backend/model_manager/metadata/fetch/huggingface.py
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
* Call CivitaiMetadata.model_validate_json() directly
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
* Second round of revisions suggested by @ryanjdick:
- Fix type mismatch in `list_all_metadata()` route.
- Do not have a default value for the model install job id
- Remove static class variable declarations from non Pydantic classes
- Change `id` field to `model_id` for the sqlite3 `model_tags` table.
- Changed AFTER DELETE triggers to ON DELETE CASCADE for the metadata and tags tables.
- Made the `id` field of the `model_metadata` table into a primary key to achieve uniqueness.
* Code cleanup suggested in PR review:
- Narrowed the declaration of the `parts` attribute of the download progress event
- Removed auto-conversion of str to Url in Url-containing sources
- Fixed handling of `InvalidModelConfigException`
- Made unknown sources raise `NotImplementedError` rather than `Exception`
- Improved status reporting on cached HuggingFace access tokens
* Multiple fixes:
- `job.total_size` returns a valid size for locally installed models
- new route `list_models` returns a paged summary of model, name,
description, tags and other essential info
- fix a few type errors
* consolidated all invokeai root pytest fixtures into a single location
* Update invokeai/backend/model_manager/metadata/metadata_store.py
Co-authored-by: psychedelicious <4822129+psychedelicious@users.noreply.github.com>
* Small tweaks in response to review comments:
- Remove flake8 configuration from pyproject.toml
- Use `id` rather than `modelId` for huggingface `ModelInfo` object
- Use `last_modified` rather than `LastModified` for huggingface `ModelInfo` object
- Add `sha256` field to file metadata downloaded from huggingface
- Add `Invoker` argument to the model installer `start()` and `stop()` routines
(but made it optional in order to facilitate use of the service outside the API)
- Removed redundant `PRAGMA foreign_keys` from metadata store initialization code.
* Additional tweaks and minor bug fixes
- Fix calculation of aggregate diffusers model size to only count the
size of files, not files + directories (which gives different unit test
results on different filesystems).
- Refactor _get_metadata() and _get_download_urls() to have distinct code paths
for Civitai, HuggingFace and URL sources.
- Forward the `inplace` flag from the source to the job and added unit test for this.
- Attach cached model metadata to the job rather than to the model install service.
* fix unit test that was breaking on windows due to CR/LF changing size of test json files
* fix ruff formatting
* a few last minor fixes before merging:
- Turn job `error` and `error_type` into properties derived from the exception.
- Add TODO comment about the reason for handling temporary directory destruction
manually rather than using tempfile.tmpdir().
* add unit tests for reporting HTTP download errors
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
Co-authored-by: psychedelicious <4822129+psychedelicious@users.noreply.github.com>
* feat: allow bfloat16 to be configurable in invoke.yaml
* fix: `torch_dtype()` util
- Use `choose_precision` to get the precision string
- Do not reference deprecated `config.full_precision` flat (why does this still exist?), if a user had this enabled it would override their actual precision setting and potentially cause a lot of confusion.
---------
Co-authored-by: psychedelicious <4822129+psychedelicious@users.noreply.github.com>
Custom nodes may be places in `$INVOKEAI_ROOT/nodes/` (configurable with `custom_nodes_dir` option).
On app startup, an `__init__.py` is copied into the custom nodes dir, which recursively loads all python files in the directory as modules (files starting with `_` are ignored). The custom nodes dir is now a python module itself.
When we `from invocations import *` to load init all invocations, we load the custom nodes dir, registering all custom nodes.
Upgrade pydantic and fastapi to latest.
- pydantic~=2.4.2
- fastapi~=103.2
- fastapi-events~=0.9.1
**Big Changes**
There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes.
**Invocations**
The biggest change relates to invocation creation, instantiation and validation.
Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie.
Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`.
With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation.
This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method.
In the end, this implementation is cleaner.
**Invocation Fields**
In pydantic v2, you can no longer directly add or remove fields from a model.
Previously, we did this to add the `type` field to invocations.
**Invocation Decorators**
With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper.
A similar technique is used for `invocation_output()`.
**Minor Changes**
There are a number of minor changes around the pydantic v2 models API.
**Protected `model_` Namespace**
All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_".
Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple.
```py
class IPAdapterModelField(BaseModel):
model_name: str = Field(description="Name of the IP-Adapter model")
base_model: BaseModelType = Field(description="Base model")
model_config = ConfigDict(protected_namespaces=())
```
**Model Serialization**
Pydantic models no longer have `Model.dict()` or `Model.json()`.
Instead, we use `Model.model_dump()` or `Model.model_dump_json()`.
**Model Deserialization**
Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions.
Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model.
```py
adapter_graph = TypeAdapter(Graph)
deserialized_graph_from_json = adapter_graph.validate_json(graph_json)
deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict)
```
**Field Customisation**
Pydantic `Field`s no longer accept arbitrary args.
Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field.
**Schema Customisation**
FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec.
This necessitates two changes:
- Our schema customization logic has been revised
- Schema parsing to build node templates has been revised
The specific aren't important, but this does present additional surface area for bugs.
**Performance Improvements**
Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node.
I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
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
The compress_level setting of PIL.Image.save(), used for PNG encoding. All settings are lossless. 0 = fastest, largest filesize, 9 = slowest, smallest filesize
Closes#4786
* 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>
Allow denying and explicitly allowing nodes. When a not-allowed node is used, a pydantic `ValidationError` will be raised.
- When collecting all invocations, check against the allowlist and denylist first. When pydantic constructs any unions related to nodes, the denied nodes will be omitted
- Add `allow_nodes` and `deny_nodes` to `InvokeAIAppConfig`. These are `Union[list[str], None]`, and may be populated with the `type` of invocations.
- When `allow_nodes` is `None`, allow all nodes, else if it is `list[str]`, only allow nodes in the list
- When `deny_nodes` is `None`, deny no nodes, else if it is `list[str]`, deny nodes in the list
- `deny_nodes` overrides `allow_nodes`
