[VoiceChat] Add accuracy validation and thinker parity test

[deepbuilder]

Stacked on sgl-project#34873 — base is rystewart-nvidia:voicechat-support, not main. Merging here first lets the VoiceChat work reach SGLang reviewers as one PR.

Two independent commits, docs and one test. No changes to model or config code.

1. Accuracy validation (874fc3b) — docs only

examples/voicechat/accuracy.md, plus one row in the Contents table.

Records measured agreement against the vLLM-Omni implementation of the same checkpoint (added in vllm-omni#5842), on a 15.61 s reference sample — 196 acoustic frames on the 12.5 Hz timeline. SGLang runs the thinker in bfloat16; the vLLM-Omni reference runs float32.

stage	comparison	result
Thinker (NemotronDuplexHForCausalLM)	frame-locked text timeline, token for token	196/196 (100%)
Audio to text, end to end	sidecar perception into the thinker, same reference timeline	196/196 (100%)

The second row is the deployed path: a WAV in, a text timeline out, through the sidecar's streaming perception encoder and the thinker together.

Three things the doc is explicit about, because each is easy to misread:

• Perception and codec are NeMo modules on both sides. nemo_audio_sidecar.py imports PerceptionCacheManager and RVQVAEModel from nemo.collections.speechlm2, and so does the reference. Comparing them measures streaming behaviour, not model correctness — the sidecar encodes frame by frame with a cache, the reference encodes the whole utterance at once. That difference is small and deterministic (cosine 0.99943, max abs 2.8e-02, identical with the perception CUDA graph on or off) and changes no tokens on this sample.
• The talker is not compared, by construction. It samples, so two implementations consume independent RNG streams and agree on only ~50% of codes even when both are correct. Measured 49.95%. No seed closes that. A real comparison would need pre-sampling distributions, teacher-forced, rather than emitted codes.
• Single sample. Exact agreement on the reference input. Not a claim that streaming perception is numerically equivalent to full-utterance encoding in general.

2. Thinker parity test (8af39fc)

test/manual/voicechat_thinker_parity.py. One new file, nothing else touched.

The thinker decodes greedily, so this is a true cross-implementation gate: the timeline can be compared against another implementation of the same model and must agree exactly. It does, 196/196.

Acoustic frames are supplied as a saved tensor rather than computed in the test, so it depends only on the thinker — no encoder, no audio stack, no sidecar. A failure points at one component.

No reference artifacts are checked in. Passing --emit without a reference runs the stage and writes its output.

Lives under test/manual/ because it needs the full checkpoint, a converted stage, and a GPU.

Still to come

Reproducibility details for the parity check are not in this PR yet: a checksum for the input WAV, the published vLLM-Omni artifacts (reference_text_tokens.json and the acoustic frames consumed by --reference-tokens and --acoustic-frames), and the vLLM-Omni commit and commands used to generate them. The input WAV and the 3.5 MB frames tensor need a download location first — see the discussion below.

Testing

• Thinker gate — 196/196 on the reference sample
• Lint clean under the repo's pinned black / isort / ruff / codespell

python/sglang/srt/models/eartts.py, python/sglang/srt/configs/eartts.py and test/registered/unit/models/test_voicechat_models.py are byte-identical to this branch's base — verified as zero diff lines.

---

CI States

Latest PR Test (Base): ❌ Run #32328257111
Latest PR Test (Extra): ❌ Run #32328257140