[Chenhao-Guan]

What does this PR do?
Fixes a RuntimeError: Expected all tensors to be on the same device... when training EomtForUniversalSegmentation.

The EomtImageProcessor returns labels (mask_labels, class_labels) as a list[torch.Tensor] because the number of masks per image varies. The standard BatchFeature.to(device) call in a training loop does not move tensors inside these lists, leaving them on the CPU.

This PR adds a robust check at the beginning of the EomtForUniversalSegmentation.forward() method. It automatically moves any provided mask_labels or class_labels to the same device as the pixel_values, ensuring all inputs are on the correct device before loss computation.

This fixes the device mismatch bug internally without requiring any changes from the user.

Fixes #42205

Before submitting
[x] This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).

[x] Did you read the contributor guideline, Pull Request section?

[x] Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case. (Fixes #42205)

[x] Did you make sure to update the documentation with your changes? (N/A - Bugfix)

[x] Did you write any new necessary tests? (Verified with issue's code)

Who can review?
@NielsRogge @molbap @merveenoyan

[github-actions]

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

run-slow: eomt, mask2former, maskformer, oneformer

[Chenhao-Guan]

additional mask2former, maskformer, oneformer added,following the suggestion of @NielsRogge. However,mask2former, maskformer, oneformer were not tested on the actual model.

[NielsRogge]

Thanks a lot! Will assign @yonigozlan for review

[yonigozlan]

Happy to add this for now. A better solution imo would be to allow the .to method of BatchFeature objects to work on list of tensors. I think there was some pushback against this in the past, but I'll check internally.
Thanks @Chenhao-Guan for contributing this!

[HuggingFaceDocBuilderDev]

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]

Thanks a lot for the fix and for tracking this down! The underlying issue (#42205) was solved generically in #42884: BatchFeature.to() now recursively moves tensors inside lists/tuples (see maybe_to in src/transformers/feature_extraction_utils.py), so mask_labels/class_labels reach the right device without per-model checks in forward(). Closing this one as superseded, but feel free to reopen if you still hit the device mismatch on main.