#!/usr/bin/env bash
#
# Llama 4 Scout 容器入口：
#   1. 幂等应用 vLLM CoreX 的 Llama4 chunked-attention 补丁
#   2. 检查模型目录
#   3. exec vLLM（信号可直接传给 vLLM）
#
set -euo pipefail

python3 /opt/llama4/apply_patch.py

MODEL_PATH="${MODEL_PATH:-/models/llama4-scout}"
SERVED_NAME="${SERVED_NAME:-llama4-scout}"
HOST="${HOST:-0.0.0.0}"
PORT="${PORT:-8091}"
TP="${TP:-8}"
MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
MAX_NUM_SEQS="${MAX_NUM_SEQS:-1}"
GPU_MEM_UTIL="${GPU_MEM_UTIL:-0.90}"
EXTRA_ARGS="${EXTRA_ARGS:-}"

if [[ ! -f "${MODEL_PATH}/config.json" ]]; then
  echo "ERROR: 模型目录不完整：${MODEL_PATH}/config.json 不存在" >&2
  echo "请确认已挂载模型，例如：" >&2
  echo "  -v /mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct:/models/llama4-scout:ro" >&2
  exit 1
fi

echo "============================================================"
echo "Llama 4 Scout"
echo "  model      : ${MODEL_PATH}"
echo "  served name: ${SERVED_NAME}"
echo "  listen     : ${HOST}:${PORT}"
echo "  TP         : ${TP}"
echo "  max len    : ${MAX_MODEL_LEN}"
echo "  max seqs   : ${MAX_NUM_SEQS}"
echo "  gpu util   : ${GPU_MEM_UTIL}"
echo "  extra      : ${EXTRA_ARGS}"
echo "============================================================"

# shellcheck disable=SC2086
exec vllm serve "${MODEL_PATH}" \
  --host "${HOST}" \
  --port "${PORT}" \
  --served-model-name "${SERVED_NAME}" \
  --tensor-parallel-size "${TP}" \
  --max-model-len "${MAX_MODEL_LEN}" \
  --max-num-seqs "${MAX_NUM_SEQS}" \
  --gpu-memory-utilization "${GPU_MEM_UTIL}" \
  --enforce-eager \
  ${EXTRA_ARGS}