diff --git a/knowledge/repos/vllm-omni/components/model-executor/_index.md b/knowledge/repos/vllm-omni/components/model-executor/_index.md index 5bb78871..aaf7734b 100644 --- a/knowledge/repos/vllm-omni/components/model-executor/_index.md +++ b/knowledge/repos/vllm-omni/components/model-executor/_index.md @@ -69,3 +69,4 @@ sources: [] - [MiniCPM-o duplex turn-end drain 规则](rules-minicpmo-turn-end-drain.md) - [MiMo-Audio 设备与平台支持](rules-mimo-audio-device.md) - [NPU model adapter rules](rules-npu-model-adapters.md) +- [vLLM-Omni model executor — upstream warmup request identification rules](rules-warmup-ids.md) diff --git a/knowledge/repos/vllm-omni/components/model-executor/rules-warmup-ids.md b/knowledge/repos/vllm-omni/components/model-executor/rules-warmup-ids.md new file mode 100644 index 00000000..afdaf0fe --- /dev/null +++ b/knowledge/repos/vllm-omni/components/model-executor/rules-warmup-ids.md @@ -0,0 +1,19 @@ +--- +title: "vLLM-Omni model executor — upstream warmup request identification rules" +created: 2026-10-23 +updated: 2026-10-23 +type: rule +tags: [vllm-omni, components, model-executor] +sources: ["PR #9264"] +--- + +# vLLM-Omni model executor — upstream warmup request identification rules + +## VLLM-OMNI-PR9264-MIXED-WARMUP-IDS — upstream warmup 请求识别必须集中于共享 helper 并覆盖全部 ID 前缀 + +- 触发：在 MRv2 runner、`OmniModelState` 或各模型 state（Fish-Omni、Higgs-Audio-v3、Moss-TTS local）中新增或修改对 upstream warmup request 的识别、text validation 豁免、preprocess/postprocess 跳过、validity 声明或 data-plane 注册逻辑。 +- 强制：识别 warmup request 只能使用 `vllm_omni/worker_v2/warmup.py` 的共享 `is_warmup_request`，它同时覆盖 legacy `_warmup_` 前缀与 vLLM 0.32 起的 `_v2_mixed_warmup_decode_`/`_v2_mixed_warmup_prefill_` 前缀；所有模型 state 与 runner 引用同一 helper。被识别的 warmup 请求走既有 warmup 处理：text 行零 codec（Fish）、保留 upstream generic text sampling/grammar/penalty 路径（Higgs）、跳过模型 preprocess/postprocess 与状态更新、仅 warmup 请求写入 fabricated validity key，且不进入 native data-plane 注册。 +- 禁止：在任何模型 state 或 runner 里散落硬编码 `_warmup_` 单前缀判断——vLLM 0.32 的 mixed-attention warmup 会被当作真实 TTS 请求处理，因缺少 text conditioning 使 Qwen3-TTS 在 FlashInfer warmup 阶段启动失败；也不得只为单个模型复制前缀列表而不回迁共享 helper。 +- 验收：识别/validity/preprocess-skip/data-plane 测试按 `_warmup_0_`、`_v2_mixed_warmup_decode_`、`_v2_mixed_warmup_prefill_` 三种 ID 参数化，并以真实请求 ID 作反例；warmup 路径断言 `preprocess`/`postprocess` 未被调用；升级 vLLM 后跑 Qwen3-TTS 两阶段 FlashInfer 启动并验证 `/health` 返回 200。^[PR #9264] + +