The problem

07 — Dual-Use: Inference Server ⇄ Gaming/Desktop

Turning a headless AI box into a desktop (or gaming machine) on demand — udev-driven auto-switching between serving models and using the GPUs interactively.

The problem

A local AI machine idles its GPUs serving models — which is great, except when you want to use the machine: to play a game, do some desktop work, or just look at it. Manually stopping the inference servers, starting the display manager, then reversing it later is a pain and easy to forget (a desktop compositor eating 1–2 GB of VRAM while the models are resident is a subtle way to get OOMs under load).

The fix on this box: a USB monitor switcher (a small USB device you plug in when you want the desktop) drives a udev rule that toggles the machine’s whole role automatically. Plug it in → headless inference stops and the desktop starts. Pull it out → the desktop stops and inference comes back. No systemctl juggling, nothing to remember.

How it works

Three small pieces:

The udev rule identifies the device by fingerprint, so nothing else plugged into the box can trigger the switch:

ACTION=="add",    SUBSYSTEM=="usb", ENV{DEVTYPE}=="usb_device", \
    ATTRS{idVendor}=="05e3", ATTRS{idProduct}=="0610", \
    RUN+="/usr/local/bin/usb-switcher-on.sh"
ACTION=="remove", SUBSYSTEM=="usb", ENV{DEVTYPE}=="usb_device", \
    ATTRS{idVendor}=="05e3", ATTRS{idProduct}=="0610", \
    RUN+="/usr/local/bin/usb-switcher-off.sh"

Find your device’s IDs with lsusb while it’s plugged in.

The on-script (headless → desktop)

#!/bin/bash
# Runs as root via udev — use absolute paths only
LOG=/var/log/usb-switcher.log
echo "[$(date '+%F %T')] switcher connected" >> "$LOG"

/usr/bin/systemctl stop llama-primary.service    2>>"$LOG"
/usr/bin/systemctl stop llama-secondary.service  2>>"$LOG"
/usr/bin/systemctl start gdm.service            2>>"$LOG"

# Switch the monitor to the DisplayPort/USB-C input
/usr/bin/ddcutil setvcp 60 0x1b 2>>"$LOG" || true

The off-script mirrors this: stop gdm, start both llama units, ddcutil setvcp 60 0x11 back to the other input.

Tips & tricks learned from running it

• Ordering matters. On plug-in, stop the inference servers before starting the display manager — free the GPUs first. On pull-out, stop the display manager before starting the servers — avoid a window where both want the GPUs. A GPU contention window here is the kind of thing that manifests as a flaky X session or a model failing to load.
• Inference services must be ordinary startable units. This whole design works because the llama servers are plain systemd units anyone can start/stop, rather than pinned to always-on. The dual-use scripts start and stop them directly.
• udev runs in a minimal environment. The scripts must use absolute paths (/usr/bin/systemctl, not systemctl) — a relative path that works in your shell will silently fail under udev.
• Use || true for the cosmetic part. ddcutil setvcp (which switches the monitor’s active input over DDC/CI) can fail if the monitor or the DRM context isn’t reachable; a failed input switch is logged but non-fatal, so the rest of the transition still happens. Verify your monitor’s VCP input values with ddcutil detect — 0x1b/0x11 are device-specific.
• Log every transition. Appending timestamped lines (plus stderr of each command) to /var/log/usb-switcher.log makes “the machine was in the wrong state” trivial to diagnose — you can see exactly which step ran and which failed.
• The default state should be the server. WantedBy=multi-user.target + the display manager stopped at boot means a plain reboot returns the box to headless inference — the desktop is the exception you opt into with the plug.
• If you don’t have a USB switcher, the same two scripts work behind any trigger: a wall switch on a USB port, a systemd path unit on a file, or just a pair of shell aliases. The device fingerprint in the udev rule is what makes it hands-free.

State model

          plug in switcher (add)                     pull out switcher (remove)
 ┌────────────────────────────────────────┐   ┌────────────────────────────────────────┐
 │ HEADLESS INFERENCING (default)         │   │  DESKTOP / GAMING (interactive)        │
 │ • gdm STOPPED                         │   │  • gdm RUNNING                         │
 │ • llama-primary RUNNING               │   │  • llama-primary STOPPED               │
 │ • llama-secondary RUNNING             │   │  • llama-secondary STOPPED             │
 └────────────────────────────────────────┘   └────────────────────────────────────────┘

Dependencies to verify on your own box

• The display manager unit (gdm.service here — could be lightdm, sddm, …) is what “desktop” means on your system.
• ddcutil is installed and can reach the monitor over DDC/CI (usually needs a DRM/X context).
• Your monitor’s input-source VCP numbering — check with ddcutil detect.
• The USB vendor/product ID of whatever device you use as the trigger.

Related

• 01 — Architecture — why the inference servers are ordinary, individually-startable units
• 03 — Primary vs Secondary — what gets stopped/started by the switch

Source Disclaimer

• AI Generated
• Human Generated
• AI Edited
• Human Edited