[Unit]
Description=llama.cpp router (wm, GTX 1080 Ti) — OpenAI-compatible API on 0.0.0.0:11343
Documentation=file:///home/coja/.config/llamacpp/README.md
# A USER unit (fl's llama.service is a system unit under a service user): it runs as coja, so the
# prebuilt binary in ~/.local/opt, the presets in ~/.config/llamacpp and the GGUFs in ~/Documents/models
# need no extra permissions. Starts with the login session (default.target); for a true at-boot start
# without logging in: `loginctl enable-linger coja`.
After=graphical-session-pre.target

[Service]
Type=simple
# Same command as the old `llamaserver` abbr. The wrapper execs the CUDA 12.8 prebuilt and raises the
# PTX JIT cache cap (Pascal is JIT-compiled from PTX; first start after a build update ≈ 50 s).
# 0.0.0.0 = every IPv4 interface (wired + wifi office LAN, ZeroTier, docker0) and still 127.0.0.1;
# keyless, by choice. IPv4 only: `localhost` → ::1 is refused, so clients use 127.0.0.1.
ExecStart=/home/coja/.local/bin/llama-server-cuda --host 0.0.0.0 --port 11343 --cors-origins localhost --models-max 1 --models-preset /home/coja/.config/llamacpp/config.ini
# Presets are read once at startup → `systemctl --user restart llama` after editing config.ini.
Restart=on-failure
RestartSec=3
# The MoE presets run their CPU experts on all 6 cores: yield to interactive apps (GPU work unaffected).
Nice=19
# Idle router holds no model (sleep-idle-seconds per preset), so keeping it up costs nothing.
StandardOutput=journal
StandardError=journal

[Install]
WantedBy=default.target