What does this PR do?

Adds a {% generation %}-marked training variant of the Qwen3-VL chat template so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

Diff vs qwen3_vl.jinja: wrap the assistant turn body — text content + tool_calls + closing <|im_end|>\n — with {% generation %} / {% endgeneration %}. The <|im_start|>assistant\n prompt cue stays outside the block; the tool and user branches are untouched, since tool responses are model input, not output.

Note: this code path is currently unreachable

SFTTrainer raises at trl/trainer/sft_trainer.py:1006 for any VLM with assistant_only_loss=True:

if self._is_vlm and args.assistant_only_loss:
    raise ValueError(
        "Assistant-only loss is not yet supported for vision-language models. ..."
    )

So the new template cannot be exercised by a real SFT training run today. This PR is forward-looking prep, submitted under the explicit sanction in #5471: "VLMs currently don't support assistant_only_loss in SFT (blocked by a separate check). These should still be tracked so templates are ready when support lands." When the guard lifts in a separate effort, Qwen3-VL training will work without an additional change here.

Validation surface, in the meantime, is the existing TestGetTrainingChatTemplate suite — prefix-preservation, behavior-unchanged, and mask-correctness checks against the rendered template string. No real SFT training was attempted.

Changes:

• trl/chat_templates/qwen3_vl_training.jinja: new training template, identical to qwen3_vl.jinja except for the {% generation %} / {% endgeneration %} markers wrapping the assistant body.
• trl/chat_template_utils.py: load the new template and add a dispatch branch in get_training_chat_template() for qwen3_vl_chat_template. Docstring updated to mention Qwen3-VL.
• tests/test_chat_template_utils.py: extend the TestGetTrainingChatTemplate parametrize with the trl-internal-testing/tiny-Qwen3VLForConditionalGeneration fixture — runs prefix-preservation, behavior-unchanged, and mask-correctness checks against the new template.
• trl/chat_templates/README.md and docs/source/chat_templates.md: short section describing the training template, mirroring the structure of the existing qwen3_training.jinja entries.

Part of #5471.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

AI writing disclosure

We welcome the use of AI tools to help with contributions. For transparency and to help us improve our review process, please indicate the level of AI involvement in this PR.

• No AI usage: the PR was written entirely by a human.
• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.
• AI-generated: the PR was mostly or fully generated by an AI tool.

Who can review?

@qgallouedec