[Bugfix] Remove incorrect torchvision requirement from PIL backend image processors by Lidang-Jiang · Pull Request #45045 · huggingface/transformers

Lidang-Jiang · 2026-03-27T08:19:19Z

Isolate dependencies, make PIL independant from Torchvision backend

PR #45029 added @requires(backends=("vision", "torch", "torchvision")) to 67 PIL backend image_processing_pil_*.py files. This causes PIL backend classes to become dummy objects when torchvision is not installed, making AutoImageProcessor unable to find any working processor — even though the PIL backend's purpose is to work without torchvision.

Fixes:

duplicate Kwarg import (PIL imported Kwargs from Torchvision equivalent file)
fix modular to inline Kwargs if used by both files
explicit import when it makes sense
explicit protection with requires(backends=("torch")) when file actually needs it.

@requires

…age processors PR huggingface#45029 added @requires(backends=("vision", "torch", "torchvision")) to 67 PIL backend image_processing_pil_*.py files. This causes PIL backend classes to become dummy objects when torchvision is not installed, making AutoImageProcessor unable to find any working processor. Fix: set @requires to ("vision",) for files that only need PIL, and ("vision", "torch") for files that also use torch directly. Also fix 5 modular source files so make fix-repo preserves the correct backends. Fixes huggingface#45042 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/transformers/models/aria/image_processing_pil_aria.py

HuggingFaceDocBuilderDev · 2026-03-27T10:10:02Z

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

ArthurZucker · 2026-03-27T10:22:10Z

I'm having a look

@requires

…ckends Per reviewer feedback: the vision-only @requires decorator is redundant for PIL backend classes since PilBackend base class already handles this. - Remove @requires(backends=("vision",)) from 43 PIL backend files - Remove unused `requires` import from 38 files (Category A) - Keep @requires(backends=("vision", "torch")) on method-level decorators (Category B: 5 files)

Lidang-Jiang · 2026-03-27T10:49:07Z

CI Failure Analysis

The 3 test failures in tests_non_model are unrelated to this PR.

Failed tests

Test	File
`TestGetDecoder::test_vision_language_model`	`tests/utils/test_modeling_utils.py`
`ModelUtilsTest::test_use_safetensors`	`tests/utils/test_modeling_utils.py`
`ModelUtilsTest::test_model_from_pretrained_dtype`	`tests/utils/test_modeling_utils.py`

Root cause

All 3 failures share the same stack trace — a KeyError in _can_set_experts_implementation() → _grouped_mm_can_dispatch() during model __init__:

modeling_utils.py:1263 → _check_and_adjust_experts_implementation
modeling_utils.py:1885 → get_correct_experts_implementation
modeling_utils.py:1938 → _grouped_mm_can_dispatch
modeling_utils.py:1733 → _can_set_experts_implementation  ← crash

This is a known pre-existing issue on main (tracked in #45003, fix in #45043). This PR only modifies @requires decorators in PIL backend image_processing_pil_*.py files and does not touch modeling_utils.py.

ArthurZucker · 2026-03-27T11:25:41Z

Thanks you agent, let me take over!

src/transformers/models/beit/image_processing_pil_beit.py

zucchini-nlp · 2026-03-27T12:12:46Z

src/transformers/models/vitmatte/image_processing_pil_vitmatte.py

 if is_torch_available():
-    import torch
+    pass


and dangling imports 😄

apart from it, looks fine to me, checked a few models

yeah might be modular

src/transformers/models/superpoint/image_processing_pil_superpoint.py

LysandreJik

love it

Lidang-Jiang · 2026-03-27T13:23:12Z

Thanks for taking over and cleaning up the imports further! I appreciate it 🙏 Just to clarify — I'm a real human, not an agent 😄 I used Claude Code as an assistive tool, but the analysis and decisions were mine. Happy to help with anything else if needed!

ArthurZucker · 2026-03-27T14:16:49Z

Oh sorry haha you were so fast I thought it was automated

Lidang-Jiang · 2026-03-27T14:45:27Z

Oh sorry haha you were so fast I thought it was automated

The Codex is really fast! Codex always rejects Claude Code's commits. I'm planning to use Codex now. Then it would be fully automated and I wouldn't need to repeatedly prompt the AI. (In other PRs, I was often rejected by Codex several times and only got merged in the end.)

ArthurZucker

Other reviewers should not have to go through it all, just check my comments and tell me!

ArthurZucker · 2026-03-28T06:38:36Z

src/transformers/models/fuyu/image_processing_pil_fuyu.py

+# Adapted from transformers.models.fuyu.image_processing_fuyu.FuyuBatchFeature
+class FuyuBatchFeature(BatchFeature):


no longer needed remove from modular for both

ArthurZucker · 2026-03-28T06:41:33Z

src/transformers/models/layoutlmv3/image_processing_pil_layoutlmv3.py

cc reviewer: only place I have a doubt maybe add require pytesseract?

ArthurZucker · 2026-03-28T06:42:57Z

src/transformers/models/mask2former/image_processing_pil_mask2former.py

+@requires(backends=("torch",))
 class Mask2FormerImageProcessorPil(PilBackend):


cc @yonigozlan 2 or 3 of them require torch -> why even have a PIL when torch is needed anyway?

ArthurZucker · 2026-03-28T06:44:18Z

src/transformers/models/oneformer/image_processing_pil_oneformer.py

weird one will see if it can be cleaner?

ArthurZucker · 2026-03-28T06:44:59Z

src/transformers/models/owlv2/image_processing_pil_owlv2.py

+@requires(backends=("torch",))
 class Owlv2ImageProcessorPil(PilBackend):


same comment, does not make sense if the entire processor needs torch -> remove it?

ArthurZucker · 2026-03-28T06:45:46Z

src/transformers/models/pix2struct/image_processing_pil_pix2struct.py


 @auto_docstring
-@requires(backends=("vision", "torch", "torchvision"))
+@requires(backends=("torch",))


ArthurZucker · 2026-03-28T06:47:40Z

src/transformers/models/sam/image_processing_pil_sam.py

+
 @auto_docstring
-@requires(backends=("vision", "torch", "torchvision"))
+@requires(backends=("torch",))


same comment

ArthurZucker · 2026-03-28T06:49:25Z

src/transformers/models/vitmatte/image_processing_pil_vitmatte.py

 if is_torch_available():
-    import torch
+    pass


yeah might be modular

ArthurZucker · 2026-03-28T06:50:11Z

src/transformers/models/zoedepth/image_processing_pil_zoedepth.py

+
 @auto_docstring
-@requires(backends=("vision", "torch", "torchvision"))
+@requires(backends=("torch",))


same comment about PIL needed at all?

ArthurZucker · 2026-03-28T06:50:51Z

utils/modular_model_converter.py

+        # Exception: for image_processing_pil files, image_processing modular classes must be inlined (not excluded),
+        # because these two files must never import from each other.


important change @yonigozlan and @Cyrilvallez

Copilot

Pull request overview

This PR fixes an unintended dependency coupling where many PIL backend image processors were marked as requiring torchvision, causing them to become unavailable (dummy objects) in environments without torchvision, which in turn breaks AutoImageProcessor/AutoProcessor resolution.

Changes:

Removes/relaxes torchvision backend requirements on PIL image processor implementations, and adds narrower @requires(backends=("torch",)) guards only where PIL code truly needs PyTorch.
Inlines/shared-copies kwargs types and helper utilities into PIL modules to avoid cross-imports between image_processing_* and image_processing_pil_*.
Updates modular conversion + docstring generation logic to handle these dependency edges more safely.

Reviewed changes

Copilot reviewed 187 out of 187 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
utils/modular_model_converter.py	Prevents cross-imports between `image_processing` and `image_processing_pil` during modular conversion; prefers modular-defined nodes when inlining.
src/transformers/utils/auto_docstring.py	Makes pipeline-example lookup resilient to missing/optional auto-model attributes.
src/transformers/image_processing_backends.py	Narrows PIL backend resize `resample` typing (but see review comment re: mismatch with implementation).
src/transformers/models/zoedepth/image_processing_zoedepth.py	Adjusts torch/torchvision import/usage patterns for ZoeDepth torchvision backend.
src/transformers/models/zoedepth/image_processing_pil_zoedepth.py	Removes torchvision requirement; inlines kwargs + resize helpers; adds torch-only requirement where needed.
src/transformers/models/yolos/image_processing_pil_yolos.py	Removes torchvision requirement; inlines kwargs; narrows torch requirements to post-processing.
src/transformers/models/vitpose/image_processing_vitpose.py	Updates import strategy for torch/torchvision backend implementation.
src/transformers/models/vitmatte/image_processing_vitmatte.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/vitmatte/image_processing_pil_vitmatte.py	Removes torchvision requirement; inlines kwargs; simplifies torch typing.
src/transformers/models/vilt/image_processing_vilt.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/vilt/image_processing_pil_vilt.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample types.
src/transformers/models/videomae/image_processing_videomae.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/videomae/image_processing_pil_videomae.py	Narrows PIL resample types.
src/transformers/models/video_llama_3/video_processing_video_llama_3.py	Removes torchvision-conditional import usage from video processor surface.
src/transformers/models/video_llama_3/modular_video_llama_3.py	Stops importing kwargs from qwen2_vl; inlines kwargs to avoid PIL↔torchvision coupling.
src/transformers/models/video_llama_3/modeling_video_llama_3.py	Import formatting cleanup.
src/transformers/models/video_llama_3/image_processing_video_llama_3.py	Removes torchvision-availability guard; adjusts typing.
src/transformers/models/video_llama_3/image_processing_pil_video_llama_3.py	Removes torchvision requirement; inlines kwargs + smart_resize into PIL file.
src/transformers/models/tvp/image_processing_tvp.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/tvp/image_processing_pil_tvp.py	Removes torchvision requirement; inlines kwargs; adjusts optional torchvision/torch handling.
src/transformers/models/textnet/image_processing_textnet.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/textnet/image_processing_pil_textnet.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample types.
src/transformers/models/swin2sr/image_processing_swin2sr.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/swin2sr/image_processing_pil_swin2sr.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample types.
src/transformers/models/superpoint/image_processing_superpoint.py	Removes torchvision-availability guard and imports torchvision functional directly.
src/transformers/models/superpoint/image_processing_pil_superpoint.py	Removes torchvision requirement; inlines kwargs; narrows torch requirement to post-processing.
src/transformers/models/superglue/image_processing_superglue.py	Removes torchvision-availability guard and imports torchvision functional directly.
src/transformers/models/superglue/image_processing_pil_superglue.py	Removes torchvision requirement; inlines validation helper; narrows torch requirement to post-processing/visualization.
src/transformers/models/smolvlm/processing_smolvlm.py	Inlines copied constants to avoid `vision` gating for text-only processor import.
src/transformers/models/smolvlm/modular_smolvlm.py	Introduces `SmolVLMImageProcessorKwargs` in modular file.
src/transformers/models/smolvlm/modeling_smolvlm.py	Import formatting cleanup.
src/transformers/models/smolvlm/image_processing_smolvlm.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/siglip2/image_processing_siglip2.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/siglip2/image_processing_pil_siglip2.py	Removes torchvision requirement; inlines kwargs + helper sizing function.
src/transformers/models/seggpt/image_processing_seggpt.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/seggpt/image_processing_pil_seggpt.py	Removes torchvision requirement; inlines kwargs + palette builder; narrows torch requirement to post-processing.
src/transformers/models/segformer/modular_segformer.py	Adds kwargs type + narrows backend requirements for PIL class.
src/transformers/models/segformer/image_processing_segformer.py	Adjusts torch functional import usage.
src/transformers/models/segformer/image_processing_pil_segformer.py	Removes torchvision requirement; inlines kwargs; narrows requires to torch+torchvision for PIL variant.
src/transformers/models/sam3_video/processing_sam3_video.py	Guards `torch` import behind `is_torch_available()`.
src/transformers/models/sam2/image_processing_sam2.py	Import formatting cleanup.
src/transformers/models/sam/image_processing_sam.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/sam/image_processing_pil_sam.py	Removes torchvision requirement; inlines kwargs; narrows requires to torch where used.
src/transformers/models/rt_detr/modular_rt_detr.py	Removes torchvision functional import; inlines kwargs; narrows PIL requires to torch.
src/transformers/models/rt_detr/image_processing_rt_detr.py	Changes torchvision import style and resample typing.
src/transformers/models/rt_detr/image_processing_pil_rt_detr.py	Removes torchvision requirement; inlines kwargs; narrows requires to torch for post-processing.
src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/qwen2_vl/image_processing_pil_qwen2_vl.py	Removes torchvision requirement; inlines kwargs + smart_resize.
src/transformers/models/prompt_depth_anything/image_processing_prompt_depth_anything.py	Removes torchvision availability check; imports torch+tvF unconditionally.
src/transformers/models/prompt_depth_anything/image_processing_pil_prompt_depth_anything.py	Removes torchvision requirement; inlines kwargs; narrows torch requirement to post-processing.
src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/pp_doclayout_v3/image_processing_pp_doclayout_v3.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/pp_doclayout_v2/image_processing_pp_doclayout_v2.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/poolformer/image_processing_poolformer.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/poolformer/image_processing_pil_poolformer.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample types.
src/transformers/models/pixtral/image_processing_pixtral.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/pixtral/image_processing_pil_pixtral.py	Removes torchvision requirement; inlines kwargs + resize math helpers.
src/transformers/models/phi4_multimodal/image_processing_phi4_multimodal.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/perception_lm/image_processing_perception_lm.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/perceiver/image_processing_pil_perceiver.py	Narrows PIL resample typing.
src/transformers/models/perceiver/image_processing_perceiver.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/paddleocr_vl/modular_paddleocr_vl.py	Removes kwargs import from qwen2_vl; inlines kwargs to avoid coupling.
src/transformers/models/paddleocr_vl/modeling_paddleocr_vl.py	Import formatting cleanup.
src/transformers/models/paddleocr_vl/image_processing_pil_paddleocr_vl.py	Removes torchvision requirement; inlines kwargs + smart_resize.
src/transformers/models/paddleocr_vl/image_processing_paddleocr_vl.py	Removes torchvision availability check; adjusts typing.
src/transformers/models/owlv2/modular_owlv2.py	Removes torchvision requirement from PIL class; adjusts typing for PIL preprocess path.
src/transformers/models/owlv2/image_processing_owlv2.py	Narrows resample typing for torch backend.
src/transformers/models/ovis2/image_processing_ovis2.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/oneformer/image_processing_oneformer.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/nougat/image_processing_pil_nougat.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample typing.
src/transformers/models/nougat/image_processing_nougat.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/mobilevit/image_processing_pil_mobilevit.py	Removes torchvision requirement; inlines kwargs; narrows torch requirement to post-processing.
src/transformers/models/mobilevit/image_processing_mobilevit.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/mobilenet_v2/image_processing_pil_mobilenet_v2.py	Removes torchvision requirement; inlines kwargs; narrows torch requirement to post-processing.
src/transformers/models/mobilenet_v2/image_processing_mobilenet_v2.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/mllama/image_processing_mllama.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/maskformer/image_processing_maskformer.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/mask2former/modular_mask2former.py	Inlines kwargs and removes torchvision requirement for PIL subclass.
src/transformers/models/mask2former/image_processing_mask2former.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/llava_onevision/modular_llava_onevision.py	Inlines kwargs for PIL variant; adjusts resample typing.
src/transformers/models/llava_onevision/modeling_llava_onevision.py	Import formatting cleanup.
src/transformers/models/llava_onevision/image_processing_pil_llava_onevision.py	Removes torchvision requirement; inlines kwargs.
src/transformers/models/llava_onevision/image_processing_llava_onevision.py	Narrows resample typing.
src/transformers/models/llava_next/image_processing_pil_llava_next.py	Removes torchvision requirement; inlines kwargs.
src/transformers/models/llava_next/image_processing_llava_next.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/llava/image_processing_pil_llava.py	Narrows PIL resample typing.
src/transformers/models/llava/image_processing_llava.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/llama4/image_processing_llama4.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/lighton_ocr/modeling_lighton_ocr.py	Import formatting cleanup.
src/transformers/models/lightglue/modular_lightglue.py	Removes torchvision requirement for PIL subclass; inlines kwargs.
src/transformers/models/lightglue/modeling_lightglue.py	Import formatting cleanup.
src/transformers/models/lightglue/image_processing_lightglue.py	Removes vision-availability gating for PIL imports; imports PIL objects directly.
src/transformers/models/lfm2_vl/image_processing_lfm2_vl.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/levit/image_processing_pil_levit.py	Narrows PIL resample typing.
src/transformers/models/levit/image_processing_levit.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/layoutlmv3/image_processing_pil_layoutlmv3.py	Removes torchvision requirement; inlines OCR helpers + kwargs; handles pytesseract availability locally.
src/transformers/models/layoutlmv3/image_processing_layoutlmv3.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/layoutlmv2/image_processing_pil_layoutlmv2.py	Removes torchvision requirement; inlines OCR helpers + kwargs; handles pytesseract availability locally.
src/transformers/models/layoutlmv2/image_processing_layoutlmv2.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/kosmos2_5/image_processing_pil_kosmos2_5.py	Removes torchvision requirement; inlines kwargs + torch patch extraction helper; adds torch-only requires.
src/transformers/models/janus/image_processing_pil_janus.py	Removes torchvision requirement; inlines kwargs.
src/transformers/models/imagegpt/image_processing_pil_imagegpt.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample typing.
src/transformers/models/imagegpt/image_processing_imagegpt.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/idefics3/image_processing_idefics3.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/idefics2/image_processing_pil_idefics2.py	Removes torchvision requirement; inlines kwargs + helpers; avoids vision gating for PIL import.
src/transformers/models/idefics2/image_processing_idefics2.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/idefics/image_processing_pil_idefics.py	Removes torchvision requirement; inlines constants + kwargs.
src/transformers/models/grounding_dino/modular_grounding_dino.py	Inlines kwargs + narrows requires for PIL post-processing to torch only.
src/transformers/models/grounding_dino/image_processing_pil_grounding_dino.py	Removes torchvision requirement; inlines kwargs; narrows torch requirements to post-processing.
src/transformers/models/grounding_dino/image_processing_grounding_dino.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/got_ocr2/image_processing_got_ocr2.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/glpn/image_processing_pil_glpn.py	Removes torchvision requirement; inlines kwargs; narrows torch requirements to post-processing.
src/transformers/models/glpn/image_processing_glpn.py	Removes torchvision availability check; imports torch+tvF unconditionally.
src/transformers/models/glm_image/processing_glm_image.py	Adds explicit `@requires(backends=("torch",))` on processor class.
src/transformers/models/glm_image/modular_glm_image.py	Adds explicit `@requires(backends=("torch",))` on processor class; imports torch unconditionally.
src/transformers/models/glm_image/modeling_glm_image.py	Imports torch unconditionally; removes `is_torch_available()` guard.
src/transformers/models/glm_image/image_processing_pil_glm_image.py	Removes torchvision requirement; inlines kwargs + smart_resize.
src/transformers/models/glm_image/image_processing_glm_image.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/glm4v/image_processing_pil_glm4v.py	Removes torchvision requirement; inlines smart_resize helper.
src/transformers/models/glm4v/image_processing_glm4v.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/glm46v/video_processing_glm46v.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/glm46v/image_processing_pil_glm46v.py	Removes torchvision requirement; inlines kwargs + smart_resize.
src/transformers/models/glm46v/image_processing_glm46v.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/gemma3/image_processing_pil_gemma3.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample typing.
src/transformers/models/gemma3/image_processing_gemma3.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/efficientloftr/modular_efficientloftr.py	Adds kwargs type; makes torch optional at import time; narrows requires to torch on post-processing.
src/transformers/models/efficientloftr/image_processing_efficientloftr.py	Adjusts PIL/torch import strategy for torchvision backend.
src/transformers/models/dpt/modular_dpt.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/dpt/image_processing_dpt.py	Imports torch.nn.functional and tvF unconditionally; removes availability guards.
src/transformers/models/donut/image_processing_pil_donut.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample typing.
src/transformers/models/dinov3_vit/image_processing_dinov3_vit.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/detr/image_processing_detr.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/depth_pro/image_processing_depth_pro.py	Removes torchvision availability check; imports tvF unconditionally.
src/transformers/models/deformable_detr/modular_deformable_detr.py	Inlines kwargs + narrows requires for PIL post-processing to torch only.
src/transformers/models/deformable_detr/modeling_deformable_detr.py	Import formatting cleanup.
src/transformers/models/deformable_detr/image_processing_pil_deformable_detr.py	Removes torchvision requirement; inlines kwargs; narrows torch requirements to post-processing.
src/transformers/models/deformable_detr/image_processing_deformable_detr.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/deepseek_vl_hybrid/modular_deepseek_vl_hybrid.py	Adds numpy typing and narrows PIL resample typing in preprocess signatures.
src/transformers/models/deepseek_vl_hybrid/image_processing_pil_deepseek_vl_hybrid.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample typing.
src/transformers/models/deepseek_vl_hybrid/image_processing_deepseek_vl_hybrid.py	Narrows PIL resample typing.
src/transformers/models/deepseek_vl/modular_deepseek_vl.py	Introduces `DeepseekVLImageProcessorKwargs` and integrates `ImagesKwargs`.
src/transformers/models/deepseek_vl/image_processing_pil_deepseek_vl.py	Removes torchvision requirement; inlines kwargs.
src/transformers/models/convnext/image_processing_pil_convnext.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample typing.
src/transformers/models/convnext/image_processing_convnext.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/conditional_detr/modular_conditional_detr.py	Inlines kwargs + narrows requires for PIL post-processing to torch only.
src/transformers/models/conditional_detr/image_processing_conditional_detr.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/cohere2_vision/modeling_cohere2_vision.py	Import formatting cleanup.
src/transformers/models/cohere2_vision/image_processing_cohere2_vision.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/chmv2/image_processing_chmv2.py	Imports torch.nn.functional and tvF unconditionally; removes availability guards.
src/transformers/models/chameleon/image_processing_chameleon.py	Imports torch and tvF unconditionally; removes availability guards.
src/transformers/models/bridgetower/image_processing_pil_bridgetower.py	Removes torchvision requirement; inlines kwargs + resize helper; narrows PIL resample typing.
src/transformers/models/bridgetower/image_processing_bridgetower.py	Moves to unconditional torchvision functional import for torchvision backend.
src/transformers/models/beit/image_processing_pil_beit.py	Removes torchvision requirement; inlines kwargs; narrows torch requirements to post-processing.
src/transformers/models/beit/image_processing_beit.py	Imports torch/tvF unconditionally; removes availability guards.
src/transformers/models/aria/modular_aria.py	Imports tvF unconditionally; removes availability guards.
src/transformers/models/aria/modeling_aria.py	Import formatting cleanup.
src/transformers/models/aria/image_processing_pil_aria.py	Removes torchvision requirement; inlines kwargs; narrows PIL resample typing.
src/transformers/models/aria/image_processing_aria.py	Moves to unconditional torchvision functional import for torchvision backend.

Comments suppressed due to low confidence (1)

src/transformers/image_processing_backends.py:543

The resample type annotation was narrowed to PILImageResampling | None, but the implementation still explicitly accepts int and also maps from torch/torchvision interpolation modes (via torch_pil_interpolation_mapping). Please keep the signature type hint aligned with actual accepted inputs (e.g., include int and tvF.InterpolationMode / Union[...]) to avoid misleading API typing.

        self,
        image: np.ndarray,
        size: SizeDict,
        resample: "PILImageResampling | None" = None,
        reducing_gap: int | None = None,
        **kwargs,
    ) -> np.ndarray:
        """Resize an image using PIL/NumPy."""
        # PIL backend only supports PILImageResampling
        if resample is not None and not isinstance(resample, (PILImageResampling, int)):
            if torch_pil_interpolation_mapping is not None and resample in torch_pil_interpolation_mapping:
                resample = torch_pil_interpolation_mapping[resample]
            else:
                resample = PILImageResampling.BILINEAR
        resample = resample if resample is not None else PILImageResampling.BILINEAR

Copilot · 2026-03-28T06:58:54Z

src/transformers/models/vitpose/image_processing_vitpose.py

+import torch
+from torchvision.transforms.v2 import functional as tvF


torch/torchvision are imported unconditionally at module import time. Since these are optional dependencies and TorchvisionBackend's @requires(backends=(...)) only takes effect after the module imports, this can raise a raw ModuleNotFoundError (and break lazy imports) when torchvision isn't installed. Please revert to guarding these imports with is_torch_available() / is_torchvision_available() (or import inside methods) so missing deps surface via the standard requires_backends error path instead of crashing the import.

Copilot · 2026-03-28T06:58:54Z

src/transformers/models/vilt/image_processing_vilt.py

 import torch
+from torchvision.transforms.v2 import functional as tvF



This module imports torchvision.transforms.v2.functional at import time. Because torchvision is optional, this can raise ModuleNotFoundError before @requires(backends=(...)) guards on classes are applied, breaking lazy imports and producing an unfriendly error. Please guard the torchvision import with is_torchvision_available() (and/or move it inside functions) and rely on requires_backends for the user-facing error.

Copilot · 2026-03-28T06:58:55Z

src/transformers/models/zoedepth/image_processing_zoedepth.py

+from torchvision.transforms.v2 import functional as tvF



from torchvision.transforms.v2 import functional as tvF is executed unconditionally (outside any is_torchvision_available() guard). Since torchvision is an optional dependency, this can raise at module import time and bypass the requires_backends mechanism. Please restore a guarded import (or defer the import until the first actual use) so missing torchvision results in a controlled, informative error rather than an import crash.

Copilot · 2026-03-28T06:58:55Z

src/transformers/models/vitmatte/image_processing_pil_vitmatte.py

 from ...utils import TensorType, auto_docstring, is_torch_available
-from ...utils.import_utils import requires
-from .image_processing_vitmatte import VitMatteImageProcessorKwargs


 if is_torch_available():
-    import torch
+    pass


is_torch_available is imported and then used only in an empty if is_torch_available(): pass block. This is dead code and can confuse readers about whether torch is actually required here. Please remove the empty guard and the unused import (or, if torch is needed for typing, use TYPE_CHECKING and quoted annotations instead).

Copilot · 2026-03-28T06:58:55Z

src/transformers/models/superpoint/image_processing_superpoint.py

+from torchvision.transforms.v2 import functional as tvF



torchvision.transforms.v2.functional is imported unconditionally at module scope. Since torchvision is optional, this can raise ModuleNotFoundError during module import and bypass the library’s normal requires_backends/dummy-object behavior. Please guard this import with is_torchvision_available() (or defer it until use) so missing torchvision produces a controlled error only when the processor is actually used.

Copilot · 2026-03-28T06:58:55Z

src/transformers/models/superglue/image_processing_superglue.py

+from torchvision.transforms.v2 import functional as tvF



This file now imports torchvision.transforms.v2.functional at module import time. In environments without torchvision, that will crash the import (and can break _LazyModule access) instead of surfacing a standard requires_backends error when the processor is used. Please restore an is_torchvision_available() guard or move the import inside the code paths that need it.

ArthurZucker · 2026-03-30T06:56:50Z

src/transformers/models/beit/image_processing_pil_beit.py


        return processed_images

+    @requires(backends=("torch",))


cc @LysandreJik updated to work as such

github-actions · 2026-03-30T06:57:51Z

[For maintainers] Suggested jobs to run (before merge)

run-slow: aria, beit, bridgetower, chameleon, chmv2, cohere2_vision, conditional_detr, convnext, deepseek_vl, deepseek_vl_hybrid, deformable_detr, depth_pro, detr

LysandreJik

ok awesome!

LysandreJik · 2026-03-30T07:02:12Z

src/transformers/models/beit/image_processing_pil_beit.py


        return processed_images

+    @requires(backends=("torch",))


As seen with yu, I don't think this was initially designed for object-methods themselves, only for module-root objects

LysandreJik · 2026-03-30T07:03:08Z

src/transformers/models/conditional_detr/image_processing_pil_conditional_detr.py

        return encoded_inputs

-    @requires(backends=("vision", "torch"))
+    @requires(backends=("torch",))


awesome that it adds to the list rather than replaces it

@requires

…age processors (huggingface#45045) * [Bugfix] Remove incorrect torchvision requirement from PIL backend image processors PR huggingface#45029 added @requires(backends=("vision", "torch", "torchvision")) to 67 PIL backend image_processing_pil_*.py files. This causes PIL backend classes to become dummy objects when torchvision is not installed, making AutoImageProcessor unable to find any working processor. Fix: set @requires to ("vision",) for files that only need PIL, and ("vision", "torch") for files that also use torch directly. Also fix 5 modular source files so make fix-repo preserves the correct backends. Fixes huggingface#45042 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * [Bugfix] Remove redundant @requires(backends=("vision",)) from PIL backends Per reviewer feedback: the vision-only @requires decorator is redundant for PIL backend classes since PilBackend base class already handles this. - Remove @requires(backends=("vision",)) from 43 PIL backend files - Remove unused `requires` import from 38 files (Category A) - Keep @requires(backends=("vision", "torch")) on method-level decorators (Category B: 5 files) * update * remove torch when its not necessary * remove if typechecking * fix import shinanigans * marvellous that's how we protect torch :) * beit is torchvisionbackend * more import cleanup * fiixup * fix-repo * update * style * fixes * up * more * fix repo * up * update * fix imports * style * fix check copies * arf * converter up * fix? * fix copies * fix for func * style * ignore * type --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Arthur <arthur.zucker@gmail.com>

@requires

…age processors (huggingface#45045) * [Bugfix] Remove incorrect torchvision requirement from PIL backend image processors PR huggingface#45029 added @requires(backends=("vision", "torch", "torchvision")) to 67 PIL backend image_processing_pil_*.py files. This causes PIL backend classes to become dummy objects when torchvision is not installed, making AutoImageProcessor unable to find any working processor. Fix: set @requires to ("vision",) for files that only need PIL, and ("vision", "torch") for files that also use torch directly. Also fix 5 modular source files so make fix-repo preserves the correct backends. Fixes huggingface#45042 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * [Bugfix] Remove redundant @requires(backends=("vision",)) from PIL backends Per reviewer feedback: the vision-only @requires decorator is redundant for PIL backend classes since PilBackend base class already handles this. - Remove @requires(backends=("vision",)) from 43 PIL backend files - Remove unused `requires` import from 38 files (Category A) - Keep @requires(backends=("vision", "torch")) on method-level decorators (Category B: 5 files) * update * remove torch when its not necessary * remove if typechecking * fix import shinanigans * marvellous that's how we protect torch :) * beit is torchvisionbackend * more import cleanup * fiixup * fix-repo * update * style * fixes * up * more * fix repo * up * update * fix imports * style * fix check copies * arf * converter up * fix? * fix copies * fix for func * style * ignore * type --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Arthur <arthur.zucker@gmail.com>

ArthurZucker reviewed Mar 27, 2026

View reviewed changes

src/transformers/models/aria/image_processing_pil_aria.py Outdated Show resolved Hide resolved

ArthurZucker added 3 commits March 27, 2026 12:14

update

1417bc1

remove torch when its not necessary

be82ad8

remove if typechecking

ed3111a

ArthurZucker added 6 commits March 27, 2026 12:29

fix import shinanigans

c078455

marvellous that's how we protect torch :)

4099913

beit is torchvisionbackend

6e011bf

more import cleanup

a1bda86

fiixup

6e906b5

fix-repo

d266fed

zucchini-nlp reviewed Mar 27, 2026

View reviewed changes

src/transformers/models/beit/image_processing_pil_beit.py Outdated Show resolved Hide resolved

zucchini-nlp reviewed Mar 27, 2026

View reviewed changes

src/transformers/models/superpoint/image_processing_pil_superpoint.py Show resolved Hide resolved

LysandreJik approved these changes Mar 27, 2026

View reviewed changes

ArthurZucker added 7 commits March 27, 2026 16:02

update

756aa7c

style

2f0267a

fixes

f9d73c9

up

678dbc5

more

5a0e02d

fix repo

e8ea722

up

b15369e

update

a5c7481

ArthurZucker added the for patch Tag issues / labels that should be included in the next patch label Mar 27, 2026

ArthurZucker added 7 commits March 27, 2026 18:11

fix imports

b747308

style

58ddf7f

fix check copies

f5956fd

arf

5ef4900

converter up

5df7f82

fix?

8d445e5

fix copies

0a2c4f9

ArthurZucker approved these changes Mar 28, 2026

View reviewed changes

ArthurZucker requested review from Copilot, molbap and yonigozlan March 28, 2026 06:54

Copilot started reviewing on behalf of ArthurZucker March 28, 2026 06:54 View session

Copilot AI reviewed Mar 28, 2026

View reviewed changes

ArthurZucker mentioned this pull request Mar 30, 2026

Fix PIL backend fallback when torchvision is unavailable #45060

Open

fix for func

e153efd

ArthurZucker reviewed Mar 30, 2026

View reviewed changes

style

2e76424

LysandreJik approved these changes Mar 30, 2026

View reviewed changes

ArthurZucker added 2 commits March 30, 2026 09:13

ignore

c2510fe

type

93239c1

ArthurZucker merged commit 2da00a3 into huggingface:main Mar 30, 2026
28 checks passed

		# Adapted from transformers.models.fuyu.image_processing_fuyu.FuyuBatchFeature
		class FuyuBatchFeature(BatchFeature):

		@requires(backends=("torch",))
		class Mask2FormerImageProcessorPil(PilBackend):

		@requires(backends=("torch",))
		class Owlv2ImageProcessorPil(PilBackend):

		# Exception: for image_processing_pil files, image_processing modular classes must be inlined (not excluded),
		# because these two files must never import from each other.

		import torch
		from torchvision.transforms.v2 import functional as tvF

Conversation

Lidang-Jiang commented Mar 27, 2026 • edited by ArthurZucker Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Isolate dependencies, make PIL independant from Torchvision backend

Fixes:

Uh oh!

Uh oh!

HuggingFaceDocBuilderDev commented Mar 27, 2026

Uh oh!

ArthurZucker commented Mar 27, 2026

Uh oh!

Lidang-Jiang commented Mar 27, 2026

CI Failure Analysis

Failed tests

Root cause

Uh oh!

ArthurZucker commented Mar 27, 2026

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

LysandreJik left a comment

Choose a reason for hiding this comment

Uh oh!

Lidang-Jiang commented Mar 27, 2026

Uh oh!

ArthurZucker commented Mar 27, 2026

Uh oh!

Lidang-Jiang commented Mar 27, 2026

Uh oh!

ArthurZucker left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Mar 28, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 28, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 28, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 28, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 28, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 28, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Lidang-Jiang commented Mar 27, 2026 •

edited by ArthurZucker

Loading