[omni] Qwen3-Omni thinker: HF bridge + /generate rollout adapter

[yxs]

Summary

RL loop wiring the Qwen3-Omni-30B-A3B thinker (text MoE) into miles GRPO, against sglang-omni's
rollout /generate: sgl-project/sglang-omni#785
and distributed weight-sync: sgl-project/sglang-omni#784
endpoints. Text-only thinker path.

Components

• tools/extract_qwen3_omni_thinker.py — extracts the standalone Qwen3-MoE text backbone from the composite omni checkpoint (drops audio/visual, model_type=qwen3_moe) for the existing bridge.
• .../megatron_to_hf/qwen3omni_moe.py — Megatron→HF converter, body.-prefixed (the namespace [RL] distributed weight-sync sgl-project/sglang-omni#784 demuxes on); --model-name qwen3omni_moe.
• miles/rollout/generate_hub/omni_thinker.py — thin rollout adapter: reuses miles' payload/parse, adds the omni fields (output_modalities=["text"], return_omni_rollout=False, repetition_penalty=1.0).
• Example GRPO script + Megatron model args + fast tests.

Notes

• The omni /generate emits temp-1 (pre-temperature) full-vocab logprobs — measured directly: with greedy decoding the returned logprobs are identical at temperature 1.0 and 2.0. miles' train recompute divides logits by rollout_temperature, so the two agree only at temp=1; the adapter asserts rollout_temperature == 1.0. repetition_penalty is forced to 1.0 (the trainer recompute can't replay a repetition penalty).
• Guards: the omni path asserts on MoE/indexer replay and image inputs (server forward-declares both — fail loud, not silent).
• Off-policy until [RL] distributed weight-sync sgl-project/sglang-omni#784: point --sglang-router-ip/port at a standalone omni server; TIS + --get-mismatch-metrics absorb the gap.

Verification

• tests/fast/test_qwen3_omni_thinker.py — 13 passed (ion-b200 miles image): extraction, body.* converter, adapter contract + guards.
• Full cross-server E2E — real omni_thinker.generate against a live [RL] add Miles-compatible /generate rollout endpoint sgl-project/sglang-omni#785 omni server (Qwen3-Omni-30B-A3B on B200): input_ids rollout returns text, output_token_logprobs length == completion_tokens, cached_tokens int (no crash), finish_reason→status, weight_version echoed; logprob convention measured (temp-1).