services:
  qwen-vlm:
    image: vllm/vllm-openai:latest
    command: >
      --model Qwen/Qwen3-VL-2B-Instruct-FP8
      --max-model-len 16384
      --max-num-batched-tokens 4096
      --gpu-memory-utilization 0.85
      --tensor-parallel-size 1
    environment:
      - HF_TOKEN=${HF_TOKEN:-}
      - NCCL_DEBUG=INFO
      - NCCL_IB_DISABLE=1
      - NCCL_P2P_LEVEL=SYS
      - NCCL_MIN_NCHANNELS=4
      - NCCL_TOPO_FILE=/dev/null
      - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
    shm_size: '2gb'
    ports:
      - "8000:8000"
    volumes:
      - ./models:/root/.cache/huggingface
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
      interval: 10s
      timeout: 5s
      retries: 5
      start_period: 120s
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    networks:
      - ocr-network

  streamlit-ui:
    build: ./streamlit
    ports:
      - "8501:8501"
    volumes:
      - ./data:/app/data
    depends_on:
      qwen-vlm:
        condition: service_started
    networks:
      - ocr-network

networks:
  ocr-network:
    driver: bridge