docker-compose.gpu.yml

# ================================================================
# docker-compose.gpu.yml — GPU 推理服务编排
# ================================================================
# 用法：docker compose -f docker-compose.yml -f docker-compose.gpu.yml up --build
#
# 在基础的 registry + ml-worker + gateway 之上，
# 增加 vllm (GPU) 和 llm-worker (CPU) 两个服务。
# ================================================================

services:

  # ── vLLM 推理引擎（唯一的 GPU 服务） ──────────────────────
  vllm:
    build:
      context: .
      dockerfile: Dockerfile.vllm
    container_name: vllm-engine
    ports:
      - "8100:8100"
    environment:
      - MODEL_NAME=Qwen/Qwen2.5-1.5B-Instruct
      - MAX_MODEL_LEN=2048
      - GPU_MEMORY_UTILIZATION=0.85
      # HuggingFace 缓存目录
      - HF_HOME=/app/models
      # 国内镜像加速下载（如果模型未缓存的话）
      - HF_ENDPOINT=https://hf-mirror.com
    volumes:
      # 模型缓存：避免每次启动都重新下载
      - hf-model-cache:/app/models
    # ★ GPU 配置 — docker compose 语法 ★
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1          # 使用 1 块 GPU
              capabilities: [gpu]
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8100/health"]
      interval: 30s
      timeout: 10s
      retries: 5
      start_period: 600s       # 首次冷启动需下载模型（~300s）+ 加载权重 + warmup

  # ── LLM Worker（CPU 代理层） ─────────────────────────────
  llm-worker:
    build:
      context: .
      dockerfile: Dockerfile.llm-worker
    container_name: llm-worker
    ports:
      - "8003:8003"
    environment:
      # ★ 容器间通信用服务名，不用 localhost ★
      - VLLM_BASE_URL=http://vllm:8100
      - VLLM_DEFAULT_MODEL=Qwen/Qwen2.5-1.5B-Instruct
    depends_on:
      vllm:
        condition: service_healthy  # 等 vLLM 加载完模型再启动
    restart: unless-stopped

  # ── Gateway 覆盖：追加 LLM Worker 依赖 ──────────────────
  gateway:
    depends_on:
      - llm-worker

volumes:
  hf-model-cache:
    driver: local