perf(qwen): project only the final token during trainside generation

[Yuanjia-Lee]

Summary

The Qwen trainside autoregressive loops only sample logits[:, -1, :], but prefill projected every prompt position through lm_head. Pass the models' native logits_to_keep=1 contract so Qwen3, Qwen2.5-VL, and Qwen3.5 materialize only [B, 1, vocab] logits.

Qwen3.5 applies the same selection on its cacheless VeOmni path, where the full growing sequence is forwarded each step. Replay remains unchanged because it needs every response-token log-probability.

Related Issue

N/A

Test Plan

• ruff check / ruff format --check / python lint/check_docstring_lines.py on the three changed AR modules — pass
• Qwen3-0.6B, Transformers 5.6: last-token logits and KV match the full forward within tolerance; prefill peak 2.50 -> 1.39 GiB (B=4, L=512, bf16)
• Transformers 5.6 tiny-model checks:
  • Qwen2.5-VL: final-token logits match the full projection within tolerance; KV length unchanged
  • Qwen3.5 cacheless path: final-token logits match the full projection within tolerance
• Live PE train: not run; sampling and cache-update paths are unchanged, and this matches the native generate() behavior.

Compatibility / Risk

No config, checkpoint, state-dict, or public API changes. The repository pins Transformers 5.6, and the HF plus VeOmni Qwen forwards all support logits_to_keep. Prompts are left-padded, so the final row is each sample's last real token.

Reviewer Notes

Open #430 edits qwen3/ar.py on the replay exact-actor path, not these generation-loop arguments.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.

[CjhHa1]

LGTM