* [WIP][LoRA] Implement hot-swapping of LoRA
This PR adds the possibility to hot-swap LoRA adapters. It is WIP.
Description
As of now, users can already load multiple LoRA adapters. They can
offload existing adapters or they can unload them (i.e. delete them).
However, they cannot "hotswap" adapters yet, i.e. substitute the weights
from one LoRA adapter with the weights of another, without the need to
create a separate LoRA adapter.
Generally, hot-swapping may not appear not super useful but when the
model is compiled, it is necessary to prevent recompilation. See #9279
for more context.
Caveats
To hot-swap a LoRA adapter for another, these two adapters should target
exactly the same layers and the "hyper-parameters" of the two adapters
should be identical. For instance, the LoRA alpha has to be the same:
Given that we keep the alpha from the first adapter, the LoRA scaling
would be incorrect for the second adapter otherwise.
Theoretically, we could override the scaling dict with the alpha values
derived from the second adapter's config, but changing the dict will
trigger a guard for recompilation, defeating the main purpose of the
feature.
I also found that compilation flags can have an impact on whether this
works or not. E.g. when passing "reduce-overhead", there will be errors
of the type:
> input name: arg861_1. data pointer changed from 139647332027392 to
139647331054592
I don't know enough about compilation to determine whether this is
problematic or not.
Current state
This is obviously WIP right now to collect feedback and discuss which
direction to take this. If this PR turns out to be useful, the
hot-swapping functions will be added to PEFT itself and can be imported
here (or there is a separate copy in diffusers to avoid the need for a
min PEFT version to use this feature).
Moreover, more tests need to be added to better cover this feature,
although we don't necessarily need tests for the hot-swapping
functionality itself, since those tests will be added to PEFT.
Furthermore, as of now, this is only implemented for the unet. Other
pipeline components have yet to implement this feature.
Finally, it should be properly documented.
I would like to collect feedback on the current state of the PR before
putting more time into finalizing it.
* Reviewer feedback
* Reviewer feedback, adjust test
* Fix, doc
* Make fix
* Fix for possible g++ error
* Add test for recompilation w/o hotswapping
* Make hotswap work
Requires https://github.com/huggingface/peft/pull/2366
More changes to make hotswapping work. Together with the mentioned PEFT
PR, the tests pass for me locally.
List of changes:
- docstring for hotswap
- remove code copied from PEFT, import from PEFT now
- adjustments to PeftAdapterMixin.load_lora_adapter (unfortunately, some
state dict renaming was necessary, LMK if there is a better solution)
- adjustments to UNet2DConditionLoadersMixin._process_lora: LMK if this
is even necessary or not, I'm unsure what the overall relationship is
between this and PeftAdapterMixin.load_lora_adapter
- also in UNet2DConditionLoadersMixin._process_lora, I saw that there is
no LoRA unloading when loading the adapter fails, so I added it
there (in line with what happens in PeftAdapterMixin.load_lora_adapter)
- rewritten tests to avoid shelling out, make the test more precise by
making sure that the outputs align, parametrize it
- also checked the pipeline code mentioned in this comment:
https://github.com/huggingface/diffusers/pull/9453#issuecomment-2418508871;
when running this inside the with
torch._dynamo.config.patch(error_on_recompile=True) context, there is
no error, so I think hotswapping is now working with pipelines.
* Address reviewer feedback:
- Revert deprecated method
- Fix PEFT doc link to main
- Don't use private function
- Clarify magic numbers
- Add pipeline test
Moreover:
- Extend docstrings
- Extend existing test for outputs != 0
- Extend existing test for wrong adapter name
* Change order of test decorators
parameterized.expand seems to ignore skip decorators if added in last
place (i.e. innermost decorator).
* Split model and pipeline tests
Also increase test coverage by also targeting conv2d layers (support of
which was added recently on the PEFT PR).
* Reviewer feedback: Move decorator to test classes
... instead of having them on each test method.
* Apply suggestions from code review
Co-authored-by: hlky <hlky@hlky.ac>
* Reviewer feedback: version check, TODO comment
* Add enable_lora_hotswap method
* Reviewer feedback: check _lora_loadable_modules
* Revert changes in unet.py
* Add possibility to ignore enabled at wrong time
* Fix docstrings
* Log possible PEFT error, test
* Raise helpful error if hotswap not supported
I.e. for the text encoder
* Formatting
* More linter
* More ruff
* Doc-builder complaint
* Update docstring:
- mention no text encoder support yet
- make it clear that LoRA is meant
- mention that same adapter name should be passed
* Fix error in docstring
* Update more methods with hotswap argument
- SDXL
- SD3
- Flux
No changes were made to load_lora_into_transformer.
* Add hotswap argument to load_lora_into_transformer
For SD3 and Flux. Use shorter docstring for brevity.
* Extend docstrings
* Add version guards to tests
* Formatting
* Fix LoRA loading call to add prefix=None
See:
https://github.com/huggingface/diffusers/pull/10187#issuecomment-2717571064
* Run make fix-copies
* Add hot swap documentation to the docs
* Apply suggestions from code review
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: hlky <hlky@hlky.ac>
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
* minor documentation fixes of the depth and normals pipelines
* update license headers
* update model checkpoints in examples
fix missing prediction_type in register_to_config in the normals pipeline
* add initial marigold intrinsics pipeline
update comments about num_inference_steps and ensemble_size
minor fixes in comments of marigold normals and depth pipelines
* update uncertainty visualization to work with intrinsics
* integrate iid
---------
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
* Update hunyuan_video.md to rectify the checkpoint id
* bfloat16
* more fixes
* don't update the checkpoint ids.
* update
* t -> T
* Apply suggestions from code review
* fix
---------
Co-authored-by: YiYi Xu <yixu310@gmail.com>
* Add server example.
* Minor updates to README.
* Add fixes after local testing.
* Apply suggestions from code review
Updates to README from code review
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
* More doc updates.
* Maybe this will work to build the docs correctly?
* Fix style issues.
* Fix toc.
* Minor reformatting.
* Move docs to proper loc.
* Fix missing tick.
* Apply suggestions from code review
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
* Sync docs changes back to README.
* Very minor update to docs to add space.
---------
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
* [docs] Replace runwayml/stable-diffusion-v1-5 with Lykon/dreamshaper-8
Updated documentation as runwayml/stable-diffusion-v1-5 has been removed from Huggingface.
* Update docs/source/en/using-diffusers/inpaint.md
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
* Replace with stable-diffusion-v1-5/stable-diffusion-v1-5
* Update inpaint.md
---------
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
* Discourage using `revision`
* `make style && make quality`
* Refactor code to use 'variant' instead of 'revision'
* `revision="bf16"` -> `variant="bf16"`
---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* Trim all the trailing white space in the whole repo
* Remove unnecessary empty places
* make style && make quality
* Trim trailing white space
* trim
---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* implement marigold depth and normals pipelines in diffusers core
* remove bibtex
* remove deprecations
* remove save_memory argument
* remove validate_vae
* remove config output
* remove batch_size autodetection
* remove presets logic
move default denoising_steps and processing_resolution into the model config
make default ensemble_size 1
* remove no_grad
* add fp16 to the example usage
* implement is_matplotlib_available
use is_matplotlib_available, is_scipy_available for conditional imports in the marigold depth pipeline
* move colormap, visualize_depth, and visualize_normals into export_utils.py
* make the denoising loop more lucid
fix the outputs to always be 4d tensors or lists of pil images
support a 4d input_image case
attempt to support model_cpu_offload_seq
move check_inputs into a separate function
change default batch_size to 1, remove any logic to make it bigger implicitly
* style
* rename denoising_steps into num_inference_steps
* rename input_image into image
* rename input_latent into latents
* remove decode_image
change decode_prediction to use the AutoencoderKL.decode method
* move clean_latent outside of progress_bar
* refactor marigold-reusable image processing bits into MarigoldImageProcessor class
* clean up the usage example docstring
* make ensemble functions members of the pipelines
* add early checks in check_inputs
rename E into ensemble_size in depth ensembling
* fix vae_scale_factor computation
* better compatibility with torch.compile
better variable naming
* move export_depth_to_png to export_utils
* remove encode_prediction
* improve visualize_depth and visualize_normals to accept multi-dimensional data and lists
remove visualization functions from the pipelines
move exporting depth as 16-bit PNGs functionality from the depth pipeline
update example docstrings
* do not shortcut vae.config variables
* change all asserts to raise ValueError
* rename output_prediction_type to output_type
* better variable names
clean up variable deletion code
* better variable names
* pass desc and leave kwargs into the diffusers progress_bar
implement nested progress bar for images and steps loops
* implement scale_invariant and shift_invariant flags in the ensemble_depth function
add scale_invariant and shift_invariant flags readout from the model config
further refactor ensemble_depth
support ensembling without alignment
add ensemble_depth docstring
* fix generator device placement checks
* move encode_empty_text body into the pipeline call
* minor empty text encoding simplifications
* adjust pipelines' class docstrings to explain the added construction arguments
* improve the scipy failure condition
add comments
improve docstrings
change the default use_full_z_range to True
* make input image values range check configurable in the preprocessor
refactor load_image_canonical in preprocessor to reject unknown types and return the image in the expected 4D format of tensor and on right device
support a list of everything as inputs to the pipeline, change type to PipelineImageInput
implement a check that all input list elements have the same dimensions
improve docstrings of pipeline outputs
remove check_input pipeline argument
* remove forgotten print
* add prediction_type model config
* add uncertainty visualization into export utils
fix NaN values in normals uncertainties
* change default of output_uncertainty to False
better handle the case of an attempt to export or visualize none
* fix `output_uncertainty=False`
* remove kwargs
fix check_inputs according to the new inputs of the pipeline
* rename prepare_latent into prepare_latents as in other pipelines
annotate prepare_latents in normals pipeline with "Copied from"
annotate encode_image in normals pipeline with "Copied from"
* move nested-capable `progress_bar` method into the pipelines
revert the original `progress_bar` method in pipeline_utils
* minor message improvement
* fix cpu offloading
* move colormap, visualize_depth, export_depth_to_16bit_png, visualize_normals, visualize_uncertainty to marigold_image_processing.py
update example docstrings
* fix missing comma
* change torch.FloatTensor to torch.Tensor
* fix importing of MarigoldImageProcessor
* fix vae offloading
fix batched image encoding
remove separate encode_image function and use vae.encode instead
* implement marigold's intial tests
relax generator checks in line with other pipelines
implement return_dict __call__ argument in line with other pipelines
* fix num_images computation
* remove MarigoldImageProcessor and outputs from import structure
update tests
* update docstrings
* update init
* update
* style
* fix
* fix
* up
* up
* up
* add simple test
* up
* update expected np input/output to be channel last
* move expand_tensor_or_array into the MarigoldImageProcessor
* rewrite tests to follow conventions - hardcoded slices instead of image artifacts
write more smoke tests
* add basic docs.
* add anton's contribution statement
* remove todos.
* fix assertion values for marigold depth slow tests
* fix assertion values for depth normals.
* remove print
* support AutoencoderTiny in the pipelines
* update documentation page
add Available Pipelines section
add Available Checkpoints section
add warning about num_inference_steps
* fix missing import in docstring
fix wrong value in visualize_depth docstring
* [doc] add marigold to pipelines overview
* [doc] add section "usage examples"
* fix an issue with latents check in the pipelines
* add "Frame-by-frame Video Processing with Consistency" section
* grammarly
* replace tables with images with css-styled images (blindly)
* style
* print
* fix the assertions.
* take from the github runner.
* take the slices from action artifacts
* style.
* update with the slices from the runner.
* remove unnecessary code blocks.
* Revert "[doc] add marigold to pipelines overview"
This reverts commit a505165150afd8dab23c474d1a054ea505a56a5f.
* remove invitation for new modalities
* split out marigold usage examples
* doc cleanup
---------
Co-authored-by: yiyixuxu <yixu310@gmail.com>
Co-authored-by: yiyixuxu <yixu310@gmail,com>
Co-authored-by: sayakpaul <spsayakpaul@gmail.com>
