Add DeepSeek-R1-Distill chat template

[qgallouedec]

Ports the R1-Distill chat template support from https://github.com/huggingface/trl-internal/pull/256, reworked to match how the other reasoning families are handled here.

• deepseek_r1_distill.jinja — the template as shipped, byte-identical across every distill checkpoint (Qwen and Llama).
• deepseek_r1_distill_template — new-style response template only, like LFM2.5 and Gemma 4. Splits out reasoning_content (the internal version folded the whole <think> block into content); the start anchor consumes the prefilled <｜Assistant｜><think>\n, so a turn without a closing </think> parses as pure content. No tool calls: the template drops tool_calls whenever content is set, so supports_tool_calling rejects it.
• deepseek_r1_distill_training.jinja — adds {% generation %} markers for SFT assistant-only loss, and stops splitting on </think> so the reasoning stays in the training target.
• Tests: new trl-internal-testing/tiny-Qwen2ForCausalLM-R1-Distill fixture, added to the training-template and tool-calling suites.

---

Note

Low Risk
Additive chat-template and test coverage only; no auth, data, or core training-path refactors beyond optional template swapping for matching tokenizers.

Overview
Adds first-class DeepSeek-R1-Distill support in TRL’s chat-template stack so SFT can keep reasoning in the loss and response parsing matches the prefilled <｜Assistant｜><think>\n format.

Ships deepseek_r1_distill.jinja (byte-identical to upstream distill checkpoints) and deepseek_r1_distill_training.jinja, which drops the native </think> split on assistant content and wraps assistant output in {% generation %} markers. get_training_chat_template and add_response_schema recognize the family via template identity and attach deepseek_r1_distill_template (reasoning_content + content; no tool-call parsing). Tool calling is explicitly not supported (native template drops tool_calls when content is set).

Docs/README list the family; tests cover trl-internal-testing/tiny-Qwen2ForCausalLM-R1-Distill in training-template and supports_tool_calling suites, plus a script to build the tiny hub fixture from DeepSeek-R1-Distill-Qwen-1.5B.

^{Reviewed by Cursor Bugbot for commit 65edd5f. 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.

[AmineDiro]

Nice addition 👍🏼