🌋 Add support for LLaVA-Next in DPOTrainer

[chenweize1998]

What does this PR do?

This PR provides a quick fix for issue #2403. Previously, DPOTrainer did not support LLaVA-Next because the required image_sizes parameter from the LLaVA-Next forward function was being removed during data processing within DPOTrainer. This update modifies DPOTrainer to retain the image_sizes parameter if it is returned by the image processor and passes it to the model when present.

While this fix resolves the issue on my end - I successfully ran DPOTrainer with LLaVA-Next - but it has not been extensively tested. I would appreciate assistance or guidance on the next steps to ensure broader compatibility and robustness.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines.
• Did you write any new necessary tests?

Who can review?

Maybe @qgallouedec

[qgallouedec]

Great! Thanks @chenweize1998! Can you try to uncomment

trl/tests/test_dpo_trainer.py

Line 1142 in ac26778

# ("trl-internal-testing/tiny-LlavaNextForConditionalGeneration",),

[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.

[qgallouedec]

Thanks @chenweize1998!

[qgallouedec]

As expected, "Tests / Tests with dev dependencies" will fail until huggingface/transformers#34953 is merged. We can safely ignore this failing test.