* add support for probing and loading SDXL VAE checkpoint files
* broaden regexp probe for SDXL VAEs
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
- Fix type errors
- Enable SilenceWarnings to be used as both a context manager and a decorator
- Remove duplicate implementation
- Check the initial verbosity on __enter__() rather than __init__()
* allow model patcher to optimize away the unpatching step when feasible
* remove lazy_offloading functionality
* allow model patcher to optimize away the unpatching step when feasible
* remove lazy_offloading functionality
* do not save original weights if there is a CPU copy of state dict
* Update invokeai/backend/model_manager/load/load_base.py
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
* documentation fixes requested during penultimate review
* add non-blocking=True parameters to several torch.nn.Module.to() calls, for slight performance increases
* fix ruff errors
* prevent crash on non-cuda-enabled systems
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
Co-authored-by: Kent Keirsey <31807370+hipsterusername@users.noreply.github.com>
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
* allow model patcher to optimize away the unpatching step when feasible
* remove lazy_offloading functionality
* allow model patcher to optimize away the unpatching step when feasible
* remove lazy_offloading functionality
* do not save original weights if there is a CPU copy of state dict
* Update invokeai/backend/model_manager/load/load_base.py
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
* documentation fixes added during penultimate review
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
Co-authored-by: Kent Keirsey <31807370+hipsterusername@users.noreply.github.com>
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
* avoid copying model back from cuda to cpu
* handle models that don't have state dicts
* add assertions that models need a `device()` method
* do not rely on torch.nn.Module having the device() method
* apply all patches after model is on the execution device
* fix model patching in latents too
* log patched tokenizer
* closes#6375
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
There are only a couple SDXL inpainting models, and my tests indicate they are not as good as SD1.5 inpainting, but at least we support them now.
- Add the config file. This matches what is used in A1111. The only difference from the non-inpainting SDXL config is the number of in-channels.
- Update the legacy config maps to use this config file.
Pending:
- Move model install calls into model manager and create passthrus in invocation_context.
- Consider splitting load_model_from_url() into a call to get the path and a call to load the path.
* introduce new abstraction layer for GPU devices
* add unit test for device abstraction
* fix ruff
* convert TorchDeviceSelect into a stateless class
* move logic to select context-specific execution device into context API
* add mock hardware environments to pytest
* remove dangling mocker fixture
* fix unit test for running on non-CUDA systems
* remove unimplemented get_execution_device() call
* remove autocast precision
* Multiple changes:
1. Remove TorchDeviceSelect.get_execution_device(), as well as calls to
context.models.get_execution_device().
2. Rename TorchDeviceSelect to TorchDevice
3. Added back the legacy public API defined in `invocation_api`, including
choose_precision().
4. Added a config file migration script to accommodate removal of precision=autocast.
* add deprecation warnings to choose_torch_device() and choose_precision()
* fix test crash
* remove app_config argument from choose_torch_device() and choose_torch_dtype()
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
When using refiner with a mask (i.e. inpainting), we don't have noise provided as an input to the node.
This situation uniquely hits a code path that wasn't reviewed when gradient denoising was implemented.
That code path does two things wrong:
- It lerp'd the input latents. This was fixed in 5a1f4cb1ce.
- It added noise to the latents an extra time. This is fixed in this change.
We don't need to add noise in `latents_from_embeddings` because we do it just a lines later in `AddsMaskGuidance`.
- Remove the extraneous call to `add_noise`
- Make `seed` a required arg. We never call the function without seed anyways. If we refactor this in the future, it will be clearer that we need to look at how seed is handled.
- Move the call to create the noise to a deeper conditional, just before we call `AddsMaskGuidance`. The created noise tensor is now only used in that function, no need to create it every time.
Note: Whether or not having both noise and latents as inputs on the node is correct is a separate conversation. This change just fixes the issue with the current setup.
- Allow user-defined precision on MPS.
- Use more explicit logic to handle all possible cases.
- Add comments.
- Remove the app_config args (they were effectively unused, just get the config using the singleton getter util)
The previous algorithm errored if the image wasn't divisible by the tile size. I've reimplemented it from scratch to mitigate this issue.
The new algorithm is simpler. We create a pool of tiles, then use them to create an image composed completely of tiles. If there is any awkwardly sized space on the edge of the image, the tiles are cropped to fit.
Finally, paste the original image over the tile image.
I've added a jupyter notebook to do a smoke test of infilling methods, and 10 test images.
The other infill algorithms can be easily tested with the notebook on the same images, though I didn't set that up yet.
Tested and confirmed this gives results just as good as the earlier infill, though of course they aren't the same due to the change in the algorithm.
Add `dump_path` arg to the converter function & save the model to disk inside the conversion function. This is the same pattern as in the other conversion functions.
* pass model config to _load_model
* make conversion work again
* do not write diffusers to disk when convert_cache set to 0
* adding same model to cache twice is a no-op, not an assertion error
* fix issues identified by psychedelicious during pr review
* following conversion, avoid redundant read of cached submodels
* fix error introduced while merging
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
The seamless logic errors when a second GPU is selected. I don't understand why, but a workaround is to skip the model patching when there there are no seamless axes specified.
This is also just a good practice regardless - don't patch the model unless we need to. Probably a negligible perf impact.
Closes#6010
"Normal" models have 4 in-channels, while "Depth" models have 5 and "Inpaint" models have 9.
We need to explicitly tell diffusers the channel count when converting models.
Closes #6058
It's possible for a model's state dict to have integer keys, though we do not actually support such models.
As part of probing, we call `key.startswith(...)` on the state dict keys. This raises an `AttributeError` for integer keys.
This logic is in `invokeai/backend/model_manager/probe.py:get_model_type_from_checkpoint`
To fix this, we can cast the keys to strings first. The models w/ integer keys will still fail to be probed, but we'll get a `InvalidModelConfigException` instead of `AttributeError`.
Closes#6044
Add `extra="forbid"` to the default settings models.
Closes#6035.
Pydantic has some quirks related to unions. This affected how the union of default settings was evaluated. See https://github.com/pydantic/pydantic/issues/9095 for a detailed description of the behaviour that this change addresses.
- Enriched dependencies to not just be a string - allows reuse of a dependency as a starter model _and_ dependency of another model. For example, all the SDXL models have the fp16 VAE as a dependency, but you can also download it on its own.
- Looked at popular models on the major model sites to select the list. No SD2 models. All hosted on HF.
For SSDs, `blake3` is about 10x faster than `blake3_single` - 3 files/second vs 30 files/second.
For spinning HDDs, `blake3` is about 100x slower than `blake3_single` - 300 seconds/file vs 3 seconds/file.
For external drives, `blake3` is always worse, but the difference is highly variable. For external spinning drives, it's probably way worse than internal.
The least offensive algorithm is `blake3_single`, and it's still _much_ faster than any other algorithm.
* add probe for SDXL controlnet models
* Update invokeai/backend/model_management/model_probe.py
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
* Update invokeai/backend/model_manager/probe.py
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
---------
Co-authored-by: Lincoln Stein <lstein@gmail.com>
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
These all support controlnet processors.
- `pil_to_cv2`
- `cv2_to_pil`
- `pil_to_np`
- `np_to_pil`
- `normalize_image_channel_count` (a readable version of `HWC3` from the controlnet repo)
- `fit_image_to_resolution` (a readable version of `resize_image` from the controlnet repo)
- `non_maximum_suppression` (a readable version of `nms` from the controlnet repo)
- `safe_step` (a readable version of `safe_step` from the controlnet repo)
The models from INITIAL_MODELS.yaml have been recreated as a structured python object. This data is served on a new route. The model sources are compared against currently-installed models to determine if they are already installed or not.
- Move base of t2i and clip_vision config models to DiffusersBase, which contains
a field to record the model variant (e.g. "fp16")
- This restore the ability to load fp16 t2i and clip_vision models
- Also add defensive coding to load the vanilla model when the fp16 model
has been replaced (or more likely, user's preferences changed since installation)
Co-authored-by: Lincoln Stein <lstein@gmail.com>
When running the configurator, the `legacy_models_conf_path` was stripped when saving the config file. Then the migration logic didn't fire correctly, and the custom models.yaml paths weren't migrated into the db.
- Rework the logic to migrate this path by adding it to the config object as a normal field that is not excluded from serialization.
- Rearrange the models.yaml migration logic to remove the legacy path after migrating, then write the config file. This way, the legacy path doesn't stick around.
- Move the schema version into the config object.
- Back up the config file before attempting migration.
- Add tests to cover this edge case
- `write_file` requires an destination file path
- `read_config` -> `merge_from_file`, if no path is provided, reads from `self.init_file_path`
- update app, tests to use new methods
- fix configurator, was overwriting config file data unexpectedly
- No longer install core conversion models. Use the HuggingFace cache to load
them if and when needed.
- Call directly into the diffusers library to perform conversions with only shallow
wrappers around them to massage arguments, etc.
- At root configuration time, do not create all the possible model subdirectories,
but let them be created and populated at model install time.
- Remove checks for missing core conversion files, since they are no
longer installed.
- This adds additional logic to the safetensors->diffusers conversion script
to check for and install missing core conversion models at runtime.
- Fixes#5934
BLAKE3 has poor performance on spinning disks when parallelized. See https://github.com/BLAKE3-team/BLAKE3/issues/31
- Replace `skip_model_hash` setting with `hashing_algorithm`. Any algorithm we support is accepted.
- Add `random` algorithm: hashes a UUID with BLAKE3 to create a random "hash". Equivalent to the previous skip functionality.
- Add `blake3_single` algorithm: hashes on a single thread using BLAKE3, fixes the aforementioned performance issue
- Update model probe to accept the algorithm to hash with as an optional arg, defaulting to `blake3`
- Update all calls of the probe to use the app's configured hashing algorithm
- Update an external script that probes models
- Update tests
- Move ModelHash into its own module to avoid circuclar import issues
