Simplify tokenization [1/N]: Remove redundant is_vlm parameter

[albertvillanova]

Simplify tokenization [1/N]: Remove redundant is_vlm parameter.

This PR refactors the tokenization logic in the DPO, KTO, and SFT trainers to simplify the handling of VLM (Vision-Language Model) processors. The main change is to remove the explicit is_vlm parameter from the tokenization functions and instead infer whether the processor is a VLM by checking its type. This reduces redundancy and streamlines the code.

Changes

Tokenization logic simplification:

• Removed the is_vlm parameter from the _tokenize and related tokenize_fn functions in dpo_trainer.py, kto_trainer.py, and sft_trainer.py. The code now determines if the processor is a VLM by checking if it is an instance of ProcessorMixin.
• Updated all calls to _tokenize and tokenize_fn to remove the is_vlm argument, passing only the necessary parameters.

Documentation updates:

• Removed references to the is_vlm parameter from function docstrings to reflect the new logic and avoid confusion.

Behavioral consistency:

• The logic for preparing multimodal messages is now consistently applied based on processor type, ensuring that the correct processing is performed for both VLM and non-VLM cases.

These changes make the codebase easier to maintain and less error-prone by reducing unnecessary parameters and centralizing the logic for determining processor type.

---

Note

Low Risk
Mechanical refactor with the same VLM branch logic; risk is only if a non-processor were misclassified, which matches existing trainer processing_class typing.

Overview
DPO, KTO, and SFT dataset preprocessing no longer threads an explicit is_vlm flag through _tokenize and the dataset.map tokenize_fn helpers.

Inside _tokenize, vision-language handling (multimodal message prep and unwrapping the processor batch dimension) is keyed off isinstance(processing_class, ProcessorMixin) instead of a caller-supplied boolean. fn_kwargs for mapping drop is_vlm / self._is_vlm, and docstrings no longer document that parameter.

Intended behavior for tokenizer vs processor paths is unchanged; this is API cleanup ahead of further tokenization refactors.

^{Reviewed by Cursor Bugbot for commit 604866f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

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.