Add Qwen3.8 support and testing

[qgallouedec]

Released 2026-08-14: Qwen/Qwen3.8-27B the dense open-weights member of the Qwen3.8 family (+ an FP8 variant), reusing the existing Qwen3_5ForConditionalGeneration architecture, so no transformers gate beyond the Qwen3.5 floor.

• Tiny model script qwen3_8_for_conditional_generation.py, mirroring the Qwen3.5 script (same layer_types / full_attention_interval workaround, same vision depth / num_heads keys), pinned to transformers==5.2.0. The checkpoint is all-bf16, so unlike Qwen3.5 no fp32 restore is needed. Config diff vs the reference is 11 entries, all scale-downs.
• Vendored qwen3_8.jinja + qwen3_8_training.jinja. The template is new: it adds a reasoning_effort flag (xhigh default / medium / low) that prepends a reasoning-effort instruction to the system prompt, and defaults preserve_thinking to enabled. Tool calls keep the Qwen3.5 Hermes-style syntax, so it reuses qwen3_5_schema / qwen3_5_template.
• Test params next to Qwen3.6 in test_chat_template_utils.py (4 lists) and test_data_utils.py. Trainer tests keep the existing Qwen3.5 models as the architecture representatives.
• Tiny model pushed: trl-internal-testing/tiny-Qwen3_5ForConditionalGeneration-3.8.

---

Note

Low Risk
Additive support mirroring Qwen3.5/3.6; changes are limited to chat templating, docs, and test fixtures with no auth or data-handling impact.

Overview
Adds Qwen3.8 to TRL’s supported chat-template families so SFT (assistant_only_loss) and GRPO (tool calls) can auto-swap training templates like other Qwen VL models.

Vendors qwen3_8.jinja (upstream-style template with reasoning_effort system instructions and default preserve_thinking) and qwen3_8_training.jinja (always emits thinking blocks for prefix-preservation; wraps assistant output in {% generation %} markers). chat_template_utils maps the stock template to existing qwen3_5_schema for tool response parsing and returns the training template from get_training_chat_template.

Includes a tiny VLM hub script (qwen3_8_for_conditional_generation.py) for trl-internal-testing/tiny-Qwen3_5ForConditionalGeneration-3.8, and extends chat-template / data-utils tests plus docs to cover Qwen3.8 alongside Qwen3.6.

^{Reviewed by Cursor Bugbot for commit 2384046. 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.