- 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
- 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
If the user specifies `torch-sdp` as the attention type in `config.yaml`, we can go ahead and use it (if available) rather than always throwing an exception.
Also, the PREVIOUS commit (@8d3885d, which was already pushed to github repo) was wrongly commented, but too late to fix without a force push or other mucking that I'm reluctant to do. That commit is actually the one that has all the changes to diffusers_pipeline.py to use additional arg down_intrablock_additional_residuals (introduced in diffusers PR https://github.com/huggingface/diffusers/pull/5362) to detangle T2I-Adapter from ControlNet inputs to main UNet.
Upgrade pydantic and fastapi to latest.
- pydantic~=2.4.2
- fastapi~=103.2
- fastapi-events~=0.9.1
**Big Changes**
There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes.
**Invocations**
The biggest change relates to invocation creation, instantiation and validation.
Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie.
Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`.
With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation.
This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method.
In the end, this implementation is cleaner.
**Invocation Fields**
In pydantic v2, you can no longer directly add or remove fields from a model.
Previously, we did this to add the `type` field to invocations.
**Invocation Decorators**
With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper.
A similar technique is used for `invocation_output()`.
**Minor Changes**
There are a number of minor changes around the pydantic v2 models API.
**Protected `model_` Namespace**
All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_".
Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple.
```py
class IPAdapterModelField(BaseModel):
model_name: str = Field(description="Name of the IP-Adapter model")
base_model: BaseModelType = Field(description="Base model")
model_config = ConfigDict(protected_namespaces=())
```
**Model Serialization**
Pydantic models no longer have `Model.dict()` or `Model.json()`.
Instead, we use `Model.model_dump()` or `Model.model_dump_json()`.
**Model Deserialization**
Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions.
Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model.
```py
adapter_graph = TypeAdapter(Graph)
deserialized_graph_from_json = adapter_graph.validate_json(graph_json)
deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict)
```
**Field Customisation**
Pydantic `Field`s no longer accept arbitrary args.
Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field.
**Schema Customisation**
FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec.
This necessitates two changes:
- Our schema customization logic has been revised
- Schema parsing to build node templates has been revised
The specific aren't important, but this does present additional surface area for bugs.
**Performance Improvements**
Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node.
I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
* Bump diffusers to 0.21.2.
* Add T2IAdapterInvocation boilerplate.
* Add T2I-Adapter model to model-management.
* (minor) Tidy prepare_control_image(...).
* Add logic to run the T2I-Adapter models at the start of the DenoiseLatentsInvocation.
* Add logic for applying T2I-Adapter weights and accumulating.
* Add T2IAdapter to MODEL_CLASSES map.
* yarn typegen
* Add model probes for T2I-Adapter models.
* Add all of the frontend boilerplate required to use T2I-Adapter in the nodes editor.
* Add T2IAdapterModel.convert_if_required(...).
* Fix errors in T2I-Adapter input image sizing logic.
* Fix bug with handling of multiple T2I-Adapters.
* black / flake8
* Fix typo
* yarn build
* Add num_channels param to prepare_control_image(...).
* Link to upstream diffusers bugfix PR that currently requires a workaround.
* feat: Add Color Map Preprocessor
Needed for the color T2I Adapter
* feat: Add Color Map Preprocessor to Linear UI
* Revert "feat: Add Color Map Preprocessor"
This reverts commit a1119a00bf.
* Revert "feat: Add Color Map Preprocessor to Linear UI"
This reverts commit bd8a9b82d8.
* Fix T2I-Adapter field rendering in workflow editor.
* yarn build, yarn typegen
---------
Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com>
Co-authored-by: psychedelicious <4822129+psychedelicious@users.noreply.github.com>
orphaned since #3550 removed the LazilyLoadedModelGroup code, probably unused since ModelCache took over responsibility for sequential offload somewhere around #3335.
This PR is to allow FP16 precision to work on Macs with MPS. In
addition, it centralizes the torch fixes/workarounds required for MPS
into a new backend utility `mps_fixes.py`. This is conditionally
imported in `api_app.py`/`cli_app.py`.
Many MANY thanks to @StAlKeR7779 for patiently working to debug and fix
these issues.
This PR is to allow FP16 precision to work on Macs with MPS. In addition, it centralizes the torch fixes/workarounds
required for MPS into a new backend utility file `mps_fixes.py`. This is conditionally imported in `api_app.py`/`cli_app.py`.
Many MANY thanks to StAlKeR7779 for patiently working to debug and fix these issues.
Only "real" conflicts were in:
invokeai/frontend/web/src/features/controlNet/components/ControlNet.tsx
invokeai/frontend/web/src/features/controlNet/store/controlNetSlice.ts
* Testing change to LatentsToText to allow setting different cfg_scale values per diffusion step.
* Adding first attempt at float param easing node, using Penner easing functions.
* Core implementation of ControlNet and MultiControlNet.
* Added support for ControlNet and MultiControlNet to legacy non-nodal Txt2Img in backend/generator. Although backend/generator will likely disappear by v3.x, right now they are very useful for testing core ControlNet and MultiControlNet functionality while node codebase is rapidly evolving.
* Added example of using ControlNet with legacy Txt2Img generator
* Resolving rebase conflict
* Added first controlnet preprocessor node for canny edge detection.
* Initial port of controlnet node support from generator-based TextToImageInvocation node to latent-based TextToLatentsInvocation node
* Switching to ControlField for output from controlnet nodes.
* Resolving conflicts in rebase to origin/main
* Refactored ControlNet nodes so they subclass from PreprocessedControlInvocation, and only need to override run_processor(image) (instead of reimplementing invoke())
* changes to base class for controlnet nodes
* Added HED, LineArt, and OpenPose ControlNet nodes
* Added an additional "raw_processed_image" output port to controlnets, mainly so could route ImageField to a ShowImage node
* Added more preprocessor nodes for:
MidasDepth
ZoeDepth
MLSD
NormalBae
Pidi
LineartAnime
ContentShuffle
Removed pil_output options, ControlNet preprocessors should always output as PIL. Removed diagnostics and other general cleanup.
* Prep for splitting pre-processor and controlnet nodes
* Refactored controlnet nodes: split out controlnet stuff into separate node, stripped controlnet stuff form image processing/analysis nodes.
* Added resizing of controlnet image based on noise latent. Fixes a tensor mismatch issue.
* More rebase repair.
* Added support for using multiple control nets. Unfortunately this breaks direct usage of Control node output port ==> TextToLatent control input port -- passing through a Collect node is now required. Working on fixing this...
* Fixed use of ControlNet control_weight parameter
* Fixed lint-ish formatting error
* Core implementation of ControlNet and MultiControlNet.
* Added first controlnet preprocessor node for canny edge detection.
* Initial port of controlnet node support from generator-based TextToImageInvocation node to latent-based TextToLatentsInvocation node
* Switching to ControlField for output from controlnet nodes.
* Refactored controlnet node to output ControlField that bundles control info.
* changes to base class for controlnet nodes
* Added more preprocessor nodes for:
MidasDepth
ZoeDepth
MLSD
NormalBae
Pidi
LineartAnime
ContentShuffle
Removed pil_output options, ControlNet preprocessors should always output as PIL. Removed diagnostics and other general cleanup.
* Prep for splitting pre-processor and controlnet nodes
* Refactored controlnet nodes: split out controlnet stuff into separate node, stripped controlnet stuff form image processing/analysis nodes.
* Added resizing of controlnet image based on noise latent. Fixes a tensor mismatch issue.
* Cleaning up TextToLatent arg testing
* Cleaning up mistakes after rebase.
* Removed last bits of dtype and and device hardwiring from controlnet section
* Refactored ControNet support to consolidate multiple parameters into data struct. Also redid how multiple controlnets are handled.
* Added support for specifying which step iteration to start using
each ControlNet, and which step to end using each controlnet (specified as fraction of total steps)
* Cleaning up prior to submitting ControlNet PR. Mostly turning off diagnostic printing. Also fixed error when there is no controlnet input.
* Added dependency on controlnet-aux v0.0.3
* Commented out ZoeDetector. Will re-instate once there's a controlnet-aux release that supports it.
* Switched CotrolNet node modelname input from free text to default list of popular ControlNet model names.
* Fix to work with current stable release of controlnet_aux (v0.0.3). Turned of pre-processor params that were added post v0.0.3. Also change defaults for shuffle.
* Refactored most of controlnet code into its own method to declutter TextToLatents.invoke(), and make upcoming integration with LatentsToLatents easier.
* Cleaning up after ControlNet refactor in TextToLatentsInvocation
* Extended node-based ControlNet support to LatentsToLatentsInvocation.
* chore(ui): regen api client
* fix(ui): add value to conditioning field
* fix(ui): add control field type
* fix(ui): fix node ui type hints
* fix(nodes): controlnet input accepts list or single controlnet
* Moved to controlnet_aux v0.0.4, reinstated Zoe controlnet preprocessor. Also in pyproject.toml had to specify downgrade of timm to 0.6.13 _after_ controlnet-aux installs timm >= 0.9.2, because timm >0.6.13 breaks Zoe preprocessor.
* Core implementation of ControlNet and MultiControlNet.
* Added first controlnet preprocessor node for canny edge detection.
* Switching to ControlField for output from controlnet nodes.
* Resolving conflicts in rebase to origin/main
* Refactored ControlNet nodes so they subclass from PreprocessedControlInvocation, and only need to override run_processor(image) (instead of reimplementing invoke())
* changes to base class for controlnet nodes
* Added HED, LineArt, and OpenPose ControlNet nodes
* Added more preprocessor nodes for:
MidasDepth
ZoeDepth
MLSD
NormalBae
Pidi
LineartAnime
ContentShuffle
Removed pil_output options, ControlNet preprocessors should always output as PIL. Removed diagnostics and other general cleanup.
* Prep for splitting pre-processor and controlnet nodes
* Refactored controlnet nodes: split out controlnet stuff into separate node, stripped controlnet stuff form image processing/analysis nodes.
* Added resizing of controlnet image based on noise latent. Fixes a tensor mismatch issue.
* Added support for using multiple control nets. Unfortunately this breaks direct usage of Control node output port ==> TextToLatent control input port -- passing through a Collect node is now required. Working on fixing this...
* Fixed use of ControlNet control_weight parameter
* Core implementation of ControlNet and MultiControlNet.
* Added first controlnet preprocessor node for canny edge detection.
* Initial port of controlnet node support from generator-based TextToImageInvocation node to latent-based TextToLatentsInvocation node
* Switching to ControlField for output from controlnet nodes.
* Refactored controlnet node to output ControlField that bundles control info.
* changes to base class for controlnet nodes
* Added more preprocessor nodes for:
MidasDepth
ZoeDepth
MLSD
NormalBae
Pidi
LineartAnime
ContentShuffle
Removed pil_output options, ControlNet preprocessors should always output as PIL. Removed diagnostics and other general cleanup.
* Prep for splitting pre-processor and controlnet nodes
* Refactored controlnet nodes: split out controlnet stuff into separate node, stripped controlnet stuff form image processing/analysis nodes.
* Added resizing of controlnet image based on noise latent. Fixes a tensor mismatch issue.
* Cleaning up TextToLatent arg testing
* Cleaning up mistakes after rebase.
* Removed last bits of dtype and and device hardwiring from controlnet section
* Refactored ControNet support to consolidate multiple parameters into data struct. Also redid how multiple controlnets are handled.
* Added support for specifying which step iteration to start using
each ControlNet, and which step to end using each controlnet (specified as fraction of total steps)
* Cleaning up prior to submitting ControlNet PR. Mostly turning off diagnostic printing. Also fixed error when there is no controlnet input.
* Commented out ZoeDetector. Will re-instate once there's a controlnet-aux release that supports it.
* Switched CotrolNet node modelname input from free text to default list of popular ControlNet model names.
* Fix to work with current stable release of controlnet_aux (v0.0.3). Turned of pre-processor params that were added post v0.0.3. Also change defaults for shuffle.
* Refactored most of controlnet code into its own method to declutter TextToLatents.invoke(), and make upcoming integration with LatentsToLatents easier.
* Cleaning up after ControlNet refactor in TextToLatentsInvocation
* Extended node-based ControlNet support to LatentsToLatentsInvocation.
* chore(ui): regen api client
* fix(ui): fix node ui type hints
* fix(nodes): controlnet input accepts list or single controlnet
* Added Mediapipe image processor for use as ControlNet preprocessor.
Also hacked in ability to specify HF subfolder when loading ControlNet models from string.
* Fixed bug where MediapipFaceProcessorInvocation was ignoring max_faces and min_confidence params.
* Added nodes for float params: ParamFloatInvocation and FloatCollectionOutput. Also added FloatOutput.
* Added mediapipe install requirement. Should be able to remove once controlnet_aux package adds mediapipe to its requirements.
* Added float to FIELD_TYPE_MAP ins constants.ts
* Progress toward improvement in fieldTemplateBuilder.ts getFieldType()
* Fixed controlnet preprocessors and controlnet handling in TextToLatents to work with revised Image services.
* Cleaning up from merge, re-adding cfg_scale to FIELD_TYPE_MAP
* Making sure cfg_scale of type list[float] can be used in image metadata, to support param easing for cfg_scale
* Fixed math for per-step param easing.
* Added option to show plot of param value at each step
* Just cleaning up after adding param easing plot option, removing vestigial code.
* Modified control_weight ControlNet param to be polistmorphic --
can now be either a single float weight applied for all steps, or a list of floats of size total_steps, that specifies weight for each step.
* Added more informative error message when _validat_edge() throws an error.
* Just improving parm easing bar chart title to include easing type.
* Added requirement for easing-functions package
* Taking out some diagnostic prints.
* Added option to use both easing function and mirror of easing function together.
* Fixed recently introduced problem (when pulled in main), triggered by num_steps in StepParamEasingInvocation not having a default value -- just added default.
---------
Co-authored-by: psychedelicious <4822129+psychedelicious@users.noreply.github.com>
Problem was that controlnet support involved adding **kwargs to method calls down in denoising loop, and AddsMaskLatents didn't accept **kwarg arg. So just changed to accept and pass on **kwargs.
- Make environment variable settings case InSenSiTive:
INVOKEAI_MAX_LOADED_MODELS and InvokeAI_Max_Loaded_Models
environment variables will both set `max_loaded_models`
- Updated realesrgan to use new config system.
- Updated textual_inversion_training to use new config system.
- Discovered a race condition when InvokeAIAppConfig is created
at module load time, which makes it impossible to customize
or replace the help message produced with --help on the command
line. To fix this, moved all instances of get_invokeai_config()
from module load time to object initialization time. Makes code
cleaner, too.
- Added `--from_file` argument to `invokeai-node-cli` and changed
github action to match. CI tests will hopefully work now.
