docker-compose.worker.yml

# ═══════════════════════════════════════════════════════════
#  MAC Worker Node — run this on each worker PC
#  Worker PCs run: vLLM (GPU inference) + optional Jupyter
#  PostgreSQL/Redis/Nginx stay on the master node only.
#
#  Steps:
#    1. Copy this file + worker_agent.py to the worker PC
#    2. Create .env.worker with MAC_ENROLL_TOKEN and MAC_MASTER_URL
#    3. docker compose -f docker-compose.worker.yml up -d
#    4. Admin approves the node in the MAC cluster panel
# ═══════════════════════════════════════════════════════════

services:

  # ── vLLM GPU Inference ─────────────────────────────────────
  vllm:
    image: vllm/vllm-openai:latest
    container_name: mac-worker-vllm
    ports:
      - "${VLLM_PORT:-8001}:8001"
    environment:
      - HF_HOME=/root/.cache/huggingface
      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
    volumes:
      - hf-cache:/root/.cache/huggingface
    command: >
      --model ${VLLM_MODEL:-Qwen/Qwen2.5-7B-Instruct-AWQ}
      --port ${VLLM_PORT:-8001}
      --gpu-memory-utilization ${VLLM_GPU_MEM:-0.85}
      --max-model-len ${VLLM_MAX_LEN:-8192}
      --trust-remote-code
      --enforce-eager
      --served-model-name ${VLLM_SERVED_NAME:-Qwen/Qwen2.5-7B-Instruct-AWQ}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped
    networks:
      - worker-net
    healthcheck:
      test: ["CMD-SHELL", "curl -sf http://localhost:${VLLM_PORT:-8001}/health || exit 1"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 120s

  # ── Jupyter Kernel Gateway (optional — for notebook offload) ──
  # Enable by setting ENABLE_NOTEBOOK=1 in .env.worker
  jupyter:
    image: jupyter/scipy-notebook:latest
    container_name: mac-worker-jupyter
    ports:
      - "${NOTEBOOK_PORT:-8888}:8888"
    environment:
      - JUPYTER_ENABLE_LAB=no
    command: >
      jupyter kernelgateway
      --KernelGatewayApp.ip=0.0.0.0
      --KernelGatewayApp.port=8888
      --KernelGatewayApp.allow_origin=*
      --KernelGatewayApp.auth_token=${JUPYTER_TOKEN:-mac-notebook-token}
    volumes:
      - notebooks:/home/jovyan/work
    restart: unless-stopped
    networks:
      - worker-net
    profiles:
      - notebook   # only starts with: docker compose --profile notebook up

  # ── Worker Agent ───────────────────────────────────────────
  worker-agent:
    image: python:3.11-slim
    container_name: mac-worker-agent
    working_dir: /app
    volumes:
      - ./worker_agent.py:/app/worker_agent.py:ro
    command: >
      sh -c "pip install --quiet httpx psutil pynvml && python worker_agent.py"
    environment:
      - MAC_MASTER_URL=${MAC_MASTER_URL}
      - MAC_ENROLL_TOKEN=${MAC_ENROLL_TOKEN:-}
      - MAC_NODE_TOKEN=${MAC_NODE_TOKEN:-}
      - MAC_WORKER_NAME=${MAC_WORKER_NAME:-Worker}
      - MAC_VLLM_PORT=${VLLM_PORT:-8001}
      - MAC_NOTEBOOK_PORT=${NOTEBOOK_PORT:-}
      - MAC_TAGS=${MAC_TAGS:-llm}
      - MAC_HEARTBEAT_SEC=${HEARTBEAT_SEC:-10}
    network_mode: host   # needs to see vLLM on localhost AND reach master
    restart: unless-stopped
    depends_on:
      vllm:
        condition: service_healthy

volumes:
  hf-cache:
  notebooks:

networks:
  worker-net:
    driver: bridge