47 lines
1.7 KiB
Bash
47 lines
1.7 KiB
Bash
|
#!/bin/bash
|
|||
|
|
# vllm_wrapper.sh —— 拦截 vllm 二进制，启动前做兼容修补后再交给真实 vllm
|
|||
|
|
#
|
|||
|
|
# 设计照搬社区已上线引擎 zhengzhongwei/vLLM-Iluvatar_bi-100-opt 的做法：
|
|||
|
|
# 把镜像里的 `vllm` 换成本脚本（真身改名 vllm_real），平台下发的
|
|||
|
|
# `vllm serve /model --port 80 ...` 就会自然流经这里，**不需要平台改命令**。
|
|||
|
|
#
|
|||
|
|
# 只在 `vllm serve <目录>` 这一种形态上做修补，其余子命令原样透传。
|
|||
|
|
set -u
|
|||
|
|
|
|||
|
|
# 与 zhengzhongwei/vLLM-Iluvatar_bi-100-opt 一致的真身路径；Dockerfile 里 vllm -> *_real
|
|||
|
|
REAL=/usr/local/corex/lib64/python3/dist-packages/bin/vllm_real
|
|||
|
|
PREFLIGHT=/opt/preflight.py
|
|||
|
|
|
|||
|
|
if [ "${1:-}" = "serve" ] && [ -n "${2:-}" ]; then
|
|||
|
|
MODEL_DIR="$2"
|
|||
|
|
shift 2
|
|||
|
|
|
|||
|
|
EXTRA=""
|
|||
|
|
if [ -f "$PREFLIGHT" ]; then
|
|||
|
|
# preflight 失败也绝不阻断：只打日志，按原命令跑
|
|||
|
|
if python3 "$PREFLIGHT" --model "$MODEL_DIR" --out /tmp/mhxc_preflight.json \
|
|||
|
|
>/tmp/mhxc_preflight.out 2>/tmp/mhxc_preflight.err; then
|
|||
|
|
EXTRA=$(python3 - /tmp/mhxc_preflight.json <<'PY'
|
|||
|
|
import json, sys, shlex
|
|||
|
|
try:
|
|||
|
|
with open(sys.argv[1], encoding="utf-8") as f:
|
|||
|
|
d = json.load(f)
|
|||
|
|
print(" ".join(shlex.quote(a) for a in (d.get("extra_args") or [])))
|
|||
|
|
except Exception:
|
|||
|
|
print("")
|
|||
|
|
PY
|
|||
|
|
)
|
|||
|
|
else
|
|||
|
|
echo "[vllm_wrapper] preflight 非 0 退出，按原命令继续"
|
|||
|
|
fi
|
|||
|
|
sed 's/^/[vllm_wrapper] /' /tmp/mhxc_preflight.err 2>/dev/null || true
|
|||
|
|
fi
|
|||
|
|
|
|||
|
|
echo "[vllm_wrapper] model=$MODEL_DIR extra=${EXTRA:-（无）}"
|
|||
|
|
echo "[vllm_wrapper] exec: vllm_real serve $MODEL_DIR $EXTRA $*"
|
|||
|
|
# shellcheck disable=SC2086
|
|||
|
|
exec "$REAL" serve "$MODEL_DIR" $EXTRA "$@"
|
|||
|
|
fi
|
|||
|
|
|
|||
|
|
exec "$REAL" "$@"
|