systemctl User Services for AI Inference

System-wide systemd units need root to install and manage. For a single-user dev GPU server, running your vLLM or Ollama via the user’s own systemd instance is cleaner – no sudo prompts, per-user config, easy lifecycle. On dedicated GPU hosting this suits solo developer boxes well.

Contents

• Enable user systemd
• Unit file
• Control
• Versus system unit

Enable

User systemd instances need lingering enabled so they survive logout:

sudo loginctl enable-linger $USER

Check: loginctl show-user $USER | grep Linger should show Linger=yes.

Unit File

Create ~/.config/systemd/user/vllm.service:

[Unit]
Description=vLLM inference server
After=default.target

[Service]
Type=simple
WorkingDirectory=%h/vllm-project
ExecStart=%h/.venvs/vllm/bin/python -m vllm.entrypoints.openai.api_server \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --port 8000
Restart=on-failure
RestartSec=10s
TimeoutStopSec=300

[Install]
WantedBy=default.target

%h expands to the user’s home directory.

Control

systemctl --user daemon-reload
systemctl --user enable vllm
systemctl --user start vllm
systemctl --user status vllm
journalctl --user -u vllm -f

All commands run as your user – no sudo required.

Versus System

Use user units when:

• Single-user server
• Developer environment – quick iteration on unit files
• You want isolation from the rest of the system

Use system-wide units when:

• Shared multi-user server
• Service must bind port below 1024
• You need boot-time start before any user logs in
• Production – system units have better log integration for monitoring

See graceful vLLM shutdown and systemd service files for AI inference.