services:
  llama4-scout:
    build:
      context: .
      dockerfile: Dockerfile
    image: llama4-scout:2026-10-23
    container_name: llama4-scout

    # vLLM 多卡 NCCL / 端口直通
    network_mode: host
    ipc: host

    # Iluvatar 运行时需要 privileged + /dev 绑定。
    # 只给 --device=/dev/iluvatar* 会在 worker init 时报
    #   cudaGetDeviceCount() ... Error 1001
    # 这是厂商驱动/容器运行时要求，不是 vLLM 本身的要求。
    privileged: true
    pid: host

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

    environment:
      MODEL_PATH: /models/llama4-scout
      SERVED_NAME: llama4-scout
      HOST: 0.0.0.0
      PORT: "8091"
      TP: "8"
      MAX_MODEL_LEN: "4096"
      MAX_NUM_SEQS: "1"
      GPU_MEM_UTIL: "0.90"
      # EXTRA_ARGS: "--pipeline-parallel-size 2"

    ulimits:
      memlock: -1
      stack: 67108864

    # 可选：某些调度场景需要；不需要可以删掉
    # cap_add:
    #   - SYS_NICE

    restart: unless-stopped