docker-compose-vllm.yml

version: '3'

services:
  h2ogpt:
    build:
      context: .
      dockerfile: Dockerfile
    restart: always
    shm_size: '2gb'
    depends_on:
      vllm:
        condition: service_healthy
    ports:
      - '${H2OGPT_PORT}:7860'
    volumes:
      - cache:/workspace/.cache
      - save:/workspace/save
    networks:
      - h2ogpt
    command:
      - /workspace/generate.py
      - --inference_server="vllm:vllm:5000"
      - --base_model=${H2OGPT_BASE_MODEL}
      - --langchain_mode=UserData
    deploy:
      resources:
        reservations:
          devices:
          - driver: nvidia
            device_ids: ['2', '3']
            capabilities: [gpu]

  vllm:
    build:
      context: .
      dockerfile: Dockerfile
    restart: always
    shm_size: '64gb'
    expose:
      - 5000
    volumes:
      - cache:/workspace/.cache
    networks:
      - h2ogpt
    entrypoint: /h2ogpt_conda/vllm_env/bin/python3.10
    command: -m vllm.entrypoints.openai.api_server --port=5000 --host=0.0.0.0 ${H2OGPT_VLLM_ARGS}
    environment:
      - NCCL_IGNORE_DISABLED_P2P=1
    healthcheck:
      test: [ "CMD", "curl", "-f", "http://0.0.0.0:5000/v1/models" ]
      interval: 30s
      timeout: 5s
      retries: 20
    deploy:
      resources:
        reservations:
          devices:
          - driver: nvidia
            device_ids: ['0', '1']
            capabilities: [gpu]

volumes:
  cache:
  save:
networks:
  h2ogpt: