[ss1127201970]

Purpose

Fixes #7941.

Higgs Audio v3 can emit multiple incomplete codec frames during EOC ramp-down. Replacing their special/pad tokens with code 0 and trimming one final frame leaves a mixed EOC/audio frame in the decoder input, causing a sentence-tail noise burst.

Keep the contiguous complete-frame prefix after reversing the delay pattern. Apply it in the shared whole-utterance and streaming helpers, so sync, token-only, MRV2 full-payload, and scalar/batched streaming entrypoints agree. Invalid frames are excluded from right context too; emitted-frame counts and context metadata reflect the actual retained frames. Complete final frames are preserved.

Rebased onto f69b1f2b19cece03f6736c7da093ea25de0e24fd, including the MRV2 refactor from #8226. This does not include the separate local codec-boundary audio-quality experiment.

Test plan and results

The CPU regression matrix covers exact codec values/order, 0/1/25/144 real frames, mixed/all/no EOC tails, holdback 0/1/4, left context 0/3, chunk size 1/25, and finish arriving with the last row or separately. It now covers 612 cases across all five entrypoints. No checkpoint is required.

python -m pytest tests/model_executor/stage_input_processors/test_higgs_audio_v3_tail.py -m 'core_model and cpu'
python -m pytest tests/model_executor/models/higgs_audio_v3/test_higgs_audio_v3.py -k StageInputProcessor

Current refresh:

• Fixed the Ruff formatting change reported by the previous CI run.
• SPDX/test-marker checks and git diff --check pass; independent code review found no blocking issues.
• Real pytest 9.1.1 was attempted against the refreshed source with installed PyTorch 2.13.0 / vLLM 0.28.0. Collection is blocked because current main imports KVConnectorBlockState, which that vLLM version does not provide. No tests ran in this attempt. A fixture-isolated retry reached the same package-import incompatibility. Full current-main pytest/GPU suites remain unverified; CI results are tracked separately.

Historical validation of the original fix on vLLM-Omni 0.28.0:

• All 300 original parameter combinations passed via direct invocation of the unchanged test bodies using an AST runner (not formal pytest collection); the original processor failed 192 cases.
• Real voice-clone synthesis on RTX 4090 D: 20 Chinese/English samples regenerated with seed 42. Model revision and chunk settings are in the issue.
• Same-code FP32 replay: 144 complete frames plus two incomplete frames became 145 decoder frames before the fix and 144 after it. Final 20 ms RMS decreased from 0.08618 to 0.000595 (~43 dB).

AI disclosure: Codex assisted with diagnosis, implementation, tests, and this write-up.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/input_output_modality_contracts.md, docs/design/module/model_integration.md, docs/design/module/ar_runtime.md.

Module owners: @alex-jw-brooks @amy-why-3459 @Gaohan123

Routing: @alex-jw-brooks via module of the changed files, semantic router; @amy-why-3459 via module of the changed files, semantic router; @Gaohan123 via module of the changed files, semantic router

@ss1127201970, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[vllm-omni-review-bot]

Omni ReviewBot triage note

Automated triage of commit 2c21046b01b5 produced:

• Priority: high. Prompt maintainer attention is suggested.

These are automated triage suggestions only — the final decision belongs to the maintainers.

[vllm-omni-review-bot]

Omni ReviewBot: superseded

The CI failure noted on 1336a970313c refers to an earlier head; the pull request now points at 2c21046b01b5.

[vllm-omni-review-bot]

Omni ReviewBot: no human activity for 8 days

@ss1127201970 this pull request has had no human commit, comment or review since 2026-09-21. Please confirm the current plan and next step. The author or a maintainer decides whether to change the PR state.

To keep it moving, any one of these is enough: push an update, reply to the open blocker, or post the current plan and timeline.

[vllm-omni-review-bot]

Omni ReviewBot routing record

Assigned Strict on zcode (GLM-5.3-Flash) under experiment fleet-strict-cursor-grok46-zcode-glm53flash-5050-c5-z10-20261002.

[ss1127201970]

Rebased onto current main and fixed the Ruff formatting failure.

Self-review: checked the shared whole-utterance and streaming helpers after #8226, including MRV2 full-payload and batched streaming. The regression matrix now covers all five entrypoints. I checked that invalid tail frames cannot enter audio or right context, valid final frames remain intact, and frame counters/context metadata follow the frames actually emitted. Independent review found no blocking issues.

SPDX, test-marker and diff checks pass. Formal pytest collection is currently blocked by an environment mismatch: main requires KVConnectorBlockState, absent from the available vLLM 0.28.0 runtime. The PR description separates this from the earlier runtime/audio validation; I have not counted collection failure as a test pass.

@vllm-omni-review-bot please review the updated diff. Maintainer review is welcome as well.

[vllm-omni-review-bot]

Omni ReviewBot attempt record

Review attempt ended as failed (step 'review' (agent.review_diff): unhandled error: RuntimeError: zcode exited 1 without a result event: statusCode: undefined } Error: Turn execution failed (traceId: da8c57d0-664f-489f-947a-78de5b6c6700) — check zcode login and the CLI version; retrying strict/zcode/GLM-5.3-Flash in 120s (try ).

[vllm-omni-review-bot]

Omni ReviewBot attempt record

Review attempt ended as failed (step 'review' (agent.review_diff): unhandled error: RuntimeError: zcode exited 1 without a result event: statusCode: undefined } Error: Turn execution failed (traceId: e6fc87b2-3eef-4e06-8886-28aad2cb5466) — check zcode login and the CLI version; retrying strict/zcode/GLM-5.3-Flash in 600s (try ).

[vllm-omni-review-bot]

Omni ReviewBot attempt record

Review attempt ended as failed (step 'review' (agent.review_diff): unhandled error: RuntimeError: zcode exited 1 without a result event: statusCode: undefined } Error: Turn execution failed (traceId: 4f1f8f28-876f-4948-b67a-c6302013abc7) — check zcode login and the CLI version; falling back to direct/cursor/auto).

[vllm-omni-review-bot]

Omni ReviewBot review

0 actionable finding(s).

CI at 2c21046b01b5 (2026-10-10T04:46:55.252558+00:00): required check(s) blocking: buildkite/vllm-omni (missing).

Note: The assigned review arm strict/zcode/GLM-5.3-Flash could not complete this review, so it was produced by the fallback arm direct/cursor/auto. It is excluded from the routing experiment.

Full review analysis

PR description

Higgs Audio v3 now keeps only the contiguous complete codec-frame prefix after delay-pattern reversal. A frame that still contains an end-of-code or pad value is excluded from decoder input, including streaming right context, and later frames are left on the original timeline. The same helper feeds the sync, token-only, and full-payload paths and the scalar and batched streaming flush, so emitted-frame counts and context metadata follow the frames that are actually retained. Complete final frames stay in the codec input.

Change flow

flowchart LR
  rows["[EXISTING] Delayed talker code rows"]:::existing
  helpers["[CHANGED] Whole-utterance and stream flush"]:::changed
  prefix["[NEW] _valid_code_prefix"]:::new
  codec["[CHANGED] Codec frames and context metadata"]:::changed
  tests["[NEW] Five-entrypoint tail matrix"]:::new
  rows --> helpers
  helpers --> prefix
  prefix --> codec
  tests --> helpers
  classDef existing fill:#e5e7eb,stroke:#6b7280,color:#111827
  classDef changed fill:#fef3c7,stroke:#d97706,color:#451a03,stroke-width:2px
  classDef new fill:#dcfce7,stroke:#16a34a,color:#052e16,stroke-width:2px
  classDef removed fill:#fee2e2,stroke:#dc2626,color:#450a0a,stroke-width:2px

Loading

No actionable findings.

---

_{🤖 This review was generated by InferMatrix Copilot, an open-source repo-maintenance agent for PR review, CI debugging and issue triage. Try it on your own repo, and ⭐ star it if it helped!}

[vllm-omni-review-bot]

Omni ReviewBot: no human activity for 7 days

@ss1127201970 this pull request has had no human commit, comment or review since 2026-10-10. Please confirm the current plan and next step. The author or a maintainer decides whether to change the PR state.

To keep it moving, any one of these is enough: push an update, reply to the open blocker, or post the current plan and timeline.

[vllm-omni-review-bot]

Omni ReviewBot: no human activity for 14 days

@ss1127201970 this pull request has had no human commit, comment or review since 2026-10-10. Please consider marking this PR as draft until work can resume. The author or a maintainer decides whether to change the PR state.

To keep it moving, any one of these is enough: push an update, reply to the open blocker, or post the current plan and timeline.