pipeline import fails with torchvision::nms operator mismatch

[paytonison]

System Info

Traceback (most recent call last):
  File "/usr/local/bin/transformers", line 5, in <module>
    from transformers.commands.transformers_cli import main
  File "/usr/local/lib/python3.11/dist-packages/transformers/commands/transformers_cli.py", line 19, in <module>
    from transformers.commands.add_new_model_like import AddNewModelLikeCommand
  File "/usr/local/lib/python3.11/dist-packages/transformers/commands/add_new_model_like.py", line 26, in <module>
    from ..models.auto.image_processing_auto import IMAGE_PROCESSOR_MAPPING_NAMES
  File "/usr/local/lib/python3.11/dist-packages/transformers/models/auto/image_processing_auto.py", line 27, in <module>
    from ...image_processing_utils import ImageProcessingMixin
  File "/usr/local/lib/python3.11/dist-packages/transformers/image_processing_utils.py", line 21, in <module>
    from .image_processing_base import BatchFeature, ImageProcessingMixin
  File "/usr/local/lib/python3.11/dist-packages/transformers/image_processing_base.py", line 26, in <module>
    from .image_utils import is_valid_image, load_image
  File "/usr/local/lib/python3.11/dist-packages/transformers/image_utils.py", line 60, in <module>
    from torchvision.transforms import InterpolationMode
  File "/usr/local/lib/python3.11/dist-packages/torchvision/__init__.py", line 10, in <module>
    from torchvision import _meta_registrations, datasets, io, models, ops, transforms, utils  # usort:skip
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/dist-packages/torchvision/_meta_registrations.py", line 163, in <module>
    @torch.library.register_fake("torchvision::nms")
     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/dist-packages/torch/library.py", line 1069, in register
    use_lib._register_fake(
  File "/usr/local/lib/python3.11/dist-packages/torch/library.py", line 219, in _register_fake
    handle = entry.fake_impl.register(
             ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/dist-packages/torch/_library/fake_impl.py", line 50, in register
    if torch._C._dispatch_has_kernel_for_dispatch_key(self.qualname, "Meta"):
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: operator torchvision::nms does not exist

Who can help?

@Rocketknight1 @ArthurZucker

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

from transformers import pipeline

pipe = pipeline("text-generation", model="google/gemma-3-270m")

Expected behavior

For text-only pipelines, transformers should not crash due to torchvision import issues.

Either:

• Skip torchvision imports unless required (e.g., when instantiating image/audio pipelines), or
• Provide a clearer error message pointing to the Torch ↔︎ Torchvision version mismatch.

Workarounds:

• Uninstalling torchvision resolves the issue for text-only models.
• Alternatively, pinning matching torch / torchvision builds (e.g., torch==2.4.1, torchvision==0.19.1) avoids the error.

[Rocketknight1]

Hi @paytonison, you're right that we should probably guard torchvision in some cases! However, errors like these are usually caused by environment incompatibilities between torch / torchvision versions, or between torchvision and CUDA.