transcription-api/docker-compose.yml

services:
  # Main service with GPU support (if available)
  # For CPU-only: Use --profile cpu or set CUDA_VISIBLE_DEVICES=""
  transcription-api:
    build: 
      context: .
      dockerfile: ${DOCKERFILE:-Dockerfile}  # Can use Dockerfile.pytorch for faster builds
    container_name: transcription-api
    environment:
      # Model configuration
      - MODEL_PATH=${MODEL_PATH:-large-v3}  # Options: tiny, base, small, medium, large, large-v2, large-v3
      
      # Model cache paths (shared with main project)
      - HF_HOME=/app/models
      - TORCH_HOME=/app/models
      - TRANSFORMERS_CACHE=/app/models
      
      # Server ports
      - GRPC_PORT=50051
      - WEBSOCKET_PORT=8765
      - ENABLE_WEBSOCKET=true
      
      # Performance tuning
      - OMP_NUM_THREADS=4
      - CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}  # GPU 0 by default
      
    ports:
      - "50051:50051"  # gRPC port
      - "8765:8765"    # WebSocket port
    
    volumes:
      # Model cache - prevents re-downloading models
      - whisper-models:/app/models
      
      # Optional: Mount SimulStreaming if available
      # - ./SimulStreaming:/app/SimulStreaming
    
    restart: unless-stopped
    
    # Resource limits (adjust based on your system)
    deploy:
      resources:
        limits:
          cpus: '4'
          memory: 8G
        reservations:
          cpus: '2'
          memory: 4G
          # GPU support (requires nvidia-docker2 or Docker 19.03+)
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    
    # Health check
    healthcheck:
      test: ["CMD", "python", "-c", "import grpc; channel = grpc.insecure_channel('localhost:50051'); channel.channel_ready()"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s

  # CPU-only service (for systems without GPU)
  transcription-api-cpu:
    profiles: ["cpu"]  # Only start with --profile cpu
    build: 
      context: .
      dockerfile: ${DOCKERFILE:-Dockerfile.pytorch}
    container_name: transcription-api-cpu
    environment:
      - MODEL_PATH=${MODEL_PATH:-base}  # Smaller model for CPU
      - HF_HOME=/app/models
      - TORCH_HOME=/app/models
      - TRANSFORMERS_CACHE=/app/models
      - GRPC_PORT=50051
      - WEBSOCKET_PORT=8765
      - ENABLE_WEBSOCKET=true
      - CUDA_VISIBLE_DEVICES=  # No GPU
    ports:
      - "50051:50051"
      - "8765:8765"
    volumes:
      - whisper-models:/app/models
    deploy:
      resources:
        limits:
          cpus: '4'
          memory: 8G
        reservations:
          cpus: '2'
          memory: 4G
          # No GPU devices for CPU profile
    healthcheck:
      test: ["CMD", "python", "-c", "import grpc; channel = grpc.insecure_channel('localhost:50051'); channel.channel_ready()"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s

volumes:
  # Share models with the main transcription project
  # This references the volume from the parent project
  whisper-models:
    external: true
    name: real-time-transcriptions_whisper-models