LlamaTray

Troubleshooting

Common LlamaTray issues: missing system tray on GNOME, llama-server not found, port conflicts and GPU monitoring gaps.

System tray is unavailable (GNOME)

GNOME does not show Qt tray icons without the AppIndicator extension. If LlamaTray logs a warning that the tray is unavailable, install and enable it:

sudo dnf install -y gnome-shell-extension-appindicator   # Fedora
gnome-extensions enable appindicatorsupport@rgcjonas.gmail.com
gnome-extensions info appindicatorsupport@rgcjonas.gmail.com

The last command should show State: ENABLED. Log out of GNOME and log back in.

If the extension UUID differs on your Fedora/Nobara release, find it with:

rpm -ql gnome-shell-extension-appindicator | grep '/metadata.json$'

and use the extension directory name as the UUID in gnome-extensions enable.

In virtual machines GNOME may refuse to activate extensions automatically. If tray rendering is still unavailable, LlamaTray safely continues in window mode — everything keeps working, just without the tray icon.

“llama-server not found”

LlamaTray looks for llama-server on your PATH and in ~/.local/bin / ~/llama.cpp/build/bin. If it reports the binary is missing:

  1. Open the built-in llama.cpp manager (bottom bar) and either build from source or install a pre-built binary, or
  2. Install llama.cpp yourself and make sure the binary lands in one of the searched locations, then start LlamaTray again.

Port is already in use

Before starting, LlamaTray checks the port and tries to clean up stale llama-server processes on it. If it still fails:

lsof -i :8080    # adjust the port to match your settings

then free the port or pick a different one in Settings (1024 – 65535).

Router Mode does nothing / model list is empty

Router Mode requires a llama-server build that supports --models-dir and the model router API. If your build is older than that, update llama.cpp (the built-in manager can rebuild it) or use Single Model mode.

GPU metrics are missing

  • NVIDIA: LlamaTray uses the NVML Python bindings first and falls back to nvidia-smi. The nvidia-ml-py package is optional; without it, the nvidia-smi fallback applies when the driver CLI is installed.
  • AMD: Metrics come from rocm-smi, falling back to the amdgpu sysfs interface.
  • When neither backend responds (e.g. no discrete GPU), CPU and RAM monitoring continue to work and GPU rows are hidden.

llama.cpp manager warnings

  • Missing SDK warning before building — the selected backend’s compiler (nvcc, hipcc, or icpx) was not found. Install the backend toolchain first, or pick a different backend (Vulkan is the most universal).
  • NVIDIA CUDA on Ubuntu/Debian — the CUDA packages require the official proprietary NVIDIA driver; the manager warns and refuses to start an invalid apt install if it’s missing.
  • AMD ROCm on Ubuntu/Debian — ROCm packages are not in the default repos there; add AMD’s official repo.radeon.com APT repository first (on dnf/pacman systems the manager installs the ROCm development packages directly).

Something else is wrong

  1. Copy the log panel contents with the one-click Copy Log button — it contains every step LlamaTray took (binary path, port check, full command, server output).
  2. Check the GitHub issues for a matching report, or open a new one with the log attached.