DPOTrainer: pick vision path by processing object type; always pass a tokenizer to chat template

[ginkyenglee]

Feature request

• Decide is_vision_model by the actual processing object type:

  • PreTrainedTokenizerBase → text-only path (tokenize_row)
  • ProcessorMixin (with .tokenizer) → vision path (process_row)

• In the chat-template step, always pass a tokenizer (if a processor was provided, use its .tokenizer). This keeps the chat templating code agnostic to multimodal processors.

Motivation

When running DPO on multimodal families (e.g., Gemma 3) with text-only data, users pass a tokenizer. The current code infers is_vision_model from model.config.model_type and subsequently calls process_row, which assumes a processor and accesses .tokenizer, leading to:

AttributeError: <TokenizerClass> has no attribute tokenizer

Your contribution

I will submit PR fot it