spark dgx部署qwen3-tts qwen3-asr语音识别和语音合成模型

经过部署测试，语音识别基本可以实时合成，语音合成速度比较慢，无法做到实时性合成。

# Qwen3 audio inference stack: TTS :8082, ASR :8001
# Switch TTS model: ./switch-tts-model.sh 0.6   or   ./switch-tts-model.sh 1.7
# Models are mounted read-only; voice/config data lives under ./volumes.
services:
  claves-qwen3-tts:
    container_name: claves-qwen3-tts
    image: ${claves_DOCKER:-}martinb78/faster-qwen3-tts-dgx-spark:latest-streaming
    restart: always
    init: true
    ports:
      - "8082:8000"
    volumes:
      - ./volumes/claves-modelscope/models/Qwen/Qwen3-TTS-12Hz-1.7B-Base:/models/Qwen3-TTS-1.7B:ro
      - ./volumes/claves-modelscope/models/Qwen/Qwen3-TTS-12Hz-0.6B-Base:/models/Qwen3-TTS-0.6B:ro
      - ./volumes/claves-qwen3-tts/config:/config:rw
    command:
      - python
      - examples/openai_server.py
      - --model
      - ${QWEN_TTS_MODEL:-/models/Qwen3-TTS-1.7B}
      - --voices
      - /config/voices.json
      - --port
      - "8000"
      - --max-seq-len
      - "2048"
    deploy:
      resources:
        limits:
          cpus: "4"
          memory: 16G
        reservations:
          memory: 1G
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    healthcheck:
      test:
        - CMD
        - python3
        - -c
        - "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=3)"
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 90s
    logging:
      driver: json-file
      options:
        max-size: "50m"
        max-file: "5"

  claves-qwen3-asr:
    container_name: claves-qwen3-asr
    # Upstream vLLM omits the audio I/O dependencies; install pinned versions
    # before starting the OpenAI-compatible transcription server.
    image: ${claves_DOCKER:-}vllm/vllm-openai:v0.29.0
    restart: always
    init: true
    shm_size: "2gb"
    ports:
      - "8001:8000"
    volumes:
      - ./volumes/claves-modelscope/models/Qwen/Qwen3-ASR-1.7B:/models/Qwen3-ASR-1.7B:ro
    entrypoint: ["/bin/bash", "-lc"]
    command:
      - >-
        python3 -m pip install --no-cache-dir -q
        --index-url ${PIP_INDEX_URL:-https://pypi.tuna.tsinghua.edu.cn/simple}
        soundfile==0.14.0 av==19.0.2 soxr==1.1.0 &&
        exec vllm serve /models/Qwen3-ASR-1.7B
        --host 0.0.0.0
        --port 8000
        --gpu-memory-utilization 0.12
        --max-model-len 8192
        --max-num-seqs 4
        --max-num-batched-tokens 4096
        --enforce-eager
    deploy:
      resources:
        limits:
          cpus: "4"
          memory: 16G
        reservations:
          memory: 1G
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    healthcheck:
      test:
        - CMD
        - python3
        - -c
        - "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=3)"
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 300s
    logging:
      driver: json-file
      options:
        max-size: "50m"
        max-file: "5"