Support VLM processors in is_chat_template_prefix_preserving

[qgallouedec]

Extend is_chat_template_prefix_preserving to accept VLM processors (not only tokenizers), and update its type hint accordingly.

Context: https://github.com/huggingface/trl/pull/5489/changes#r3087655676

Why

For VLMs, processor.apply_chat_template is not just an alias for processor.tokenizer.apply_chat_template. Checking prefix-preservation on the inner tokenizer can therefore diverge from what actually happens at training time. We want to call the check on the processor whenever one is available.

---

Note

Low Risk
Small, well-scoped change to a utility check and its tests; main risk is introducing an extra PIL/image dependency path when running the prefix-preservation check for processors.

Overview
is_chat_template_prefix_preserving now accepts either a PreTrainedTokenizer or a VLM ProcessorMixin and runs the prefix check via processing_class.apply_chat_template.

For processors, the check now builds multimodal messages (including a dummy image) via prepare_multimodal_messages so image-token expansion is exercised, and tests add a new require_vision case validating prefix-preservation on a processor template.

^{Reviewed by Cursor Bugbot for commit fee8f7e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 355700a58c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[albertvillanova]

Thanks.