llama-4-bi150/docker-compose.yml

services:
  llama4:
    build:
      context: .
      dockerfile: Dockerfile
    image: llama4-bi150:${TAG:-2026-10-26}
    container_name: llama4

    # 镜像本身不设置启动命令，启动方式全部由这里决定。
    entrypoint: ["/usr/local/bin/llama4-entrypoint.sh"]
    working_dir: /opt/llama4

    # 容器内固定监听 80，宿主机端口由 HOST_PORT 决定（默认 24396）。
    expose:
      - "80"
    ports:
      - "${HOST_PORT:-24396}:80"

    # bridge 网络 + 端口映射；ipc host 供 vLLM/PyTorch 多卡共享内存。
    ipc: host

    # Iluvatar 运行时要求
    privileged: true
    pid: host

    volumes:
      - /dev:/dev
      - ${MODEL_HOST_DIR:-/mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct}:/models/llama4:ro

    environment:
      MODEL_PATH: /models/llama4
      AUTO_DETECT: "${AUTO_DETECT:-1}"
      HOST: "${HOST:-0.0.0.0}"

      # 默认暴露全部 16 张卡；entrypoint 会根据模型选择实际 TP*PP。
      CUDA_VISIBLE_DEVICES: "${GPUS:-0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15}"

      # 容器内端口固定 80，模型名固定 llm（AGENTS.md 5.2 / 5.5）。
      PORT: "80"
      SERVED_NAME: "${SERVED_NAME:-llm}"

      # 留空时由 entrypoint 根据 config.json 自动判断。
      # 需要手动覆盖时直接设置同名环境变量即可。
      TP: "${TP:-}"
      PP: "${PP:-}"

      MAX_MODEL_LEN: "${MAX_MODEL_LEN:-4096}"
      MAX_NUM_SEQS: "${MAX_NUM_SEQS:-1}"
      GPU_MEM_UTIL: "${GPU_MEM_UTIL:-0.90}"
      ENFORCE_EAGER: "${ENFORCE_EAGER:-1}"
      PYTHONUNBUFFERED: "1"
      EXTRA_ARGS: "${EXTRA_ARGS:-}"

    ulimits:
      memlock: -1
      stack: 67108864

    restart: unless-stopped