Consolidate graph processing logic into session processor.
With graphs as the unit of work, and the session queue distributing graphs, we no longer need the invocation queue or processor.
Instead, the session processor dequeues the next session and processes it in a simple loop, greatly simplifying the app.
- Remove `graph_execution_manager` service.
- Remove `queue` (invocation queue) service.
- Remove `processor` (invocation processor) service.
- Remove queue-related logic from `Invoker`. It now only starts and stops the services, providing them with access to other services.
- Remove unused `invocation_retrieval_error` and `session_retrieval_error` events, these are no longer needed.
- Clean up stats service now that it is less coupled to the rest of the app.
- Refactor cancellation logic - cancellations now originate from session queue (i.e. HTTP cancel endpoint) and are emitted as events. Processor gets the events and sets the canceled event. Access to this event is provided to the invocation context for e.g. the step callback.
- Remove `sessions` router; it provided access to `graph_executions` but that no longer exists.
`GraphInvocation` is a node that can contain a whole graph. It is removed for a number of reasons:
1. This feature was unused (the UI doesn't support it) and there is no plan for it to be used.
The use-case it served is known in other node execution engines as "node groups" or "blocks" - a self-contained group of nodes, which has group inputs and outputs. This is a planned feature that will be handled client-side.
2. It adds substantial complexity to the graph processing logic. It's probably not enough to have a measurable performance impact but it does make it harder to work in the graph logic.
3. It allows for graphs to be recursive, and the improved invocations union handling does not play well with it. Actually, it works fine within `graph.py` but not in the tests for some reason. I do not understand why. There's probably a workaround, but I took this as encouragement to remove `GraphInvocation` from the app since we don't use it.
The change to `Graph.nodes` and `GraphExecutionState.results` validation requires some fanagling to get the OpenAPI schema generation to work. See new comments for a details.
We use pydantic to validate a union of valid invocations when instantiating a graph.
Previously, we constructed the union while creating the `Graph` class. This introduces a dependency on the order of imports.
For example, consider a setup where we have 3 invocations in the app:
- Python executes the module where `FirstInvocation` is defined, registering `FirstInvocation`.
- Python executes the module where `SecondInvocation` is defined, registering `SecondInvocation`.
- Python executes the module where `Graph` is defined. A union of invocations is created and used to define the `Graph.nodes` field. The union contains `FirstInvocation` and `SecondInvocation`.
- Python executes the module where `ThirdInvocation` is defined, registering `ThirdInvocation`.
- A graph is created that includes `ThirdInvocation`. Pydantic validates the graph using the union, which does not know about `ThirdInvocation`, raising a `ValidationError` about an unknown invocation type.
This scenario has been particularly problematic in tests, where we may create invocations dynamically. The test files have to be structured in such a way that the imports happen in the right order. It's a major pain.
This PR refactors the validation of graph nodes to resolve this issue:
- `BaseInvocation` gets a new method `get_typeadapter`. This builds a pydantic `TypeAdapter` for the union of all registered invocations, caching it after the first call.
- `Graph.nodes`'s type is widened to `dict[str, BaseInvocation]`. This actually is a nice bonus, because we get better type hints whenever we reference `some_graph.nodes`.
- A "plain" field validator takes over the validation logic for `Graph.nodes`. "Plain" validators totally override pydantic's own validation logic. The validator grabs the `TypeAdapter` from `BaseInvocation`, then validates each node with it. The validation is identical to the previous implementation - we get the same errors.
`BaseInvocationOutput` gets the same treatment.
- 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
- Rename old "model_management" directory to "model_management_OLD" in order to catch
dangling references to original model manager.
- Caught and fixed most dangling references (still checking)
- Rename lora, textual_inversion and model_patcher modules
- Introduce a RawModel base class to simplfy the Union returned by the
model loaders.
- Tidy up the model manager 2-related tests. Add useful fixtures, and
a finalizer to the queue and installer fixtures that will stop the
services and release threads.
- ModelMetadataStoreService is now injected into ModelRecordStoreService
(these two services are really joined at the hip, and should someday be merged)
- ModelRecordStoreService is now injected into ModelManagerService
- Reduced timeout value for the various installer and download wait*() methods
- Introduced a Mock modelmanager for testing
- Removed bare print() statement with _logger in the install helper backend.
- Removed unused code from model loader init file
- Made `locker` a private variable in the `LoadedModel` object.
- Fixed up model merge frontend (will be deprecated anyway!)
- Replace legacy model manager service with the v2 manager.
- Update invocations to use new load interface.
- Fixed many but not all type checking errors in the invocations. Most
were unrelated to model manager
- Updated routes. All the new routes live under the route tag
`model_manager_v2`. To avoid confusion with the old routes,
they have the URL prefix `/api/v2/models`. The old routes
have been de-registered.
- Added a pytest for the loader.
- Updated documentation in contributing/MODEL_MANAGER.md
- Implement new model loader and modify invocations and embeddings
- Finish implementation loaders for all models currently supported by
InvokeAI.
- Move lora, textual_inversion, and model patching support into
backend/embeddings.
- Restore support for model cache statistics collection (a little ugly,
needs work).
- Fixed up invocations that load and patch models.
- Move seamless and silencewarnings utils into better location
- Cache stat collection enabled.
- Implemented ONNX loading.
- Add ability to specify the repo version variant in installer CLI.
- If caller asks for a repo version that doesn't exist, will fall back
to empty version rather than raising an error.
We have two different classes named `ModelInfo` which might need to be used by API consumers. We need to export both but have to deal with this naming collision.
The `ModelInfo` I've renamed here is the one that is returned when a model is loaded. It's the object least likely to be used by API consumers.
Replace `delete_on_startup: bool` & associated logic with `ephemeral: bool` and `TemporaryDirectory`.
The temp dir is created inside of `output_dir`. For example, if `output_dir` is `invokeai/outputs/tensors/`, then the temp dir might be `invokeai/outputs/tensors/tmpvj35ht7b/`.
The temp dir is cleaned up when the service is stopped, or when it is GC'd if not properly stopped.
In the event of a catastrophic crash where the temp files are not cleaned up, the user can delete the tempdir themselves.
This situation may not occur in normal use, but if you kill the process, python cannot clean up the temp dir itself. This includes running the app in a debugger and killing the debugger process - something I do relatively often.
Tests updated.
- The default is to not delete on startup - feels safer.
- The two services using this class _do_ delete on startup.
- The class has "ephemeral" removed from its name.
- Tests & app updated for this change.
`_delete_all` logged how many items it deleted, and had to be called _after_ service start bc it needed access to logger.
Move the logger call to the startup method and return the the deleted stats from `_delete_all`. This lets `_delete_all` be called at any time.
Turns out they are just different enough in purpose that the implementations would be rather unintuitive. I've made a separate ObjectSerializer service to handle tensors and conditioning.
Refined the class a bit too.
Turns out `ItemStorageABC` was almost identical to `PickleStorageBase`. Instead of maintaining separate classes, we can use `ItemStorageABC` for both.
There's only one change needed - the `ItemStorageABC.set` method must return the newly stored item's ID. This allows us to let the service handle the responsibility of naming the item, but still create the requisite output objects during node execution.
The naming implementation is improved here. It extracts the name of the generic and appends a UUID to that string when saving items.
- New generic class `PickleStorageBase`, implements the same API as `LatentsStorageBase`, use for storing non-serializable data via pickling
- Implementation `PickleStorageTorch` uses `torch.save` and `torch.load`, same as `LatentsStorageDisk`
- Add `tensors: PickleStorageBase[torch.Tensor]` to `InvocationServices`
- Add `conditioning: PickleStorageBase[ConditioningFieldData]` to `InvocationServices`
- Remove `latents` service and all `LatentsStorage` classes
- Update `InvocationContext` and all usage of old `latents` service to use the new services/context wrapper methods
This class works the same way as `WithMetadata` - it simply adds a `board` field to the node. The context wrapper function is able to pull the board id from this. This allows image-outputting nodes to get a board field "for free", and have their outputs automatically saved to it.
This is a breaking change for node authors who may have a field called `board`, because it makes `board` a reserved field name. I'll look into how to avoid this - maybe by naming this invoke-managed field `_board` to avoid collisions?
Supporting changes:
- `WithBoard` is added to all image-outputting nodes, giving them the ability to save to board.
- Unused, duplicate `WithMetadata` and `WithWorkflow` classes are deleted from `baseinvocation.py`. The "real" versions are in `fields.py`.
- Remove `LinearUIOutputInvocation`. Now that all nodes that output images also have a `board` field by default, this node is no longer necessary. See comment here for context: https://github.com/invoke-ai/InvokeAI/pull/5491#discussion_r1480760629
- Without `LinearUIOutputInvocation`, the `ImagesInferface.update` method is no longer needed, and removed.
Note: This commit does not bump all node versions. I will ensure that is done correctly before merging the PR of which this commit is a part.
Note: A followup commit will implement the frontend changes to support this change.
- The config is already cached by the config class's `get_config()` method.
- The config mutates itself in its `root_path` property getter. Freezing the class makes any attempt to grab a path from the config error. Unfortunately this means we cannot easily freeze the class without fiddling with the inner workings of `InvokeAIAppConfig`, which is outside the scope here.
Update all invocations to use the new context. The changes are all fairly simple, but there are a lot of them.
Supporting minor changes:
- Patch bump for all nodes that use the context
- Update invocation processor to provide new context
- Minor change to `EventServiceBase` to accept a node's ID instead of the dict version of a node
- Minor change to `ModelManagerService` to support the new wrapped context
- Fanagling of imports to avoid circular dependencies
Methods `get_node` and `complete` were typed as returning a dynamically created unions `InvocationsUnion` and `InvocationOutputsUnion`, respectively.
Static type analysers cannot work with dynamic objects, so these methods end up as effectively un-annotated, returning `Unknown`.
They now return `BaseInvocation` and `BaseInvocationOutput`, respectively, which are the superclasses of all members of each union. This gives us the best type annotation that is possible.
Note: the return types of these methods are never introspected, so it doesn't really matter what they are at runtime.
The change to memory session storage brings a subtle behaviour change.
Previously, we serialized and deserialized everything (e.g. field state, invocation outputs, etc) constantly. The meant we were effectively working with deep-copied objects at all time. We could mutate objects freely without worrying about other references to the object.
With memory storage, objects are now passed around by reference, and we cannot handle them in the same way.
This is problematic for nodes that mutate their own inputs. There are two ways this causes a problem:
- An output is used as input for multiple nodes. If the first node mutates the output object while `invoke`ing, the next node will get the mutated object.
- The invocation cache stores live python objects. When a node mutates an output pulled from the cache, the next node that uses the cached object will get the mutated object.
The solution is to deep-copy a node's inputs as they are set, effectively reproducing the same behaviour as we had with the SQLite session storage. Nodes can safely mutate their inputs and those changes never leave the node's scope.
Closes #5665
The stats service was logging error messages when attempting to retrieve stats for a graph that it wasn't tracking. This was rather noisy.
Instead of logging these errors within the service, we now will just raise the error and let the consumer of the service decide whether or not to log. Our usage of the service at this time is to suppress errors - we don't want to log anything to the console.
Note: With the improvements in the previous two commits, we shouldn't get these errors moving forward, but I still think this change is correct.
When an invocation is canceled, we consider the graph canceled. Log its graph's stats before resetting its graph's stats. No reason to not log these stats.
We also should stop the profiler at this point, because this graph is finished. If we don't stop it manually, it will stop itself and write the profile to disk when it is next started, but the resultant profile will include more than just its target graph.
Now we get both stats and profiles for canceled graphs.
When an invocation errored, we clear the stats for the whole graph. Later on, we check the graph for errors and see the failed invocation, and we consider the graph failed. We then attempts to log the stats for the failed graph.
Except now the failed graph has no stats, and the stats raises an error.
The user sees, in the terminal:
- An invocation error
- A stats error (scary!)
- No stats for the failed graph (uninformative!)
What the user should see:
- An invocation error
- Graph stats
The fix is simple - don't reset the graph stats when an invocation has an error.
- `ItemStorageMemory.get` now throws an `ItemNotFoundError` when the requested `item_id` is not found.
- Update docstrings in ABC and tests.
The new memory item storage implementation implemented the `get` method incorrectly, by returning `None` if the item didn't exist.
The ABC typed `get` as returning `T`, while the SQLite implementation typed `get` as returning `Optional[T]`. The SQLite implementation was referenced when writing the memory implementation.
This mismatched typing is a violation of the Liskov substitution principle, because the signature of the implementation of `get` in the implementation is wider than the abstract class's definition. Using `pyright` in strict mode catches this.
In `invocation_stats_default`, this introduced an error. The `_prune_stats` method calls `get`, expecting the method to throw if the item is not found. If the graph is no longer stored in the bounded item storage, we will call `is_complete()` on `None`, causing the error.
Note: This error condition never arose the SQLite implementation because it parsed the item with pydantic before returning it, which would throw if the item was not found. It implicitly threw, while the memory implementation did not.
* Port the command-line tools to use model_manager2
1.Reimplement the following:
- invokeai-model-install
- invokeai-merge
- invokeai-ti
To avoid breaking the original modeal manager, the udpated tools
have been renamed invokeai-model-install2 and invokeai-merge2. The
textual inversion training script should continue to work with
existing installations. The "starter" models now live in
`invokeai/configs/INITIAL_MODELS2.yaml`.
When the full model manager 2 is in place and working, I'll rename
these files and commands.
2. Add the `merge` route to the web API. This will merge two or three models,
resulting a new one.
- Note that because the model installer selectively installs the `fp16` variant
of models (rather than both 16- and 32-bit versions as previous),
the diffusers merge script will choke on any huggingface diffuserse models
that were downloaded with the new installer. Previously-downloaded models
should continue to merge correctly. I have a PR
upstream https://github.com/huggingface/diffusers/pull/6670 to fix
this.
3. (more important!)
During implementation of the CLI tools, found and fixed a number of small
runtime bugs in the model_manager2 implementation:
- During model database migration, if a registered models file was
not found on disk, the migration would be aborted. Now the
offending model is skipped with a log warning.
- Caught and fixed a condition in which the installer would download the
entire diffusers repo when the user provided a single `.safetensors`
file URL.
- Caught and fixed a condition in which the installer would raise an
exception and stop the app when a request for an unknown model's metadata
was passed to Civitai. Now an error is logged and the installer continues.
- Replaced the LoWRA starter LoRA with FlatColor. The former has been removed
from Civitai.
* fix ruff issue
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
Initially I wanted to show how many sessions were being deleted. In hindsight, this is not great:
- It requires extra logic in the migrator, which should be as simple as possible.
- It may be alarming to see "Clearing 224591 old sessions".
The app still reports on freed space during the DB startup logic.
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.
