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

    # vLLM 多卡 NCCL / 端口直通
    network_mode: host
    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: "1"
      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}"

      # 留空时由 entrypoint 根据 config.json 自动判断。
      # 如果需要手动覆盖，取消注释并填值。
      # TP: "${TP:-}"
      # PP: "${PP:-}"
      # PORT: "${PORT:-}"
      # SERVED_NAME: "${SERVED_NAME:-}"

      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}"
      EXTRA_ARGS: "${EXTRA_ARGS:-}"

    ulimits:
      memlock: -1
      stack: 67108864

    restart: unless-stopped