#!/usr/bin/env bash
#
# 一键构建 + 启动 Llama 4 Scout（8 卡）
#
#   MODEL_HOST_DIR=/mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct \
#   ./run.sh
#
set -euo pipefail

IMAGE="${IMAGE:-llama4-scout:2026-10-23}"
NAME="${NAME:-llama4-scout}"
PORT="${PORT:-8091}"
MODEL_HOST_DIR="${MODEL_HOST_DIR:-/mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct}"

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

echo "==> build ${IMAGE}"
docker build -t "${IMAGE}" .

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

# 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。
docker run -d --name "${NAME}" \
  --privileged \
  --pid host \
  --network host \
  --ipc host \
  -v /dev:/dev \
  -v "${MODEL_HOST_DIR}:/models/llama4-scout:ro" \
  --ulimit memlock=-1 \
  --ulimit stack=67108864 \
  -e TP=8 \
  -e MAX_MODEL_LEN=4096 \
  -e MAX_NUM_SEQS=1 \
  -e GPU_MEM_UTIL=0.90 \
  --restart unless-stopped \
  "${IMAGE}"

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