services:
  wyoming-granite-stt:
    build: .
    image: wyoming-granite-stt:latest
    container_name: wyoming-granite-stt
    restart: unless-stopped
    ports:
      - "10300:10300"
    environment:
      # Hugging Face cache directory (optional, set to use a persistent volume)
      - HF_HOME=/data/hf
      # Hugging Face token for gated models (optional)
      # - HF_TOKEN=hf_...
    volumes:
      # Persistent cache for downloaded models
      - ./data:/data
    # GPU: requires nvidia-container-toolkit installed on the host
    #   https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html
    # runtime: nvidia
    #
    # Without GPU, use --device cpu --dtype float32 below
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    command: >
      --uri tcp://0.0.0.0:10300
      --model ibm-granite/granite-speech-4.1-2b
      --device cuda
      --dtype bfloat16
      --num-beams 1
      --language en-US