Fixes #48556

AutoProcessor.from_pretrained("OpenGVLab/InternVL3-1B-hf") without torchvision installed raises AttributeError: 'NoneType' object has no attribute '__module__' at video_processing_auto.py:363, instead of the "requires torchvision" error added in #44125.

has_local_code is true when the config type is a key in VIDEO_PROCESSOR_MAPPING, but without torchvision every value in that mapping is None. For InternVL video_processor_class is also None — its image processor is GotOCRImageProcessor and no GotOCRVideoProcessor exists to infer from, which line 330 already handles — so (None or None).__module__ raises. Salesforce/instructblip-vicuna-7b fails the same way, and InstructBLIP is image-only.

This mirrors image_processing_auto.py:655-664, which guards the None and computes explicit_local_code only inside if has_remote_code: — the sole place it is used.

After the change both models reach the intended error:

ValueError: OpenGVLab/InternVL3-1B-hf requires `torchvision` to be installed.

The test mirrors test_unavailable_backend_error_mentions_missing_dependency in tests/models/auto/test_image_processing_auto.py. It uses patch.dict(VIDEO_PROCESSOR_MAPPING._extra_content, ...) to force the None value while torchvision is installed in CI; happy to change that if you prefer a different seam.

Differentiation from existing PRs

#44125 added the torchvision message at the end of this function but did not guard this earlier path. #45045 fixed a spurious torchvision requirement on PIL-backend image processors, a different code path. No open PR touches explicit_local_code in video_processing_auto.py.

Tests

pytest tests/models/auto/test_video_processing_auto.py
  12 passed

pytest tests/models/auto/test_processing_auto.py tests/models/auto/test_image_processing_auto.py
  41 passed, 5 skipped

pytest tests/models/auto/test_video_processing_auto.py -k unavailable_backend
  before: 1 failed (AttributeError: 'NoneType' object has no attribute '__module__')
  after:  1 passed

ruff check  - All checks passed
ruff format - 2 files already formatted

Reproduced on transformers 5.16.1 and main @ c93057d, Python 3.11.15, macOS arm64, torchvision not installed.

AI usage disclosure

Per CONTRIBUTING: this change was made with AI assistance. I hit the defect while loading a Qwen VL processor for a fine-tuning run on a machine without torchvision.

• Drafted and reproduced with Claude Opus 5.
• Reviewed by GLM 5.3 and Muse Spark 1.3 as judges, and by me before submitting.
• Coordination link: AutoProcessor.from_pretrained raises AttributeError: 'NoneType' object has no attribute '__module__' when torchvision is missing #48556
• I have reviewed every changed line, validated the behaviour end to end on both affected models, and run the test commands above.