[Unit]
Description=vLLM (Qwen3.8-Flash-Next-W4A16 on 4x RTX 3090, TP2xPP2+EP, MTP, 262K)
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
User=root
EnvironmentFile=-/etc/default/vllm
# start.sh is the single source of truth; the checkpoint goes first (CKPT in /etc/default/vllm)
WorkingDirectory=/opt/vllm
# The compile cache is $CACHE_ROOT, default /opt/vllm/.vllm-cache (on disk: it survives a
# reboot). A cold cache costs ~6 min of kernel compile at the next start, so do not use tmpfs here.
# Set CACHE_ROOT in /etc/default/vllm to move it.
ExecStart=/opt/vllm/start.sh
# MTP + cudagraphs do not always exit on SIGTERM: give --shutdown-timeout 10 room, then take the cgroup
KillMode=control-group
TimeoutStopSec=20
Restart=on-failure
RestartSec=15
LimitMEMLOCK=infinity
LimitNOFILE=1048576
# start.sh writes the engine and the PLE offload child to $LOG (/var/log/vllm.log by default);
# systemd/vllm.logrotate rotates it. The pre-serve lines stay in the journal: journalctl -u vllm

[Install]
WantedBy=multi-user.target