Fix #3982: Fix DPO Trainer support for Gemma 3 vision models

[akshay-babbar]

What

This PR addresses an issue with the DPO Trainer's handling of vision-language models, specifically for Gemma 3. The changes enhance model type detection to properly support image-text-to-text models.

Fixes #3982

Changes made:

• Import MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES from transformers' modeling_auto
• Update the is_vision_model check to include both vision-to-sequence and image-text-to-text model types
• Add pixel_values and pixel_attention_mask to the signature columns to properly process vision inputs

Testing

• Unit test added: test_dpo_trainer_gemma3_vision_model_detection in verifies that Gemma3 models are correctly identified as vision models and processed through the appropriate pipeline
  • Integration verification: Confirmed fix resolves the original issue where DPOTrainer incorrectly routed Gemma3 models through tokenizer path instead of processor path
  • End-to-end validation: Successfully instantiated DPOTrainer with Gemma3 vision models and verified is_vision_model=True

The test suite ensures that Gemma3 and other image-text-to-text models are properly detected and routed through the vision processing pipeline, preventing the processor/tokenizer confusion that caused training failures.

Review Request 🙏

I've included unit tests and verified the fix resolves the original issue. Would appreciate maintainer review when possible.

Thanks!

[qgallouedec]

MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES and AutoModelForVision2Seq are actually deprecated, can you simply replace it by MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES?

[akshay-babbar]

Sure, done!