knowledge(vllm-omni): 2 change(s) from intake

Author: tzhouam

knowledge/repos/vllm-omni/components/serving/rules.md

@@ -1,10 +1,10 @@
 ---
 title: "Serving 规则"
 created: 2026-07-20
-updated: 2026-09-05
+updated: 2026-10-23
 type: rule
 tags: [vllm-omni, components, serving]
-sources: ["Issue #5369", "PR #3576", "PR #4583", "PR #4718", "PR #4834", "PR #4905", "PR #4912", "PR #5085", "PR #5157", "PR #5374", "PR #5670", "PR #5682", "PR #5713", "PR #5732", "PR #5746", "PR #5752", "PR #5843", "PR #5957", "PR #6008", "PR #6138", "PR #6202", "Issue #5811", "PR #6150", "claude-workflow-starter-private@09dca46", "zuiho-kai/claude-workflow-starter@c217fc6", .pre-commit-config.yaml, vllm_omni/entrypoints/async_omni.py, vllm_omni/entrypoints/omni_base.py, vllm_omni/entrypoints/openai/api_server.py, vllm_omni/entrypoints/openai/diffusion_request_utils.py, vllm_omni/entrypoints/openai/serving_chat.py, vllm_omni/entrypoints/openai/serving_speech.py, vllm_omni/entrypoints/openai/serving_video.py, vllm_omni/entrypoints/openai/video_api_utils.py, vllm_omni/entrypoints/openai/tts_adapters/, vllm_omni/engine/async_omni_engine.py, vllm_omni/engine/orchestrator.py, vllm_omni/engine/stage_pool.py, vllm_omni/engine/cfg_companion_tracker.py, vllm_omni/metrics/prometheus.py, tests/dfx/reliability/test_reliability_qwen3_omni.py, tests/engine/test_orchestrator_error_handling.py, tests/entrypoints/test_async_omni.py, tests/entrypoints/test_omni_entrypoints.py, tests/entrypoints/openai_api/test_api_server_guards.py, tests/entrypoints/openai_api/test_audex_serving_guards.py, tests/entrypoints/openai_api/test_omni_sleep_wakeup.py, tests/entrypoints/openai_api/test_serving_speech.py, tests/entrypoints/openai_api/test_tts_detection.py, tests/entrypoints/openai_api/test_video_api_utils.py, tests/entrypoints/openai_api/test_video_server.py, tests/tools/test_check_tts_adapter.py, tools/pre_commit/check_tts_adapter.py, "PR #4795", "PR #4755", "PR #3805", "PR #5878", "PR #6070", "PR #6122", "PR #4499", "PR #6050", "PR #6329", "PR #5999", "PR #5445", "PR #6288", "PR #6622", "PR #6499", "PR #6529", "PR #6776", recipes/MiniMaxAI/MiniMax-H3.md]
+sources: ["Issue #5369", "PR #3576", "PR #4583", "PR #4718", "PR #4834", "PR #4905", "PR #4912", "PR #5085", "PR #5157", "PR #5374", "PR #5670", "PR #5682", "PR #5713", "PR #5732", "PR #5746", "PR #5752", "PR #5843", "PR #5957", "PR #6008", "PR #6138", "PR #6202", "Issue #5811", "PR #6150", claude-workflow-starter-private@09dca46, zuiho-kai/claude-workflow-starter@c217fc6, .pre-commit-config.yaml, vllm_omni/entrypoints/async_omni.py, vllm_omni/entrypoints/omni_base.py, vllm_omni/entrypoints/openai/api_server.py, vllm_omni/entrypoints/openai/diffusion_request_utils.py, vllm_omni/entrypoints/openai/serving_chat.py, vllm_omni/entrypoints/openai/serving_speech.py, vllm_omni/entrypoints/openai/serving_video.py, vllm_omni/entrypoints/openai/video_api_utils.py, vllm_omni/entrypoints/openai/tts_adapters/, vllm_omni/engine/async_omni_engine.py, vllm_omni/engine/orchestrator.py, vllm_omni/engine/stage_pool.py, vllm_omni/engine/cfg_companion_tracker.py, vllm_omni/metrics/prometheus.py, tests/dfx/reliability/test_reliability_qwen3_omni.py, tests/engine/test_orchestrator_error_handling.py, tests/entrypoints/test_async_omni.py, tests/entrypoints/test_omni_entrypoints.py, tests/entrypoints/openai_api/test_api_server_guards.py, tests/entrypoints/openai_api/test_audex_serving_guards.py, tests/entrypoints/openai_api/test_omni_sleep_wakeup.py, tests/entrypoints/openai_api/test_serving_speech.py, tests/entrypoints/openai_api/test_tts_detection.py, tests/entrypoints/openai_api/test_video_api_utils.py, tests/entrypoints/openai_api/test_video_server.py, tests/tools/test_check_tts_adapter.py, tools/pre_commit/check_tts_adapter.py, "PR #4795", "PR #4755", "PR #3805", "PR #5878", "PR #6070", "PR #6122", "PR #4499", "PR #6050", "PR #6329", "PR #5999", "PR #5445", "PR #6288", "PR #6622", "PR #6499", "PR #6529", "PR #6776", recipes/MiniMaxAI/MiniMax-H3.md, "PR #6824", "PR #9282"]
 confidence: high
 ---
 
@@ -237,3 +237,12 @@ confidence: high
 - 强制：INFO 级别只记录 request ID、模型类型、voice clone、reference image 和参数等元数据；用户输入仅在 `self.request_logger` 存在时通过 `logger.debug` 输出，并按 `max_log_len` 预留日志前缀长度进行截断；未配置上限时使用 200 字符默认上限。
 - 禁止：在 INFO 日志中写入原始 `text` 或 `prompt`；无条件输出 DEBUG 用户内容；使用固定预览长度替代请求日志开关；在未配置 `max_log_len` 时输出无界用户输入。
 - 验收：覆盖 TTS、diffusion TTS、audio generation 和 diffusion chat，断言 INFO 不含用户内容；启用 `--enable-log-requests` 时 DEBUG 才输出内容，且配置与未配置 `max_log_len` 时均遵守截断上限。 ^[PR #6329]
+
+## VLLM-OMNI-PR9282-DUPLEX-APPEND-GATE — 全双工响应门只统计未完成的 response-bound append 任务
+
+- 触发：修改 `vllm_omni/engine/duplex/session/context.py` 中 append 任务的跟踪与清理、`has_response_bound_append_tasks` / `cancel_append_tasks`，或 barge-in/VAD 之后新回合的响应门判定。
+- 强制：响应门只统计 `response_bound` 且尚未 `task.done()` 的 append 任务；已运行完成但仍留在 `append_tasks` 里等待 done callback 清理的任务，不得在 barge-in 取消前一响应后继续关闭下一语音回合的响应门。活动响应与未播放输出仍由各自独立的门控制。
+- 禁止：仅凭任务仍出现在 `append_tasks` 中（不检查 `task.done()`）就让响应门保持关闭；不得用 sleep、时序假设或削减参数化来掩盖该窗口。
+- 验收：`tests/engine/duplex/test_session_manager.py` 必须覆盖真实 asyncio 任务已完成但 done callback 尚未运行的窗口，并对 response_bound 与非 response_bound 两种情况断言门状态；运行 duplex 相关 core_model CPU 测试（`tests/engine/duplex` 与 `tests/model_executor/models/qwen3_omni/test_duplex_plugin.py`）确认多轮音频用例不再间歇性失败；`docs/design/fullduplex.md` 同步该门语义。^[PR #9282]
+
+<!-- kb:rule status=active since=v0.30.0 -->

knowledge/repos/vllm-omni/models/minimax-h3/rules-loading.md

@@ -1,10 +1,10 @@
 ---
 title: "MiniMax H3 加载规则"
 created: 2026-09-02
-updated: 2026-09-04
+updated: 2026-10-23
 type: rule
 tags: [vllm-omni, models, diffusion]
-sources: ["PR #5706", "PR #5737", "PR #5824", vllm_omni/diffusion/models/minimax_h3/encoder.py, vllm_omni/diffusion/models/minimax_h3/minimax_h3_transformer.py, vllm_omni/diffusion/models/minimax_h3/pipeline_minimax_h3.py, tests/diffusion/models/minimax_h3/test_minimax_h3_contract.py, "PR #5910", "PR #6213", "PR #6445", "PR #6486"]
+sources: ["PR #5706", "PR #5737", "PR #5824", vllm_omni/diffusion/models/minimax_h3/encoder.py, vllm_omni/diffusion/models/minimax_h3/minimax_h3_transformer.py, vllm_omni/diffusion/models/minimax_h3/pipeline_minimax_h3.py, tests/diffusion/models/minimax_h3/test_minimax_h3_contract.py, "PR #5910", "PR #6213", "PR #6445", "PR #6486", "PR #8801"]
 confidence: high
 ---
 
@@ -72,3 +72,12 @@ H3 的该 contract 当前只为 BF16 no-AllGather DLO final-layout consumer 接
 区分 TP rank/size 与 SP layout，等价 DP replicas 才共享；TP/SP/layout/model revision 改变时必须以
 `preferred` 重新产生 artifact，`required` 只消费 exact hit。它不扩大 direct checkpoint mmap、
 online quantization 或 AllGather 的支持范围。^[PR #6486]
+
+## VLLM-OMNI-PR8801-H3-NATIVE-PARTITION — 本地 MiniMax-H3 判定必须让 native 分区布局优先于 modular 标记文件
+
+- 触发：修改 `vllm_omni/model_executor/models/minimax_h3/checkpoint.py` 的 `is_minimax_h3_modular`、`_NATIVE_PARTITIONS`、`resolve_minimax_h3_model_root`/`resolve_minimax_h3_diffusion_model_path`，或本地 checkpoint 的 native/FastH3 modular 判定顺序。
+- 强制：本地目录必须先按 native 分区判定——`FL2VA/model_index.json` 或 `Ref2VA/model_index.json` 存在即判 native（返回 `False`），之后才检查 `modular_model_index.json`/`fastvideo_inference.json` 标记。依据是官方 `MiniMaxAI/MiniMax-H3` 仓库会在根目录同时携带 `modular_model_index.json` 与 `FL2VA/`/`Ref2VA/` 分区，而 FastH3 modular release 没有分区；误判会让 pipeline init 因找不到 `fastvideo_inference.json` 抛 `FileNotFoundError`。native root 下 encoder resolver 必须返回 `FL2VA/text_encoder`，diffusion resolver 返回 `FL2VA`。Hub ID 走独立分支，不受该本地目录判定影响。
+- 禁止：仅凭 modular 标记文件存在就把含分区的本地快照判成 FastH3 modular；把没有 `model_index.json` 的空分区目录当作 native 证据；改动判定顺序时影响 Hub ID 分支或破坏既有 FastH3 本地/Hub 检测语义。
+- 验收：`tests/diffusion/models/minimax_h3/test_fasth3_checkpoint.py` 必须同时覆盖混合根 case（`FL2VA/`+`Ref2VA/`+`modular_model_index.json` 判为 native，且两个 resolver 返回分区路径）与对照 case（仅空分区目录加 modular 标记仍判为 modular），并保持既有 FastH3 本地与 Hub 检测测试全部通过。^[PR #8801]
+
+<!-- kb:rule status=active since=v0.30.0 -->