llama-4-bi150/run.sh

#!/usr/bin/env bash
#
# Llama 4 Scout / Maverick 一键构建 + 启动（Iluvatar BI-V150）
#
#   ./run.sh scout
#   ./run.sh maverick
#
# 常用覆盖项：
#   GPUS=8,9,10,11,12,13,14,15 ./run.sh scout
#   MODEL_HOST_DIR=/path/to/model ./run.sh maverick
#   EXTRA_ARGS="--max-model-len 8192" ./run.sh scout
#   EXTRA_DOCKER_ARGS="-v /mnt/disk1:/data:ro" ./run.sh scout
#   BUILD=0 ./run.sh scout          # 使用已有镜像，不重新构建
#   DRY_RUN=1 ./run.sh scout        # 只打印 docker run 命令
#
set -euo pipefail

cd "$(dirname "$0")"

usage() {
  cat >&2 <<'EOF'
用法: ./run.sh <scout|maverick>

预设默认值在 presets/<name>.env 中，所有值都可以用环境变量覆盖：
  IMAGE NAME PORT MODEL_HOST_DIR MODEL_MOUNT SERVED_NAME
  TP PP GPUS MAX_MODEL_LEN MAX_NUM_SEQS GPU_MEM_UTIL
  ENFORCE_EAGER EXTRA_ARGS EXTRA_DOCKER_ARGS
  BUILD=0/1  DRY_RUN=0/1
EOF
}

VARIANT="${1:-}"
if [[ -z "${VARIANT}" || "${VARIANT}" == "-h" || "${VARIANT}" == "--help" ]]; then
  usage
  exit 2
fi

PRESET="presets/${VARIANT}.env"
if [[ ! -f "${PRESET}" ]]; then
  echo "ERROR: 未知预设: ${VARIANT}" >&2
  usage
  exit 2
fi

# shellcheck disable=SC1090
source "${PRESET}"

IMAGE="${IMAGE:-${DEFAULT_IMAGE}}"
NAME="${NAME:-${DEFAULT_NAME}}"
PORT="${PORT:-${DEFAULT_PORT}}"
MODEL_HOST_DIR="${MODEL_HOST_DIR:-${DEFAULT_MODEL_HOST_DIR}}"
MODEL_MOUNT="${MODEL_MOUNT:-/models/llama4}"
SERVED_NAME="${SERVED_NAME:-${DEFAULT_SERVED_NAME}}"
TP="${TP:-${DEFAULT_TP}}"
PP="${PP:-${DEFAULT_PP}}"
GPUS="${GPUS:-${CUDA_VISIBLE_DEVICES:-${DEFAULT_GPUS}}}"
MAX_MODEL_LEN="${MAX_MODEL_LEN:-${DEFAULT_MAX_MODEL_LEN}}"
MAX_NUM_SEQS="${MAX_NUM_SEQS:-${DEFAULT_MAX_NUM_SEQS}}"
GPU_MEM_UTIL="${GPU_MEM_UTIL:-${DEFAULT_GPU_MEM_UTIL}}"
ENFORCE_EAGER="${ENFORCE_EAGER:-1}"
EXTRA_ARGS="${EXTRA_ARGS:-}"
EXTRA_DOCKER_ARGS="${EXTRA_DOCKER_ARGS:-}"
BUILD="${BUILD:-1}"
DRY_RUN="${DRY_RUN:-0}"

GPUS="${GPUS// /}"

if [[ ! -f "${MODEL_HOST_DIR}/config.json" ]]; then
  echo "ERROR: 模型目录不完整: ${MODEL_HOST_DIR}/config.json" >&2
  echo "请先下载模型，或设置 MODEL_HOST_DIR 指向正确目录。" >&2
  exit 1
fi

EXPECTED_GPUS=$((TP * PP))
if [[ -n "${GPUS}" ]]; then
  IFS=',' read -r -a GPU_LIST <<< "${GPUS}"
  if (( ${#GPU_LIST[@]} < EXPECTED_GPUS )); then
    echo "ERROR: TP*PP=${EXPECTED_GPUS}，但 GPUS 只给了 ${#GPU_LIST[@]} 张: ${GPUS}" >&2
    exit 1
  fi
  if (( ${#GPU_LIST[@]} > EXPECTED_GPUS )); then
    echo "WARN: TP*PP=${EXPECTED_GPUS}，但 GPUS 给了 ${#GPU_LIST[@]} 张；vLLM 只会使用其中 ${EXPECTED_GPUS} 张。" >&2
  fi
fi

DOCKER_EXTRA=()
if [[ -n "${EXTRA_DOCKER_ARGS}" ]]; then
  # 按空白拆分；路径中不要带空格。
  read -r -a DOCKER_EXTRA <<< "${EXTRA_DOCKER_ARGS}"
fi

DOCKER_RUN=(
  docker run -d --name "${NAME}"
  --privileged
  --pid host
  --network host
  --ipc host
  -v /dev:/dev
  -v "${MODEL_HOST_DIR}:${MODEL_MOUNT}:ro"
  --ulimit memlock=-1
  --ulimit stack=67108864
  -e CUDA_VISIBLE_DEVICES="${GPUS}"
  -e MODEL_PATH="${MODEL_MOUNT}"
  -e SERVED_NAME="${SERVED_NAME}"
  -e PORT="${PORT}"
  -e TP="${TP}"
  -e PP="${PP}"
  -e MAX_MODEL_LEN="${MAX_MODEL_LEN}"
  -e MAX_NUM_SEQS="${MAX_NUM_SEQS}"
  -e GPU_MEM_UTIL="${GPU_MEM_UTIL}"
  -e ENFORCE_EAGER="${ENFORCE_EAGER}"
  -e EXTRA_ARGS="${EXTRA_ARGS}"
  "${DOCKER_EXTRA[@]}"
  --restart unless-stopped
  "${IMAGE}"
)

if [[ "${DRY_RUN}" == "1" ]]; then
  printf '%q ' "${DOCKER_RUN[@]}"
  printf '\n'
  exit 0
fi

if [[ "${BUILD}" == "1" ]]; then
  echo "==> build ${IMAGE}"
  docker build -t "${IMAGE}" .
else
  echo "==> skip build (BUILD=0), image=${IMAGE}"
fi

echo "==> recreate ${NAME}"
docker rm -f "${NAME}" 2>/dev/null || true

echo "==> run ${VARIANT}: GPUs=${GPUS}, TP=${TP}, PP=${PP}, model=${MODEL_HOST_DIR}"
"${DOCKER_RUN[@]}"

echo "==> wait for http://127.0.0.1:${PORT}/health"
for _ in $(seq 1 120); do
  if curl -sf "http://127.0.0.1:${PORT}/health" >/dev/null 2>&1; then
    echo "ready: http://127.0.0.1:${PORT}/v1/models"
    exit 0
  fi
  sleep 5
done

echo "ERROR: 服务 10 分钟未就绪，查看日志：" >&2
echo "  docker logs -f ${NAME}" >&2
exit 1