[qgallouedec]

Adds Mistral Medium 3.5's chat template (mistralai/Mistral-Medium-3.5-128B) and its training variant, so assistant_only_loss=True works for it. The original is already prefix-preserving: the training template only wraps the assistant turn (content, tool calls and </s>) in {% generation %} markers.

It also fixes processors that return tensors from apply_chat_template (PixtralProcessor defaults to return_tensors="pt"): get_training_chat_template died with "Boolean value of Tensor with more than one value is ambiguous" and is_chat_template_stop_token_trained returned False. Same unbatching in GRPO's and Distillation's _get_tool_suffix_ids.

• New tiny model trl-internal-testing/tiny-Mistral3ForConditionalGeneration (script added), used in the tool-calling and training-template tests.

---

Note

Low Risk
Changes are scoped to chat-template plumbing and VLM processor edge-case handling, with no auth or training-core logic altered beyond safer ID normalization.

Overview
Adds Mistral Medium 3.5 (mistralai/Mistral-Medium-3.5-128B) to TRL’s bundled chat templates: a reference mistral_medium_3_5.jinja and a training variant that only wraps the assistant turn (content, tool calls, EOS) in {% generation %} markers, since the base template is already prefix-preserving for GRPO tools.

Pixtral / Mistral processors can return torch.Tensor IDs from apply_chat_template even when lists are expected. The PR normalizes those tensors to lists in prefix-preserving checks, stop-token training checks, and GRPO/Distillation _get_tool_suffix_ids, fixing template selection and tool-suffix logic for this family.

Docs and check_chat_templates.py list the new family; a tiny Mistral3 hub script and test parametrization exercise tool-calling and training-template behavior (transformers ≥ 5.0, vision marks).

^{Reviewed by Cursor Bugbot for commit bc46223. Bugbot is set up for automated code reviews on this repo. Configure here.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-27T14:08:51.516192Z	e3753dd	PR opened

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e3753dd2e8

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".

[chatgpt-codex-connector]

Register Mistral response parsing before enabling tool training

When this processor is passed to GRPOTrainer or DistillationTrainer with tools, the constructors call add_response_schema before reaching this training-template mapping. The new tiny processor is copied directly from the checkpoint without adding a response_template or response_schema, while add_response_schema has no branch for mistral_medium_3_5_chat_template; initialization therefore raises Unrecognized chat template and never reaches either this mapping or the newly fixed _get_tool_suffix_ids path. Add the corresponding response parser mapping and cover the trainer's public initialization path rather than only supports_tool_calling.

AGENTS.md reference: AGENTS.md:L66-L70

Useful? React with 👍 / 👎.

[qgallouedec]

Not in this PR. A response parser alone moves the failure further down: GRPO's prompt tokenization also breaks on Pixtral processors, which return tensors, so the tool loop then fails on Tensor + list. Supporting GRPO with tools on Mistral 3 needs both, plus a place in the GRPO VLM tests, so it gets its own PR. This one is about SFT (the training template) and the template checks.

[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.

[albertvillanova]

Thanks! I checked the stored template against the Hub, the training-template diff, and the new tests on an H100. With the tensor-to-list conversion removed from chat_template_utils.py, 12 of the 13 mistral3 tests fail, so the fix is needed and covered. With assistant_only_loss=True, the mask covers both assistant turns and their </s>, and nothing from the system or user turns.

Two comments:

1. The tiny model is now FP8, and with the current transformers release it can't be trained: LoRA fails with "quantized with fp8 but that quantization method do not support training", and full fine-tuning with "cannot perform fine-tuning on purely quantized models". This PR only loads the processor, so nothing breaks here. With huggingface/transformers#50031, LoRA on the unmodified tiny trains (SFT with assistant_only_loss=True), so the GRPO tests planned in the follow-up can use a PEFT config and be gated on the transformers release that ships it.

2. Nit on the skip reason "Ministral 3 was introduced in transformers-5.0.0": these tests only load the processor, and what needs 5.0 there is the tokenizer. The tiny's tokenizer class is TokenizersBackend, added in 5.0.0 (huggingface/transformers#40936); on 4.57.6, AutoProcessor.from_pretrained fails with "Tokenizer class TokenizersBackend does not exist". Maybe "Mistral Medium 3.5 tokenizer requires transformers>=5.0.0", like the GLM4 and LFM2.5 params next to it?

[cursor]

Cursor Bugbot has reviewed your changes using default effort 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 b7aa518. Configure here.}