feat: serve MOSS diarization across deployment stack

Signed-off-by: LauraGPT <[email protected]>

.github/workflows/test-moss-adapter.yml

@@ -5,7 +5,13 @@ on:
     paths:
       - "funasr/models/moss_transcribe_diarize/**"
       - "funasr/auto/auto_model.py"
+      - "funasr/bin/_server_app.py"
+      - "funasr/bin/server.py"
+      - "examples/openai_api/**"
+      - "README*.md"
+      - "docs/index.rst"
       - "docs/moss_transcribe_diarize*.md"
+      - "web-pages/product-site/data/deployments.json"
       - "tests/test_moss_transcribe_diarize_*.py"
       - ".github/workflows/test-moss-adapter.yml"
   push:
@@ -14,7 +20,13 @@ on:
     paths:
       - "funasr/models/moss_transcribe_diarize/**"
       - "funasr/auto/auto_model.py"
+      - "funasr/bin/_server_app.py"
+      - "funasr/bin/server.py"
+      - "examples/openai_api/**"
+      - "README*.md"
+      - "docs/index.rst"
       - "docs/moss_transcribe_diarize*.md"
+      - "web-pages/product-site/data/deployments.json"
       - "tests/test_moss_transcribe_diarize_*.py"
       - ".github/workflows/test-moss-adapter.yml"
 
@@ -45,9 +57,14 @@ jobs:
             funasr/models/moss_transcribe_diarize/model.py \
             tests/test_moss_transcribe_diarize_model.py \
             tests/test_moss_transcribe_diarize_docs.py
-          python -m py_compile funasr/models/moss_transcribe_diarize/model.py
+          python -m py_compile \
+            funasr/models/moss_transcribe_diarize/model.py \
+            funasr/bin/_server_app.py \
+            funasr/bin/server.py \
+            examples/openai_api/server.py
       - name: Run adapter and documentation contracts
         run: |
           python -m pytest -q \
             tests/test_moss_transcribe_diarize_model.py \
-            tests/test_moss_transcribe_diarize_docs.py
+            tests/test_moss_transcribe_diarize_docs.py \
+            tests/test_server_app_openai_segments.py

README.md

@@ -251,8 +251,12 @@ pip install torch torchaudio
 pip install funasr vllm fastapi uvicorn python-multipart
 funasr-server --device cuda
 # → POST /v1/audio/transcriptions at localhost:8000
+# Joint long-form ASR + anonymous speaker labels (offline HTTP):
+funasr-server --model moss-transcribe-diarize --device cuda:0
 ```
 
+[MOSS service, Docker, Kubernetes, vLLM, SGLang, LocalAI, and FunClip guide →](./docs/moss_transcribe_diarize.md)
+
 Verify it with a public sample:
 
 ```bash

README_ja.md

@@ -140,11 +140,15 @@ pip install funasr
 # OpenAI互換API（推奨）
 pip install funasr fastapi uvicorn python-multipart
 funasr-server --device cuda
+# オフライン長時間音声 ASR + 匿名 speaker label:
+funasr-server --model moss-transcribe-diarize --device cuda:0
 
 # Dockerストリーミングサービス
 docker pull registry.cn-hangzhou.aliyuncs.com/funasr_repo/funasr:funasr-runtime-sdk-online-cpu-0.1.12
 ```
 
+[MOSS service / Docker / Kubernetes / vLLM / SGLang / LocalAI / FunClip guide →](./docs/moss_transcribe_diarize.md)
+
 CPU/エッジで Python なしのオフライン ASR が必要な場合は、llama.cpp / GGUF ランタイムを使えます：[funasr.com/deploy/llama-cpp](https://www.funasr.com/en/deploy/llama-cpp.html) · [Fun-ASR-Nano-GGUF](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-GGUF) · [SenseVoiceSmall-GGUF](https://huggingface.co/FunAudioLLM/SenseVoiceSmall-GGUF)。
 
 [Colab quickstart →](./examples/colab/README_ja.md) · [OpenAI API example →](./examples/openai_api/README_ja.md) · [Client recipes →](./examples/openai_api/CLIENTS.md) · [Workflow recipes →](./examples/openai_api/WORKFLOWS.md) · [Postman collection →](./examples/openai_api/POSTMAN.md) · [OpenAPI spec →](./examples/openai_api/OPENAPI.md) · [Security guide →](./examples/openai_api/SECURITY.md) · [Deployment matrix →](./docs/deployment_matrix_ja.md) · [デプロイドキュメント →](./runtime/readme.md) · [Agent連携 →](https://modelscope.github.io/FunASR/agent.html)

README_ko.md

@@ -140,11 +140,15 @@ pip install funasr
 # OpenAI 호환 API (권장)
 pip install funasr fastapi uvicorn python-multipart
 funasr-server --device cuda
+# 오프라인 장시간 ASR + 익명 speaker label:
+funasr-server --model moss-transcribe-diarize --device cuda:0
 
 # Docker 스트리밍 서비스
 docker pull registry.cn-hangzhou.aliyuncs.com/funasr_repo/funasr:funasr-runtime-sdk-online-cpu-0.1.12
 ```
 
+[MOSS service / Docker / Kubernetes / vLLM / SGLang / LocalAI / FunClip guide →](./docs/moss_transcribe_diarize.md)
+
 CPU/엣지에서 Python 없이 오프라인 ASR만 필요하다면 llama.cpp / GGUF 런타임을 사용할 수 있습니다: [funasr.com/deploy/llama-cpp](https://www.funasr.com/en/deploy/llama-cpp.html) · [Fun-ASR-Nano-GGUF](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-GGUF) · [SenseVoiceSmall-GGUF](https://huggingface.co/FunAudioLLM/SenseVoiceSmall-GGUF).
 
 [Colab quickstart →](./examples/colab/README_ko.md) · [OpenAI API example →](./examples/openai_api/README_ko.md) · [Client recipes →](./examples/openai_api/CLIENTS.md) · [Workflow recipes →](./examples/openai_api/WORKFLOWS.md) · [Postman collection →](./examples/openai_api/POSTMAN.md) · [OpenAPI spec →](./examples/openai_api/OPENAPI.md) · [Security guide →](./examples/openai_api/SECURITY.md) · [Deployment matrix →](./docs/deployment_matrix_ko.md) · [배포 문서 →](./runtime/readme.md) · [Agent 연동 →](https://modelscope.github.io/FunASR/agent.html)

README_zh.md

@@ -275,8 +275,12 @@ funasr *.wav --output-format srt --output-dir ./output
 pip install funasr fastapi uvicorn python-multipart
 funasr-server --model sensevoice --device cuda
 # → POST /v1/audio/transcriptions，地址 localhost:8000
+# 离线长音频联合转写 + 匿名说话人标签：
+funasr-server --model moss-transcribe-diarize --device cuda:0
 ```
 
+[MOSS 服务、Docker、Kubernetes、vLLM、SGLang、LocalAI 与 FunClip 部署指南 →](./docs/moss_transcribe_diarize_zh.md)
+
 使用公开样例音频验证服务：
 
 ```bash

docs/index.rst

@@ -73,6 +73,8 @@ Overview
 
    ./deployment_matrix.md
    ./deployment_matrix_zh.md
+   ./moss_transcribe_diarize.md
+   ./moss_transcribe_diarize_zh.md
    ./runtime/readme.md
    ./runtime/docs/SDK_tutorial_online.md
    ./runtime/docs/SDK_tutorial.md

docs/moss_transcribe_diarize.md

@@ -95,6 +95,44 @@ clients keep the exact tagged generation in `raw_text`; in structured mode,
 `raw_text` is the cleaned `text` returned by vLLM and the authoritative speaker
 metadata is in `sentence_info`.
 
+## FunASR OpenAI-compatible service
+
+The built-in offline HTTP service exposes the same normalized result through
+`/v1/audio/transcriptions`. It loads the pinned Transformers revision and does
+not attach an external VAD or speaker model:
+
+```bash
+python -m pip install "transformers>=5.6,<6" fastapi uvicorn python-multipart
+funasr-server --model moss-transcribe-diarize --device cuda:0 --port 8000
+
+curl -fsS http://127.0.0.1:8000/v1/audio/transcriptions \
+  -F file=@meeting.wav \
+  -F model=moss-transcribe-diarize \
+  -F response_format=verbose_json
+```
+
+The response contains `text`, audio `duration`, and `segments` with `start`,
+`end`, `text`, and anonymous `speaker` labels. The request does not need
+`spk=true`; if a generic client sends it, the service still uses MOSS's native
+labels and does not start a second diarization pipeline.
+
+For a reproducible GPU container, build from the repository root with
+`examples/openai_api/docker-compose.moss.yml`. Kubernetes operators can build
+the same `funasr-moss-api:local` image and apply
+`examples/openai_api/kubernetes/funasr-moss-api.yaml`; replace the local image
+reference with the immutable digest from their registry before rollout.
+
+MOSS is an offline long-form model, so it is not exposed by FunASR's realtime
+WebSocket service. Use the HTTP endpoint for complete files. FunClip consumes
+the same `sentence_info` contract for speaker-aware subtitles and clips.
+
+The built-in service path was reproduced on one H100 80GB with Transformers
+`5.16.0.dev0` and Torch `2.11.0+cu130`. The pinned model processed the bundled
+6.000-second sample (`ea03e1f473ad1618a03da3327a545369cb8f6f06cb0f4115535e5a866167d47e`)
+through the real HTTP endpoint and returned one non-empty monotonic segment
+labelled `S01`, with `duration=6.0`. This is a functional service-contract
+smoke, not an accuracy, throughput, concurrency, or production-capacity claim.
+
 ## Choose a serving path
 
 | Path | Environment | Response contract | Use when |

docs/moss_transcribe_diarize_zh.md

@@ -85,6 +85,41 @@ speaker `segments` 直接映射到 `sentence_info`。认证可通过 `vllm_api_k
 此时 `raw_text` 保留原始标签生成；结构化模式下，`raw_text` 是 vLLM 返回的清理后
 `text`，权威说话人信息位于 `sentence_info`。
 
+## FunASR OpenAI 兼容服务
+
+内置离线 HTTP 服务通过 `/v1/audio/transcriptions` 返回同一套标准化结果。
+服务加载固定的 Transformers revision，不会外挂 VAD 或第二套说话人模型：
+
+```bash
+python -m pip install "transformers>=5.6,<6" fastapi uvicorn python-multipart
+funasr-server --model moss-transcribe-diarize --device cuda:0 --port 8000
+
+curl -fsS http://127.0.0.1:8000/v1/audio/transcriptions \
+  -F file=@meeting.wav \
+  -F model=moss-transcribe-diarize \
+  -F response_format=verbose_json
+```
+
+响应包含 `text`、音频 `duration`，以及带 `start`、`end`、`text` 和匿名
+`speaker` 标签的 `segments`。请求不需要设置 `spk=true`；即使通用客户端传入
+该字段，服务也只使用 MOSS 原生标签，不会启动第二套说话人流水线。
+
+GPU 容器可在仓库根目录使用
+`examples/openai_api/docker-compose.moss.yml` 构建。Kubernetes 运维可构建同一
+`funasr-moss-api:local` 镜像，再应用
+`examples/openai_api/kubernetes/funasr-moss-api.yaml`；上线前应将本地镜像名
+替换为内部镜像仓库的不可变 digest。
+
+MOSS 是离线长音频模型，不接入 FunASR 的实时 WebSocket 服务；完整文件请使用
+HTTP 接口。FunClip 通过同一套 `sentence_info` 契约生成带说话人信息的字幕与
+片段。
+
+内置服务路径已在一张 H100 80GB、Transformers `5.16.0.dev0`、Torch
+`2.11.0+cu130` 环境复现。固定 revision 的模型通过真实 HTTP 接口处理仓库自带
+6.000 秒样例（`ea03e1f473ad1618a03da3327a545369cb8f6f06cb0f4115535e5a866167d47e`），
+返回一个非空、时间单调且标记为 `S01` 的 segment，`duration=6.0`。这只是服务
+契约 smoke test，不代表准确率、吞吐、并发或生产容量。
+
 ## 选择服务后端
 
 | 路径 | 环境 | 响应契约 | 适用场景 |

examples/openai_api/CLIENTS.md

@@ -20,6 +20,10 @@ If the server is on another machine, replace `localhost` with the reachable host
 | `paraformer` | Mandarin production transcription | Includes VAD and punctuation. |
 | `paraformer-en` | English transcription | Smaller English-only route. |
 | `fun-asr-nano` | LLM-based ASR experiments | Pair with vLLM for higher throughput deployments. |
+| `moss-transcribe-diarize` | Offline multi-speaker long-form audio | Returns timestamps and anonymous speaker labels in `verbose_json`; do not add external VAD or diarization. |
+
+For MOSS server, Docker, Kubernetes, vLLM, SGLang Omni, LocalAI, and FunClip
+paths, use the [dedicated deployment guide](../../docs/moss_transcribe_diarize.md).
 
 ## Python OpenAI SDK
 

examples/openai_api/Dockerfile.moss

@@ -0,0 +1,27 @@
+FROM pytorch/pytorch:2.9.1-cuda12.8-cudnn9-runtime
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1 \
+    FUNASR_MODEL=moss-transcribe-diarize \
+    FUNASR_DEVICE=cuda:0 \
+    FUNASR_PORT=8000
+
+WORKDIR /opt/funasr
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends ffmpeg git libsndfile1 \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY . /opt/funasr
+
+RUN python -m pip install --upgrade pip \
+    && python -m pip install . fastapi "uvicorn[standard]" python-multipart \
+    && python -m pip install "transformers>=5.6,<6"
+
+EXPOSE 8000
+
+HEALTHCHECK --interval=30s --timeout=5s --start-period=10m --retries=3 \
+    CMD python -c "import os, urllib.request; urllib.request.urlopen('http://127.0.0.1:%s/health' % os.getenv('FUNASR_PORT', '8000'), timeout=3).read()" || exit 1
+
+CMD ["sh", "-c", "funasr-server --host 0.0.0.0 --port ${FUNASR_PORT:-8000} --device ${FUNASR_DEVICE:-cuda:0} --model ${FUNASR_MODEL:-moss-transcribe-diarize}"]

examples/openai_api/GRADIO.md

@@ -51,6 +51,10 @@ For OpenAI SDK clients, remember that SDK base URLs include `/v1`; this Gradio d
 | `paraformer` | Mandarin-oriented production transcription. |
 | `paraformer-en` | English-only compatibility checks. |
 | `fun-asr-nano` | LLM-based ASR and vLLM experiments. |
+| `moss-transcribe-diarize` | Offline long-form transcription with timestamps and anonymous speaker labels; select `verbose_json`. |
+
+MOSS requires the dedicated GPU service environment and is not a realtime
+microphone model. See the [MOSS deployment guide](../../docs/moss_transcribe_diarize.md).
 
 See the [model selection guide](../../docs/model_selection.md) for a deeper comparison.
 

examples/openai_api/GRADIO_zh.md

@@ -51,6 +51,10 @@ OpenAI SDK 的 base URL 需要包含 `/v1`；这个 Gradio demo 使用的是不
 | `paraformer` | 中文生产转写。 |
 | `paraformer-en` | 英文兼容性检查。 |
 | `fun-asr-nano` | LLM-based ASR 和 vLLM 实验。 |
+| `moss-transcribe-diarize` | 离线长音频转写、时间戳和匿名说话人标签；请选择 `verbose_json`。 |
+
+MOSS 需要专用 GPU 服务环境，不适合作为实时麦克风模型。详见
+[MOSS 部署指南](../../docs/moss_transcribe_diarize_zh.md)。
 
 更完整的对比见 [模型选择指南](../../docs/model_selection_zh.md)。
 

examples/openai_api/JAVASCRIPT.md

@@ -11,6 +11,11 @@ cd examples/openai_api
 python server.py --model sensevoice --device cuda --port 8000
 ```
 
+For offline multi-speaker files, start the isolated MOSS service with
+`funasr-server --model moss-transcribe-diarize --device cuda:0`, set
+`FUNASR_MODEL=moss-transcribe-diarize`, and request `verbose_json`. See the
+[MOSS deployment guide](../../docs/moss_transcribe_diarize.md).
+
 Then verify the service from another terminal:
 
 ```bash

examples/openai_api/JAVASCRIPT_zh.md

@@ -11,6 +11,11 @@ cd examples/openai_api
 python server.py --model sensevoice --device cuda --port 8000
 ```
 
+处理离线多人音频时，使用
+`funasr-server --model moss-transcribe-diarize --device cuda:0` 启动独立服务，设置
+`FUNASR_MODEL=moss-transcribe-diarize` 并请求 `verbose_json`。详见
+[MOSS 部署指南](../../docs/moss_transcribe_diarize_zh.md)。
+
 在另一个终端验证服务：
 
 ```bash

examples/openai_api/OPENAPI.md

@@ -38,7 +38,7 @@ Replace them with the URL reachable from your app, container, or workflow runtim
 | Field | Type | Required | Notes |
 |---|---|---|---|
 | `file` | binary | yes | Audio file such as wav, mp3, flac, m4a, ogg, or webm. |
-| `model` | string | no | Defaults to `sensevoice`; available aliases are listed by `/v1/models`. |
+| `model` | string | no | Defaults to `sensevoice`; `/v1/models` also lists `moss-transcribe-diarize` for offline long-form transcription with native anonymous speaker labels. |
 | `language` | string | no | Optional language hint. |
 | `response_format` | string | no | Use `json` or `verbose_json`. |
 
@@ -51,3 +51,7 @@ curl -fsS http://localhost:8000/openapi.json > /tmp/funasr-openapi-live.json
 ```
 
 The live FastAPI schema may include framework-specific validation details; this checked-in spec keeps the public integration surface small and stable.
+
+MOSS requires its isolated Transformers environment. Start it with
+`funasr-server --model moss-transcribe-diarize --device cuda:0` and request
+`response_format=verbose_json`; see the [deployment guide](../../docs/moss_transcribe_diarize.md).

examples/openai_api/OPENAPI_zh.md

@@ -38,7 +38,7 @@
 | Field | Type | Required | 说明 |
 |---|---|---|---|
 | `file` | binary | yes | 音频文件，例如 wav、mp3、flac、m4a、ogg 或 webm。 |
-| `model` | string | no | 默认 `sensevoice`；可用别名由 `/v1/models` 返回。 |
+| `model` | string | no | 默认 `sensevoice`；`/v1/models` 也会列出用于离线长音频与原生匿名说话人标签的 `moss-transcribe-diarize`。 |
 | `language` | string | no | 可选语言提示。 |
 | `response_format` | string | no | 使用 `json` 或 `verbose_json`。 |
 
@@ -51,3 +51,7 @@ curl -fsS http://localhost:8000/openapi.json > /tmp/funasr-openapi-live.json
 ```
 
 实时 FastAPI schema 可能包含框架级校验细节；仓库中的静态规范保留更小、更稳定的公开集成面。
+
+MOSS 需要独立 Transformers 环境。使用
+`funasr-server --model moss-transcribe-diarize --device cuda:0` 启动，并请求
+`response_format=verbose_json`；详见 [部署指南](../../docs/moss_transcribe_diarize_zh.md)。

examples/openai_api/README.md

@@ -92,6 +92,12 @@ curl http://localhost:8000/v1/audio/transcriptions \
 | `paraformer` | 120x realtime | 15x realtime | zh/en | Punctuation |
 | `paraformer-en` | 120x realtime | 15x realtime | en | English only |
 | `fun-asr-nano` | 17x realtime | 3.6x realtime | zh/en/ja + Chinese dialects/accents | LLM-based, timestamps |
+| `moss-transcribe-diarize` | Validate on target GPU | Offline only | Multilingual | Long-form timestamps + anonymous speaker labels |
+
+MOSS uses a pinned third-party HF revision and must not be combined with an
+external VAD or speaker model. See the [complete MOSS deployment guide](../../docs/moss_transcribe_diarize.md)
+for `funasr-server`, Docker Compose, Kubernetes, vLLM, SGLang Omni, LocalAI,
+and FunClip paths.
 
 ## API Endpoints
 

examples/openai_api/README_ja.md

@@ -89,6 +89,11 @@ curl http://localhost:8000/v1/audio/transcriptions \
 | `paraformer` | 120x realtime | 15x realtime | zh/en | 句読点復元 |
 | `paraformer-en` | 120x realtime | 15x realtime | en | 英語認識 |
 | `fun-asr-nano` | 17x realtime | 3.6x realtime | 中英日 + 中国語方言・地域アクセント | LLM-based、タイムスタンプ |
+| `moss-transcribe-diarize` | 対象 GPU で要検証 | オフラインのみ | 多言語 | 長時間 timestamp + 匿名 speaker label |
+
+MOSS は固定 revision の third-party HF model を使用し、外部 VAD / speaker model
+とは併用しません。[MOSS deployment guide](../../docs/moss_transcribe_diarize.md) に
+`funasr-server`、Docker、Kubernetes、vLLM、SGLang、LocalAI、FunClip をまとめています。
 
 ## API エンドポイント
 

examples/openai_api/README_ko.md

@@ -89,6 +89,11 @@ curl http://localhost:8000/v1/audio/transcriptions \
 | `paraformer` | 120x realtime | 15x realtime | zh/en | 문장부호 복원 |
 | `paraformer-en` | 120x realtime | 15x realtime | en | 영어 인식 |
 | `fun-asr-nano` | 17x realtime | 3.6x realtime | 중영일 + 중국어 방언/지역 억양 | LLM-based, 타임스탬프 |
+| `moss-transcribe-diarize` | 대상 GPU 검증 필요 | 오프라인 전용 | 다국어 | 장시간 timestamp + 익명 speaker label |
+
+MOSS는 고정 revision의 third-party HF model을 사용하며 외부 VAD / speaker model과
+결합하지 않습니다. `funasr-server`, Docker, Kubernetes, vLLM, SGLang, LocalAI,
+FunClip 경로는 [MOSS deployment guide](../../docs/moss_transcribe_diarize.md)를 참고하세요.
 
 ## API 엔드포인트
 

examples/openai_api/README_zh.md

@@ -89,6 +89,11 @@ curl http://localhost:8000/v1/audio/transcriptions \
 | `paraformer` | 120x realtime | 15x realtime | zh/en | 标点恢复 |
 | `paraformer-en` | 120x realtime | 15x realtime | en | 英文识别 |
 | `fun-asr-nano` | 17x realtime | 3.6x realtime | 中/英/日 + 中文方言/口音 | LLM-based，时间戳 |
+| `moss-transcribe-diarize` | 需在目标 GPU 验证 | 仅离线 | 多语言 | 长音频时间戳 + 匿名说话人标签 |
+
+MOSS 使用固定 revision 的第三方 HF 模型，不能外挂 VAD 或说话人模型。完整的
+`funasr-server`、Docker Compose、Kubernetes、vLLM、SGLang Omni、LocalAI 与
+FunClip 路径见 [MOSS 部署指南](../../docs/moss_transcribe_diarize_zh.md)。
 
 ## API 端点
 

examples/openai_api/WORKFLOWS.md

@@ -25,6 +25,11 @@ If the workflow engine runs in Docker, `localhost` usually means the workflow co
 
 Before configuring a low-code tool, you can import the [Postman collection](POSTMAN.md) and run health, model-list, and transcription requests from a GUI. For schema-driven imports, use the [OpenAPI spec](OPENAPI.md). Set `FUNASR_BASE_URL`, choose a local audio file for the multipart `file` field, and keep `MODEL_ALIAS=sensevoice` for the first test.
 
+For offline multi-speaker meetings, set `MODEL_ALIAS=moss-transcribe-diarize`
+and keep `response_format=verbose_json` so downstream nodes receive native
+anonymous speaker segments. Use the [MOSS deployment guide](../../docs/moss_transcribe_diarize.md)
+for the isolated GPU service and file-duration limits.
+
 ## Multipart HTTP request
 
 Every workflow engine eventually needs to send this request shape:

examples/openai_api/WORKFLOWS_zh.md

@@ -25,6 +25,10 @@ curl -fsS "$FUNASR_BASE_URL/v1/models"
 
 在配置低代码工具前，可以先导入 [Postman collection](POSTMAN_zh.md)，从图形界面跑通 health、模型列表和转写请求；需要按 schema 导入时可使用 [OpenAPI spec](OPENAPI_zh.md)。设置 `FUNASR_BASE_URL`，在 multipart `file` 字段选择本地音频文件，第一次测试建议保持 `MODEL_ALIAS=sensevoice`。
 
+处理离线多人会议时，将 `MODEL_ALIAS` 改为 `moss-transcribe-diarize`，并保留
+`response_format=verbose_json`，下游节点即可收到模型原生匿名说话人 segments。
+独立 GPU 服务与文件时长边界见 [MOSS 部署指南](../../docs/moss_transcribe_diarize_zh.md)。
+
 ## Multipart HTTP 请求
 
 所有工作流引擎最终都需要发出下面这种请求：

examples/openai_api/docker-compose.moss.yml

@@ -0,0 +1,26 @@
+services:
+  funasr-moss-api:
+    build:
+      context: ../..
+      dockerfile: examples/openai_api/Dockerfile.moss
+    image: funasr-moss-api:local
+    ports:
+      - "${FUNASR_HOST_PORT:-8000}:8000"
+    environment:
+      FUNASR_PORT: "8000"
+      FUNASR_DEVICE: "${FUNASR_DEVICE:-cuda:0}"
+      FUNASR_MODEL: moss-transcribe-diarize
+      HF_HOME: /root/.cache/huggingface
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    volumes:
+      - funasr-moss-cache:/root/.cache
+    shm_size: "8gb"
+
+volumes:
+  funasr-moss-cache:

examples/openai_api/kubernetes/README.md

@@ -57,6 +57,11 @@ For in-cluster clients, use `http://funasr-api.speech.svc.cluster.local:8000` as
 | Memory request | `8Gi` | Tune after observing startup and real audio workloads. |
 | Startup probe | 10 minutes | Increase if your registry, model hub, or storage backend is slow. |
 
+The `moss-transcribe-diarize` alias uses a separate GPU image and resource profile. Build
+`examples/openai_api/Dockerfile.moss`, then apply `funasr-moss-api.yaml` after
+replacing `funasr-moss-api:local` with your registry's immutable image digest.
+The complete runtime matrix is in the [MOSS deployment guide](../../../docs/moss_transcribe_diarize.md).
+
 ## GPU notes
 
 The example Dockerfile is CPU-first. For GPU clusters you need to adapt the image to CUDA-capable PyTorch/FunASR dependencies, then add your cluster's GPU scheduling fields, for example:

examples/openai_api/kubernetes/README_zh.md

@@ -57,6 +57,11 @@ python smoke_test.py --base-url http://localhost:8000
 | 内存 request | `8Gi` | 根据启动过程和真实音频负载观测结果调整。 |
 | Startup probe | 10 分钟 | registry、模型下载或存储后端较慢时增大。 |
 
+`moss-transcribe-diarize` 使用独立的 GPU 镜像与资源配置。先构建
+`examples/openai_api/Dockerfile.moss`，将 `funasr-moss-api.yaml` 中的
+`funasr-moss-api:local` 替换为内部镜像仓库的不可变 digest 后再应用。完整运行矩阵
+见 [MOSS 部署指南](../../../docs/moss_transcribe_diarize_zh.md)。
+
 ## GPU 说明
 
 示例 Dockerfile 默认面向 CPU。GPU 集群需要先把镜像改成 CUDA-capable PyTorch/FunASR 依赖，再根据集群增加 GPU 调度字段，例如：

examples/openai_api/kubernetes/funasr-moss-api.yaml

@@ -0,0 +1,86 @@
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: funasr-moss-cache
+spec:
+  accessModes:
+    - ReadWriteOnce
+  resources:
+    requests:
+      storage: 40Gi
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: funasr-moss-api
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: funasr-moss-api
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: funasr-moss-api
+    spec:
+      containers:
+        - name: api
+          image: funasr-moss-api:local
+          imagePullPolicy: IfNotPresent
+          env:
+            - name: FUNASR_MODEL
+              value: moss-transcribe-diarize
+            - name: FUNASR_DEVICE
+              value: cuda:0
+            - name: FUNASR_PORT
+              value: "8000"
+          ports:
+            - name: http
+              containerPort: 8000
+          startupProbe:
+            httpGet:
+              path: /health
+              port: http
+            periodSeconds: 10
+            timeoutSeconds: 5
+            failureThreshold: 60
+          readinessProbe:
+            httpGet:
+              path: /health
+              port: http
+            periodSeconds: 10
+          resources:
+            requests:
+              cpu: "4"
+              memory: 24Gi
+              nvidia.com/gpu: "1"
+            limits:
+              memory: 48Gi
+              nvidia.com/gpu: "1"
+          volumeMounts:
+            - name: cache
+              mountPath: /root/.cache
+            - name: dshm
+              mountPath: /dev/shm
+      volumes:
+        - name: cache
+          persistentVolumeClaim:
+            claimName: funasr-moss-cache
+        - name: dshm
+          emptyDir:
+            medium: Memory
+            sizeLimit: 8Gi
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: funasr-moss-api
+spec:
+  selector:
+    app.kubernetes.io/name: funasr-moss-api
+  ports:
+    - name: http
+      port: 8000
+      targetPort: http

examples/openai_api/openapi.json

@@ -181,7 +181,8 @@
               "sensevoice",
               "paraformer",
               "paraformer-en",
-              "fun-asr-nano"
+              "fun-asr-nano",
+              "moss-transcribe-diarize"
             ],
             "default": "sensevoice",
             "description": "FunASR model alias."

examples/openai_api/server.py

@@ -6,6 +6,7 @@ Works with any agent framework that supports OpenAI audio API.
 
 Usage:
     python server.py --model sensevoice --device cuda --port 8000
+    python server.py --model moss-transcribe-diarize --device cuda:0 --port 8000
 
 Then use with any OpenAI-compatible client:
     curl http://localhost:8000/v1/audio/transcriptions \
@@ -54,6 +55,13 @@ MODEL_CONFIGS = {
         "vad_model": "fsmn-vad",
         "vad_kwargs": {"max_single_segment_time": 30000},
     },
+    "moss-transcribe-diarize": {
+        "model": "OpenMOSS-Team/MOSS-Transcribe-Diarize",
+        "model_revision": "e8681d68e7042738ffca8ac8212bc8fcb1131ab8",
+        "hub": "hf",
+        "backend": "hf",
+        "trust_remote_code": True,
+    },
 }
 
 
@@ -99,7 +107,7 @@ async def transcribe(
     
     Accepts the same parameters as OpenAI's /v1/audio/transcriptions:
     - file: Audio file (wav, mp3, flac, m4a, ogg, webm)
-    - model: Model to use (sensevoice, paraformer, fun-asr-nano)
+    - model: Model to use (sensevoice, paraformer, fun-asr-nano, moss-transcribe-diarize)
     - language: Optional language hint
     - response_format: json or verbose_json
     """

funasr/bin/_server_app.py

@@ -1,7 +1,8 @@
 """FunASR Server — unified vLLM-based inference service.
 
 Provides OpenAI-compatible API (/v1/audio/transcriptions) and REST API (/asr).
-Uses vLLM for Fun-ASR-Nano (GPU) or falls back to AutoModel for non-LLM models (SenseVoice/Paraformer).
+Uses vLLM for Fun-ASR-Nano (GPU) or AutoModel for supported models, including
+MOSS-Transcribe-Diarize's joint transcription and anonymous speaker labels.
 """
 
 import io
@@ -32,6 +33,8 @@ PACKAGE_VERSION = (Path(__file__).resolve().parents[1] / "version.txt").read_tex
 
 
 _LANGUAGE_TAG_RE = re.compile(r"<\|(zh|en|yue|ja|ko)\|>")
+MOSS_MODEL_REVISION = "e8681d68e7042738ffca8ac8212bc8fcb1131ab8"
+NATIVE_DIARIZATION_MODELS = {"moss-transcribe-diarize"}
 
 
 def extract_language_from_asr_text(text):
@@ -257,6 +260,13 @@ def create_app(
             "vad_model": "fsmn-vad",
             "punc_model": "ct-punc",
         },
+        "moss-transcribe-diarize": {
+            "model": "OpenMOSS-Team/MOSS-Transcribe-Diarize",
+            "model_revision": MOSS_MODEL_REVISION,
+            "hub": "hf",
+            "backend": "hf",
+            "trust_remote_code": True,
+        },
     }
 
     def _load_spk_model():
@@ -335,7 +345,7 @@ def create_app(
         if app.state.model_path:
             cfg["model"] = app.state.model_path
             cfg["hub"] = app.state.hub
-        elif app.state.hub:
+        elif app.state.hub and "hub" not in cfg:
             cfg["hub"] = app.state.hub
         cfg["device"] = device
         cfg["disable_update"] = True
@@ -434,7 +444,7 @@ def create_app(
                 segments.append(segment)
         if not segments and text:
             segments = build_openai_fallback_segments(text, duration)
-        if use_spk and segments:
+        if use_spk and model_name not in NATIVE_DIARIZATION_MODELS and segments:
             audio_data, sr = sf.read(audio_path)
             attach_speaker_labels(
                 audio_data,

funasr/bin/server.py

@@ -5,6 +5,7 @@ Usage:
     funasr-server                          # default: sensevoice on cuda:0, port 8000
     funasr-server --device cpu --port 9000
     funasr-server --model paraformer
+    funasr-server --model moss-transcribe-diarize --device cuda:0
     funasr-server --model-path /path/to/local/model
     funasr-server --model-path username/paraformer --hub hf
     funasr-server --cors-origin http://localhost:3000
@@ -31,6 +32,7 @@ Examples:
   funasr-server                          # Start with SenseVoice on GPU
   funasr-server --device cpu             # Start on CPU
   funasr-server --model paraformer       # Use Paraformer model
+  funasr-server --model moss-transcribe-diarize --device cuda:0
   funasr-server --port 9000             # Custom port
   funasr-server --model-path /path/to/local/model  # Use local model
   funasr-server --model-path username/model --hub hf  # Use HuggingFace model
@@ -45,7 +47,14 @@ Then use with OpenAI SDK:
     parser.add_argument("--host", default="0.0.0.0", help="Bind address (default: 0.0.0.0)")
     parser.add_argument("--port", type=int, default=8000, help="Port (default: 8000)")
     parser.add_argument("--device", default="cuda", help="Device: cuda, cpu, mps (default: cuda)")
-    parser.add_argument("--model", default="auto", help="Pre-load model: auto (GPU=fun-asr-nano, CPU=sensevoice), sensevoice, paraformer, fun-asr-nano")
+    parser.add_argument(
+        "--model",
+        default="auto",
+        help=(
+            "Pre-load model: auto (GPU=fun-asr-nano, CPU=sensevoice), "
+            "sensevoice, paraformer, fun-asr-nano, moss-transcribe-diarize"
+        ),
+    )
     parser.add_argument("--model-path", default=None, help="Local model path or model ID (overrides --model)")
     parser.add_argument("--hub", default="ms", help="Model hub: ms (ModelScope), hf (HuggingFace) (default: ms)")
     parser.add_argument(

tests/test_moss_transcribe_diarize_docs.py

@@ -1,3 +1,4 @@
+import json
 from pathlib import Path
 
 import pytest
@@ -78,3 +79,82 @@ def test_all_deployment_matrices_link_moss_guide(matrix: Path) -> None:
         "FunASR 소유 model 또는 `AutoModel` backend가 아닙니다",
     ):
         assert stale_claim not in moss_row
+
+
+def test_moss_guides_cover_funasr_service_and_offline_boundaries() -> None:
+    for guide in GUIDES:
+        text = guide.read_text(encoding="utf-8")
+        assert "funasr-server --model moss-transcribe-diarize" in text
+        assert "response_format=verbose_json" in text
+        assert "examples/openai_api/docker-compose.moss.yml" in text
+        assert "examples/openai_api/kubernetes/funasr-moss-api.yaml" in text
+        assert "WebSocket" in text
+        assert "FunClip" in text
+
+
+def test_moss_service_recipes_are_pinned_and_gpu_scoped() -> None:
+    dockerfile = ROOT / "examples" / "openai_api" / "Dockerfile.moss"
+    compose = ROOT / "examples" / "openai_api" / "docker-compose.moss.yml"
+    kubernetes = (
+        ROOT / "examples" / "openai_api" / "kubernetes" / "funasr-moss-api.yaml"
+    )
+
+    assert dockerfile.is_file()
+    assert compose.is_file()
+    assert kubernetes.is_file()
+    docker_text = dockerfile.read_text(encoding="utf-8")
+    compose_text = compose.read_text(encoding="utf-8")
+    kubernetes_text = kubernetes.read_text(encoding="utf-8")
+    for text in (docker_text, compose_text, kubernetes_text):
+        assert "moss-transcribe-diarize" in text
+    assert "transformers>=5.6,<6" in docker_text
+    assert "capabilities: [gpu]" in compose_text
+    assert "nvidia.com/gpu" in kubernetes_text
+
+
+def test_github_pages_indexes_moss_deployment_guide() -> None:
+    index = (ROOT / "docs" / "index.rst").read_text(encoding="utf-8")
+    assert "./moss_transcribe_diarize.md" in index
+    assert "./moss_transcribe_diarize_zh.md" in index
+
+
+def test_readmes_link_the_runnable_moss_service() -> None:
+    readmes = [
+        ROOT / "README.md",
+        ROOT / "README_zh.md",
+        ROOT / "README_ja.md",
+        ROOT / "README_ko.md",
+        ROOT / "examples" / "openai_api" / "README.md",
+        ROOT / "examples" / "openai_api" / "README_zh.md",
+        ROOT / "examples" / "openai_api" / "README_ja.md",
+        ROOT / "examples" / "openai_api" / "README_ko.md",
+    ]
+    for readme in readmes:
+        text = readme.read_text(encoding="utf-8")
+        assert "moss-transcribe-diarize" in text
+        assert "moss_transcribe_diarize" in text
+
+
+def test_openai_consumer_docs_expose_moss_alias_and_boundaries() -> None:
+    paths = [
+        "CLIENTS.md",
+        "GRADIO.md",
+        "GRADIO_zh.md",
+        "OPENAPI.md",
+        "OPENAPI_zh.md",
+        "WORKFLOWS.md",
+        "WORKFLOWS_zh.md",
+        "JAVASCRIPT.md",
+        "JAVASCRIPT_zh.md",
+        "kubernetes/README.md",
+        "kubernetes/README_zh.md",
+    ]
+    root = ROOT / "examples" / "openai_api"
+    for relative in paths:
+        text = (root / relative).read_text(encoding="utf-8")
+        assert "moss-transcribe-diarize" in text
+        assert "moss_transcribe_diarize" in text
+
+    spec = json.loads((root / "openapi.json").read_text(encoding="utf-8"))
+    model = spec["components"]["schemas"]["TranscriptionRequest"]["properties"]["model"]
+    assert "moss-transcribe-diarize" in model["enum"]

tests/test_server_app_openai_segments.py

@@ -442,6 +442,53 @@ def test_spk_request_lazily_loads_and_reuses_speaker_model(monkeypatch):
     ]
 
 
+def test_moss_service_uses_pinned_joint_transcription_config(monkeypatch):
+    module = load_server_app(monkeypatch)
+    DummyAutoModel = install_dummy_funasr(monkeypatch)
+
+    module.create_app(device="cuda:0", preload_model="moss-transcribe-diarize")
+
+    config = DummyAutoModel.instances[-1]
+    assert config["model"] == "OpenMOSS-Team/MOSS-Transcribe-Diarize"
+    assert config["model_revision"] == "e8681d68e7042738ffca8ac8212bc8fcb1131ab8"
+    assert config["hub"] == "hf"
+    assert config["backend"] == "hf"
+    assert config["trust_remote_code"] is True
+    assert "vad_model" not in config
+    assert "spk_model" not in config
+
+
+def test_moss_verbose_json_preserves_native_speaker_segments(monkeypatch):
+    module = load_server_app(monkeypatch)
+    DummyAutoModel = install_dummy_funasr(
+        monkeypatch,
+        generated_result={
+            "text": "hello again",
+            "sentence_info": [
+                {"start": 100, "end": 900, "spk": "S01", "text": "hello"},
+                {"start": 950, "end": 1700, "spk": "S02", "sentence": "again"},
+            ],
+        },
+    )
+    monkeypatch.setattr(module.sf, "info", lambda path: types.SimpleNamespace(duration=1.7))
+    app = module.create_app(device="cuda:0", preload_model="moss-transcribe-diarize")
+    transcribe = app.routes[("POST", "/v1/audio/transcriptions")]
+
+    response = asyncio.run(
+        transcribe(
+            file=DummyUpload(),
+            model="moss-transcribe-diarize",
+            language=None,
+            response_format="verbose_json",
+            spk=True,
+        )
+    )
+
+    assert [segment["speaker"] for segment in response["segments"]] == ["S01", "S02"]
+    assert [segment["text"] for segment in response["segments"]] == ["hello", "again"]
+    assert not any(config.get("model") == "cam++" for config in DummyAutoModel.instances)
+
+
 def test_server_versions_follow_package_version(monkeypatch):
     expected = (REPO_ROOT / "funasr" / "version.txt").read_text().strip()
     module = load_server_app(monkeypatch)

web-pages/product-site/data/deployments.json

@@ -1,6 +1,6 @@
 {
   "schema_version": 1,
-  "verified": "2026-08-30",
+  "verified": "2026-09-01",
   "deployments": [
     {
       "id": "moss-transcribe-diarize",
@@ -13,7 +13,7 @@
       "models": ["OpenMOSS-Team/MOSS-Transcribe-Diarize (third-party Apache-2.0 model)"],
       "operating_systems": ["Linux"],
       "interfaces": ["OpenAI-compatible HTTP", "vLLM", "SGLang Omni", "Transformers", "LocalAI / moss-transcribe.cpp"],
-      "tested": {"funasr": "AutoModel vLLM + SGLang adapters; third-party model@e8681d68", "runtime": "vLLM 0.27.1 / Torch 2.13.0+cu129 / H100 80GB", "verified": "2026-08-30"},
+      "tested": {"funasr": "AutoModel HF + vLLM + SGLang adapters and OpenAI HTTP service; third-party model@e8681d68", "runtime": "Transformers 5.16.0.dev0 + Torch 2.11.0+cu130 and vLLM 0.27.1 + Torch 2.13.0+cu129 / H100 80GB", "verified": "2026-09-01"},
       "commands": {
         "install": [
           "uv venv --python 3.12 .venv-moss && curl -fL https://github.com/vllm-project/vllm/releases/download/v0.27.1/vllm-0.27.1%2Bcu129-cp38-abi3-manylinux_2_28_x86_64.whl -o vllm-0.27.1+cu129-cp38-abi3-manylinux_2_28_x86_64.whl && echo 'bf0d52faa2a51e7a01c6856a7a8a2d1307fd0ff711415d34168a67ffac0fa47b  vllm-0.27.1+cu129-cp38-abi3-manylinux_2_28_x86_64.whl' | sha256sum -c - && uv pip install --python .venv-moss/bin/python --torch-backend=auto \"vllm[audio] @ file://$PWD/vllm-0.27.1+cu129-cp38-abi3-manylinux_2_28_x86_64.whl\"",
@@ -33,6 +33,31 @@
         ]
       },
       "runtime_paths": [
+        {
+          "id": "funasr-server",
+          "tested": "Transformers 5.16.0.dev0 / Torch 2.11.0+cu130 / H100 80GB; real HTTP response verified 2026-09-01",
+          "translations": {
+            "zh": {"name": "FunASR OpenAI 兼容服务", "summary": "直接使用内置 funasr-server 加载固定 revision 的 HF 后端，返回带匿名说话人标签的 verbose_json；可使用专用 Docker Compose 与 Kubernetes GPU 配方。"},
+            "en": {"name": "FunASR OpenAI-compatible service", "summary": "Load the pinned HF backend directly with funasr-server and return verbose_json with anonymous speaker labels; dedicated Docker Compose and Kubernetes GPU recipes are included."}
+          },
+          "commands": {
+            "install": [
+              "python -m pip install 'transformers>=5.6,<6' fastapi uvicorn python-multipart"
+            ],
+            "launch": [
+              "funasr-server --model moss-transcribe-diarize --device cuda:0 --port 8000",
+              "docker compose -f examples/openai_api/docker-compose.moss.yml up --build"
+            ],
+            "health": [
+              "curl -fsS http://127.0.0.1:8000/health",
+              "curl -fsS http://127.0.0.1:8000/v1/models"
+            ],
+            "smoke": [
+              "curl -fsS http://127.0.0.1:8000/v1/audio/transcriptions -F file=@runtime/llama.cpp/tests/sample.wav -F model=moss-transcribe-diarize -F response_format=verbose_json | tee /tmp/funasr-moss-transcription.json",
+              "python - <<'PY'\nimport json\n\nwith open('/tmp/funasr-moss-transcription.json', encoding='utf-8') as stream:\n    payload = json.load(stream)\nsegments = payload.get('segments', [])\nassert payload.get('text', '').strip() and segments, payload\nassert all(item.get('speaker') and item.get('start') <= item.get('end') for item in segments), payload\nprint(payload['text'], sorted({item['speaker'] for item in segments}))\nPY"
+            ]
+          }
+        },
         {
           "id": "vllm",
           "tested": "vLLM 0.27.1 / Torch 2.13.0+cu129 / H100 80GB; FunASR adapter verified",

web-pages/product-site/tests/test_output.py

@@ -128,7 +128,7 @@ def test_detail_commands_come_from_registry(built_site):
             assert command in rendered
 
 
-def test_moss_detail_renders_separate_vllm_and_sglang_runtime_paths(built_site):
+def test_moss_detail_renders_all_service_runtime_paths(built_site):
     for relative in (
         'deploy/moss-transcribe-diarize.html',
         'en/deploy/moss-transcribe-diarize.html',
@@ -136,8 +136,14 @@ def test_moss_detail_renders_separate_vllm_and_sglang_runtime_paths(built_site):
         soup = read_soup(built_site / relative)
         paths = soup.select('[data-runtime-path]')
 
-        assert [path['data-runtime-path'] for path in paths] == ['vllm', 'sglang-omni']
+        assert [path['data-runtime-path'] for path in paths] == [
+            'funasr-server',
+            'vllm',
+            'sglang-omni',
+        ]
         rendered = '\n'.join(path.get_text('\n') for path in paths)
+        assert 'funasr-server --model moss-transcribe-diarize' in rendered
+        assert 'docker-compose.moss.yml' in rendered
         assert 'vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize' in rendered
         assert 'sgl-omni serve' in rendered
         assert 'response_format=diarized_json' in rendered

web-pages/product-site/tests/test_registry.py

@@ -107,15 +107,24 @@ def test_moss_transcribe_diarize_contract_tracks_third_party_upstream(valid_regi
         'OpenMOSS-Team/MOSS-Transcribe-Diarize (third-party Apache-2.0 model)'
     ]
     assert entry['tested'] == {
-        'funasr': 'AutoModel vLLM + SGLang adapters; third-party model@e8681d68',
-        'runtime': 'vLLM 0.27.1 / Torch 2.13.0+cu129 / H100 80GB',
-        'verified': '2026-08-30',
+        'funasr': 'AutoModel HF + vLLM + SGLang adapters and OpenAI HTTP service; third-party model@e8681d68',
+        'runtime': 'Transformers 5.16.0.dev0 + Torch 2.11.0+cu130 and vLLM 0.27.1 + Torch 2.13.0+cu129 / H100 80GB',
+        'verified': '2026-09-01',
     }
     assert 'LocalAI / moss-transcribe.cpp' in entry['interfaces']
     assert {'cpu', 'desktop-edge-gpu'} <= set(entry['hardware'])
 
     runtime_paths = {path['id']: path for path in entry['runtime_paths']}
-    assert set(runtime_paths) == {'vllm', 'sglang-omni'}
+    assert set(runtime_paths) == {'funasr-server', 'vllm', 'sglang-omni'}
+    service_commands = '\n'.join(
+        command
+        for group in ('install', 'launch', 'health', 'smoke')
+        for command in runtime_paths['funasr-server']['commands'][group]
+    )
+    assert 'funasr-server --model moss-transcribe-diarize' in service_commands
+    assert 'docker-compose.moss.yml' in service_commands
+    assert 'response_format=verbose_json' in service_commands
+    assert 'real HTTP response verified 2026-09-01' in runtime_paths['funasr-server']['tested']
     assert runtime_paths['vllm']['tested'] == (
         'vLLM 0.27.1 / Torch 2.13.0+cu129 / H100 80GB; FunASR adapter verified'
     )
@@ -222,7 +231,7 @@ def test_audio_cpp_contract_tracks_mainline_nano_and_sensevoice(valid_registry):
     entry = next(item for item in valid_registry['deployments'] if item['id'] == 'audio-cpp')
     llama_cpp = next(item for item in valid_registry['deployments'] if item['id'] == 'llama-cpp')
 
-    assert valid_registry['verified'] == '2026-08-30'
+    assert valid_registry['verified'] == '2026-09-01'
     assert entry['maturity'] == 'community-verified'
     assert entry['selector_rank'] > llama_cpp['selector_rank']
     assert entry['tested'] == {
@@ -391,7 +400,7 @@ def test_sensevoice_native_server_contract_tracks_merged_runtime(valid_registry)
     llama_cpp = next(item for item in valid_registry['deployments'] if item['id'] == 'llama-cpp')
     audio_cpp = next(item for item in valid_registry['deployments'] if item['id'] == 'audio-cpp')
 
-    assert valid_registry['verified'] == '2026-08-30'
+    assert valid_registry['verified'] == '2026-09-01'
     assert entry['maturity'] == 'production-verified'
     assert llama_cpp['selector_rank'] < entry['selector_rank'] < audio_cpp['selector_rank']
     assert entry['tested'] == {
@@ -485,9 +494,9 @@ def test_runtime_paths_require_unique_ids_translations_and_commands(valid_regist
         item for item in data['deployments']
         if item['id'] == 'moss-transcribe-diarize'
     )
-    entry['runtime_paths'][1]['id'] = 'vllm'
-    del entry['runtime_paths'][0]['translations']['en']['summary']
-    del entry['runtime_paths'][1]['commands']['health']
+    entry['runtime_paths'][2]['id'] = 'vllm'
+    del entry['runtime_paths'][1]['translations']['en']['summary']
+    del entry['runtime_paths'][2]['commands']['health']
 
     errors = validate_registry(data)