knowledge: daily intake batch 16

[tzhouam]

Daily knowledge-intake batch from omni-reviewbot: executable rules distilled from merged PRs and copilot bugfix runs, append-only per the knowledge contract. Merging this PR is the human promotion gate — review the rules like any other knowledge edit.

Proposed rules

• knowledge/repos/vllm-omni/components/diffusion/rules-attention.md → DIFF-1ai (sources: PR #7047)
• knowledge/repos/vllm-omni/components/diffusion/rules-component-lifecycle.md → DIFF-2ag (sources: PR #7047)
• knowledge/repos/vllm-omni/models/qwen-omni/rules.md → QOMNI-1h (sources: PR #7345)
• knowledge/repos/vllm-omni/components/model-executor/rules-platform-backends.md → EXEC-10f (sources: PR #7393)
• knowledge/repos/vllm-omni/models/qwen3-tts/rules.md → Q3TTS-5a (sources: PR #6639)
• knowledge/repos/vllm-omni/ci/rules-test-fixtures.md → OMNI-CI-2i (sources: PR #7534)
• knowledge/repos/vllm-omni/models/mammoth-moda2/rules.md → MAMMO-1c (sources: PR #7102)
• knowledge/repos/vllm-omni/components/serving/rules-request-input.md → SERV-4r (sources: PR #7463)

Source events

• merged_pr PR #7047: [Model][MiniMax-H3] Multi-GPU serving fixes and layer-wise offload component selection
• merged_pr PR #7345: [Bugfix][Qwen3-Omni] Fix synchronous Thinker pipeline parallelism
• merged_pr PR #7477: [Qwen2.5-Omni] Avoid redundant embedding computation
• merged_pr PR #6639: [Bugfix][Hardware][Ascend] Fix Qwen3-TTS code predictor dtype and nested graph replay
• merged_pr PR #7531: [BugFix] Fix importing renamed module vllm.entrypoints.openai.engine.…
• merged_pr PR #7534: [Tests] Skip global GPU cleanup for parallel online diffusion tests
• merged_pr PR #7102: [Performance] Optimize MammothModa2 AR → DiT hidden-state transfer
• merged_pr PR #7463: [Model] Stop MiniCPM-o Thinker at TTS boundaries
• merged_pr PR #7098: [Misc][HiggsAudioV3] Diagnose skipped reference audio substitution
• merged_pr PR #7242: [Frontend][Benchmark] Add duplex performance metrics for OmniInteract and Omni-DuplexEval benchmark
• merged_pr PR #6898: [Bugfix][Qwen3-TTS] Gate async-chunk Code2Wav on non_streaming_mode (#4371)
• merged_pr PR #7108: [Model] Add VoxCPM2 startup LoRA adapter support
• merged_pr PR #7271: [Bugfix][MiniCPM-o] Attach a unit's whole frame group and size the vision slot budget like Stage0
• merged_pr PR #7338: [Kernel] Fuse Ming streaming ISTFT overlap-add and normalization
• merged_pr PR #7192: [Hardware][Ascend][Model] Support MOSS-TTS-Nano on Ascend A2
• merged_pr PR #6334: [Bugfix] Don't shutdown Engine on py_generator=True
• merged_pr PR #7541: [Bugfix] Use current OpenAI audio event names in duplex client metrics
• merged_pr PR #7313: [Refactor][Diffusion] Complete generic backend plan consumption
• merged_pr PR #7393: [Bugfix] Restore the async PP sampled-token broadcast in the AR runner
• merged_pr PR #7397: [CI][ROCm] Add non-blocking engine and model executor GPU coverage

Public SDK curation result

{
  "batch_id": "sha256:e7016e66b0eb270966d97636a09ab53d1786fbc06320dc5d5ac7fc592afb2b3c",
  "success": true,
  "attempted": 8,
  "applied": 8,
  "accepted_indexes": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7
  ],
  "rejected_indexes": [],
  "updated_document_ids": [
    "knowledge/repos/vllm-omni/ci/rules-test-fixtures.md",
    "knowledge/repos/vllm-omni/components/diffusion/rules-attention.md",
    "knowledge/repos/vllm-omni/components/diffusion/rules-component-lifecycle.md",
    "knowledge/repos/vllm-omni/components/model-executor/rules-platform-backends.md",
    "knowledge/repos/vllm-omni/components/serving/rules-request-input.md",
    "knowledge/repos/vllm-omni/models/mammoth-moda2/rules.md",
    "knowledge/repos/vllm-omni/models/qwen-omni/rules.md",
    "knowledge/repos/vllm-omni/models/qwen3-tts/rules.md"
  ],
  "updated_on": "2026-09-16",
  "validators": [
    {
      "validator_id": "knowledge/tools/check_knowledge_tree.py",
      "passed": true,
      "status": "passed",
      "returncode": 0,
      "output": "提醒：文件接近拆分线：general/review/guides/review-execution-contract.md (183 个非空行，21548 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/ci/rules.md (243 个非空行，27775 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/configuration/rules.md (174 个非空行，30403 bytes)\n提醒：目录有 15 个普通页面，应该考虑分类：repos/vllm-omni/components/configuration\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules-attention.md (196 个非空行，26488 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules-checkpoint-loading.md (211 个非空行，26702 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules-output-lifecycle.md (147 个非空行，21047 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules-system-runtime.md (189 个非空行，27138 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules.md (172 个非空行，27612 bytes)\n提醒：目录有 22 个普通页面，应该考虑分类：repos/vllm-omni/components/diffusion\n提醒：文件接近拆分线：repos/vllm-omni/components/model-executor/rules-bridge-batch.md (134 个非空行，18619 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/model-executor/rules-platform-backends.md (93 个非空行，17449 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/model-executor/rules.md (193 个非空行，31439 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/scheduler/rules.md (299 个非空行，32496 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/serving/rules-engine-lifecycle.md (246 个非空行，32271 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/serving/rules.md (196 个非空行，31197 bytes)\n提醒：目录有 12 个普通页面，应该考虑分类：repos/vllm-omni/components/serving\n提醒：文件接近拆分线：repos/vllm-omni/models/hunyuan-image3/rules.md (137 个非空行，25387 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/minicpm-o-4-5/rules.md (258 个非空行，30441 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/minimax-h3/rules-cache-task.md (204 个非空行，28481 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/minimax-h3/rules-deployment.md (236 个非空行，27707 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/minimax-h3/rules.md (183 个非空行，32421 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/qwen3-tts/rules.md (196 个非空行，22331 bytes)\n知识目录检查通过：0 个错误，23 个提醒"
    },
    {
      "validator_id": "knowledge/tools/check_wiki_lint.py",
      "passed": true,
      "status": "passed",
      "returncode": 0,
      "output": "wiki lint 通过：0 个错误，0 个提醒（共 388 页）"
    }
  ],
  "rolled_back": false
}

The vllm-omni release audit runs in CI on this PR as usual.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.