unit 骨架
unit 骨架
[Unit]
Description=vLLM Server
After=network-online.target nvidia-persistenced.service
[Service]
ExecStart=/opt/venv/bin/vllm serve <模型路径> --port 8000
Restart=on-failure
RestartSec=5
EnvironmentFile=/etc/llm/env
TimeoutStopSec=90
[Install]
WantedBy=multi-user.target改 unit 后先 daemon-reload 再 restart；参数走 env 文件便于审计与回滚。