Fix add_response_schema for VLM processors

[qgallouedec]

PR #5323 added VLM support for parse_response in GRPOTrainer via runtime workarounds: a schema-propagation block in _generate_and_score_completions, a compound has_response_schema check in __init__, and a getattr-unwrap at the decode gate.

The root cause was in add_response_schema: when given a processor, it set response_schema on the outer processor, but parse_response is a tokenizer method that reads self.response_schema from the inner tokenizer. Setting it on the processor only made hasattr checks pass without enabling parsing: #5323 was plastering over this leak.

Fix

add_response_schema now accepts PreTrainedTokenizer | ProcessorMixin and unwraps to .tokenizer before setting the schema, mirroring parse_response on the read side.

With the schema set where it's needed, the workarounds from #5323 collapse:

• _generate_and_score_completions: propagation block removed, decode gate simplifies to a single tokenizer check
• __init__: has_response_schema becomes a single getattr on the unwrapped tokenizer
• isinstance(..., PreTrainedTokenizerBase) exclusion at the decode gate removed — works uniformly for tokenizers and processors

Same simplification applied to experimental DPPOTrainer for consistency.

Tests

• TestAddResponseSchema split into test_add_response_schema (LLMs, AutoTokenizer) and test_add_response_schema_vlm (VLMs, AutoProcessor). The VLM test asserts the schema lands on the inner tokenizer.
• TestParseResponse gains a _load helper that dispatches to AutoTokenizer / AutoProcessor based on model name (*ForCausalLM vs *ForConditionalGeneration) and sets self.is_vlm. VLM tokenization goes through prepare_multimodal_messages to structure content, and all apply_chat_template calls are normalized to tokenize=True, return_dict=True.

---

Note

Medium Risk
Touches response parsing and tool-calling decode paths used during GRPO/DPPO rollouts; mistakes could break tool-call extraction or completion decoding for multimodal processors.

Overview
Fixes add_response_schema to accept either a tokenizer or a VLM ProcessorMixin and always set response_schema on the underlying tokenizer (while matching templates against the processor’s top-level chat_template). parse_response/call sites are updated to treat processors and tokenizers uniformly.

Simplifies GRPO/DPPO completion decoding gates by checking response_schema only on the unwrapped tokenizer (removing prior processor-specific workarounds), and expands tests to cover VLM processors via AutoProcessor, including multimodal message preparation and consistent apply_chat_template(tokenize=True, return_dict=True) usage.

^{Reviewed by Cursor Bugbot for commit 90153c2. 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: 7561aecd98

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".

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9f41214. Configure here.}

[cursor]

Inconsistent getattr dispatch vs isinstance in same file

Low Severity

parse_response uses getattr(processing_class, "tokenizer", processing_class) for processor-vs-tokenizer dispatch, while add_response_schema in the same file now uses the cleaner isinstance(processing_class, ProcessorMixin) check for the exact same pattern. The project rules say to avoid hasattr and getattr when a cleaner alternative exists — and this PR already demonstrates that alternative a few hundred lines above.

Additional Locations (1)

• trl/chat_template_utils.py#L326-L330

 

^{Triggered by project rule: ../.ai/AGENTS.md}

^{Reviewed by Cursor Bugbot for commit 9f41214. Configure here.}

[albertvillanova]

Thanks.

[albertvillanova]

I think I understand now! This is the fixing change!