* init commit
* style
* take comments into account
* mrege with main and simplify
* nits
* final
* small fixes
* fix
* super small update!
* add another test
* up up
* update
* fixes
* sort them by default
* Use canonical get_size_with_aspect_ratio (with max_size) from transformers.image_transforms to fix#37939
* Fix import sorting/style
* Fix import order
* Refactor: use canonical get_size_with_aspect_ratio across image processors (except YOLOS)
This commit updates image processing utilities in multiple model processors to use the shared
transformers.image_transforms.get_size_with_aspect_ratio for consistent resizing logic and
aspect ratio handling.
YOLOS processors are intentionally left unchanged in this commit to preserve their current
behavior and avoid breaking model-specific padding/resizing assumptions. YOLOS will be updated
in a dedicated follow-up PR once compatibility is fully verified.
* ruff fixes
* Fix check_copies.py references for get_size_with_aspect_ratio to use canonical transformers.image_transforms version
---------
Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com>
* Fix flash_attention.py: wrong argument passing for attn_implementation
The name of the attn type argument for `_flash_attention_forward()` should be `implementation`, instead of `attn_implementation` which currently uses in the function call. This would result in wrong type specification.
* modify the kwargs inside _flash_attention_forward
* fix the doc
* fix typo
---------
Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>
* update all models
* fix copies
* skip aria tests
* update other models
* skip should be in test, not tester
* i think this is more descriptive as a name
* find and replace for new models
The main content of this PR is to fix a bug in the delete_adapter method
of the PeftAdapterMixin. Previously, it did not take into account
auxiliary modules from PEFT, e.g. those added by modules_to_save. This
PR fixes this oversight.
Note that the PR uses a new functionality from PEFT that exposes
integration functions like delete_adapter. Those will be contained in
the next PEFT release, 0.18.0 (yet unreleased). Therefore, the bug is
only fixed when users have a PEFT version fullfilling this requirement.
I ensured that with old PEFT versions, the integration still works the
same as previously. The newly added test for this is skipped if the PEFT
version is too low.
(Note: I tested locally with that the test will pass with PEFT 0.18.0)
While working on this, I also cleaned up the following:
- The active_adapter property has been deprecated for more than 2 years
(#26407). It is safe to remove it now.
- There were numerous small errors or outdated pieces of information in
the docstrings, which have been addressed.
When PEFT < 0.18.0 is used, although we cannot delete modules_to_save,
we can still detect them and warn about it.
* support aux loss in qwen3vlmoe
* update qwen3vl processor test!
* add integration tests for qwen3vl-30a3
* remove duplicated decorator
* code clean
* fix consistency
* do not inherit from nn.Linear for better quantization
* pass check
