services:
  text-generation-inference:
    image: ghcr.io/huggingface/text-generation-inference:latest
    network_mode: "host"
    volumes:
      - $HOME/.cache/huggingface:/data
    ports:
      - "5009:5009"
    devices:
      - nvidia.com/gpu=all
    environment:
      - CUDA_VISIBLE_DEVICES=0
      - HF_HOME=/data
      - HF_DATASETS_CACHE=/data
      - HF_MODULES_CACHE=/data
      - HF_HUB_CACHE=/data
    command: ["--dtype", "bfloat16", "--usage-stats", "on", "--sharded", "false", "--model-id", "meta-llama/Llama-3.1-8B-Instruct", "--port", "5009", "--cuda-memory-fraction", "0.3"]
    deploy:
      resources:
        reservations:
          devices:
          - driver: nvidia
            # that's the closest analogue to --gpus; provide
            # an integer amount of devices or 'all'
            count: 1
            # Devices are reserved using a list of capabilities, making
            # capabilities the only required field. A device MUST
            # satisfy all the requested capabilities for a successful
            # reservation.
            capabilities: [gpu]
    runtime: nvidia
    healthcheck:
      test: ["CMD", "curl", "-f", "http://text-generation-inference:5009/health"]
      interval: 5s
      timeout: 5s
      retries: 30
  llamastack:
    depends_on:
      text-generation-inference:
        condition: service_healthy
    image: llamastack/distribution-tgi
    network_mode: "host"
    volumes:
      - ~/.llama:/root/.llama
      # Link to TGI run.yaml file
      - ./run.yaml:/root/my-run.yaml
    ports:
      - "5000:5000"
    # Hack: wait for TGI server to start before starting docker
    entrypoint: bash -c "sleep 60; python -m llama_stack.distribution.server.server --yaml_config /root/my-run.yaml"
    restart_policy:
      condition: on-failure
      delay: 3s
      max_attempts: 5
      window: 60s