GitLab

* FIX set_lora_device when target layers differ

Resolves #11833

Fixes a bug that occurs after calling set_lora_device when multiple LoRA
adapters are loaded that target different layers.

Note: Technically, the accompanying test does not require a GPU because
the bug is triggered even if the parameters are already on the
corresponding device, i.e. loading on CPU and then changing the device
to CPU is sufficient to cause the bug. However, this may be optimized
away in the future, so I decided to test with GPU.

* Update docstring to warn about device mismatch

* Extend docstring with an example

* Fix docstring

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* initial commit

* initial commit

* initial commit

* fix import

* fix prefix

* remove print

* Apply style fixes

---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

* update: FluxKontextInpaintPipeline support

* fix: Refactor code, remove mask_image_latents and ruff check

* feat: Add test case and fix with pytest

* Apply style fixes

* copies

---------
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

* add `WandVACETransformer3DModel` in`SINGLE_FILE_LOADABLE_CLASSES`

* add rename keys for `VACE`

add rename keys for `VACE`

* fix typo

Sincere thanks to @nitinmukesh 🙇‍♂️

* support for `1.3B VACE` model

Sincere thanks to @nitinmukesh again🙇

‍♂️

* update

* update

* Apply style fixes

---------
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

* use real instead of complex tensors in Wan2.1 RoPE

* remove the redundant type conversion

* unpack rotary_emb

* register rotary embedding frequencies as non-persistent buffers

* Apply style fixes

---------
Co-authored-by: Aryan <aryan@huggingface.co>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

* update

* update

* update docs

remove print

* ENH Improve speed of expanding LoRA scales

Resolves #11816

The following call proved to be a bottleneck when setting a lot of LoRA
adapters in diffusers:

https://github.com/huggingface/diffusers/blob/cdaf84a708eadf17d731657f4be3fa39d09a12c0/src/diffusers/loaders/peft.py#L482

This is because we would repeatedly call unet.state_dict(), even though
in the standard case, it is not necessary:

https://github.com/huggingface/diffusers/blob/cdaf84a708eadf17d731657f4be3fa39d09a12c0/src/diffusers/loaders/unet_loader_utils.py#L55



This PR fixes this by deferring this call, so that it is only run when
it's necessary, not earlier.

* Small fix

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* feat: use exclude modules to loraconfig.

* version-guard.

* tests and version guard.

* remove print.

* describe the test

* more detailed warning message + shift to debug

* update

* update

* update

* remove test

* fix: lora unloading behvaiour

* fix

* update

* update

* add test

* address review comments

* update

* fixes

* change decorator order to fix tests

* try fix

* fight tests

fix

* support flux kontext

* make fix-copies

* add example

* add tests

* update docs

* update

* add note on integrity checker

* initial commit

* initial commit

* add readme section and fixes in the training script.

* add test

* rectify ckpt_id

* fix ckpt

* fixes

* change id

* update

* Update examples/dreambooth/train_dreambooth_lora_flux_kontext.py
Co-authored-by: Aryan <aryan@huggingface.co>

* Update examples/dreambooth/README_flux.md

---------
Co-authored-by: Aryan <aryan@huggingface.co>
Co-authored-by: linoytsaban <linoy@huggingface.co>
Co-authored-by: Linoy Tsaban <57615435+linoytsaban@users.noreply.github.com>

* support flux kontext

* make fix-copies

* add example

* add tests

* update docs

* update

* add note on integrity checker

* make fix-copies issue

* add copied froms

* make style

* update repository ids

* more copied froms

* [rfc][compile] compile method for DiffusionPipeline

* Apply suggestions from code review
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Apply style fixes

* Update docs/source/en/optimization/fp16.md

* check

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

* update

* update

* update

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* post release v0.34.0

* code quality

---------
Co-authored-by: YiYi Xu <yixu310@gmail.com>

* update

* update

* update

* raise as early as possible in group offloading

* remove check from ModuleGroup

up

* minor cleanups in the lora docs.

* Apply suggestions from code review
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* format docs

* fix copies

---------
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

enable cpu offloading of new pipelines on XPU & use device agnostic empty to make pipelines work on XPU (#11671)

* commit 1
Signed-off-by: YAO Matrix <matrix.yao@intel.com>

* patch 2
Signed-off-by: YAO Matrix <matrix.yao@intel.com>

* Update pipeline_pag_sana.py

* Update pipeline_sana.py

* Update pipeline_sana_controlnet.py

* Update pipeline_sana_sprint_img2img.py

* Update pipeline_sana_sprint.py

* fix style
Signed-off-by: YAO Matrix <matrix.yao@intel.com>

* fix fat-thumb while merge conflict
Signed-off-by: YAO Matrix <matrix.yao@intel.com>

* fix ci issues
Signed-off-by: YAO Matrix <matrix.yao@intel.com>

---------
Signed-off-by: YAO Matrix <matrix.yao@intel.com>
Co-authored-by: Ilyas Moutawwakil <57442720+IlyasMoutawwakil@users.noreply.github.com>

Fix dimensionality in `apply_rotary_emb` functions' comments.

update

* start

* updates

* update

* update

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* start implementing disk offloading in group.

* delete diff file.

* updates.patch

* offload_to_disk_path

* check if safetensors already exist.

* add test and clarify.

* updates

* update todos.

* update more docs.

* update docs

* factor out stuff from load_lora_adapter().

* simplifying text encoder lora loading.

* fix peft.py

* fix logging locations.

* formatting

* fix

* update

* update

* update

update

update

add is_compileable property to quantizers.

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* updte

* update

* update

* update

change to 2025 licensing for remaining

* ⚡

️ Speed up method `AutoencoderKLWan.clear_cache` by 886%

**Key optimizations:**
- Compute the number of `WanCausalConv3d` modules in each model (`encoder`/`decoder`) **only once during initialization**, store in `self._cached_conv_counts`. This removes unnecessary repeated tree traversals at every `clear_cache` call, which was the main bottleneck (from profiling).
- The internal helper `_count_conv3d_fast` is optimized via a generator expression with `sum` for efficiency.

All comments from the original code are preserved, except for updated or removed local docstrings/comments relevant to changed lines.  
**Function signatures and outputs remain unchanged.**

* Apply style fixes

* Apply suggestions from code review
Co-authored-by: Aryan <contact.aryanvs@gmail.com>

* Apply style fixes

---------
Co-authored-by: codeflash-ai[bot] <148906541+codeflash-ai[bot]@users.noreply.github.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Aryan <aryan@huggingface.co>
Co-authored-by: Aryan <contact.aryanvs@gmail.com>
Co-authored-by: Aseem Saxena <aseem.bits@gmail.com>

update

* show how metadata stuff should be incorporated in training scripts.

* typing

* fix

---------
Co-authored-by: Linoy Tsaban <57615435+linoytsaban@users.noreply.github.com>