docker-compose.rocm.yml

Rama master · réplica de lectura del repositorio original.

# Hearthwave — AMD ROCm deployment (build-from-source).
# Standalone: run directly, no base file needed:
#   bash setup-rocm.sh           # one-time host prep + model staging
#   docker compose -f docker-compose.rocm.yml up --build
#
# Runs the STT final pass on an AMD GPU via ROCm. The streaming pass stays on
# CPU. Requires: amdgpu kernel driver, the host 'render' group, and (for
# unsupported GPUs like the Radeon 680M / gfx1035) HSA_OVERRIDE_GFX_VERSION.
# setup-rocm.sh writes RENDER_GID / HSA_OVERRIDE_GFX_VERSION / PULSE_UID to .env.
name: hearthwave

services:
  backend:
    build:
      context: .
      dockerfile: backend/Dockerfile
      args:
        COMPUTE_BACKEND: rocm
    volumes:
      - ./data:/data
      - ./Voices:/app/Voices:ro
      - ./Models:/app/backend/Models:ro
      - /run/user/${PULSE_UID:-1000}/pulse:/run/pulse:ro
    environment:
      - PULSE_SERVER=unix:/run/pulse/native
      # gfx override so unsupported GPUs present as a supported gfx1030.
      - HSA_OVERRIDE_GFX_VERSION=${HSA_OVERRIDE_GFX_VERSION:-10.3.0}
    devices:
      - /dev/snd:/dev/snd
      - /dev/kfd:/dev/kfd      # ROCm compute device
      - /dev/dri:/dev/dri      # GPU render nodes
    group_add:
      - "${RENDER_GID:-992}"   # host 'render' group GID (getent group render)
    restart: unless-stopped
    networks:
      - radio
    healthcheck:
      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8765/health')"]
      interval: 10s
      timeout: 5s
      retries: 5
      start_period: 20s

  frontend:
    build:
      context: ./frontend
      dockerfile: Dockerfile
    ports:
      - "80:8080"
    volumes:
      - ./data:/data:ro
    depends_on:
      backend:
        condition: service_started
    restart: unless-stopped
    networks:
      - radio

networks:
  radio:
    driver: bridge