services:
  qwen-vllm:
    build:
      context: .
      dockerfile: Dockerfile
    container_name: qwen3-vl-4b-instruct
    restart: unless-stopped
    ports:
      - "8000:8000"  # vLLM API
      - "8501:8501"  # Streamlit UI
    environment:
      - HF_HOME=/root/.cache/huggingface
      - VLLM_WORKER_MULTIPROC_METHOD=spawn
      # Optional: Add HuggingFace token if model requires authentication
      # - HF_TOKEN=your_token_here
    volumes:
      # Mount HuggingFace cache to persist model
      - huggingface_cache:/root/.cache/huggingface
      # Mount custom configs if needed
      - ./config:/app/config:ro
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    runtime: nvidia
    shm_size: '16gb'  # Shared memory for PyTorch dataloader workers
    ipc: host  # Host IPC for better performance
    stdin_open: true
    tty: true

volumes:
  huggingface_cache:
    driver: local

networks:
  default:
    name: qwen-vllm-network