Universal llama.cpp Monitor

A Chrome extension that shows live telemetry for llama.cpp-based engines - llama-server, Unsloth Studio, and llama.cpp forks. Two views: a full-tab dashboard and a side panel.

GPU telemetry is NVIDIA-only (read via nvidia-smi); host CPU/RAM, slots, context fill, speed, and the request inspector work against any llama.cpp-compatible server.

What it shows

• GPU hardware - per-GPU VRAM, temperature, power, utilization, and the combined VRAM pool.
• Host - CPU load and system RAM.
• Slots & context - active slots and context-window fill (33.4k / 84k · 39.8%).
• Speed - decode and prefill tokens/s from the server's /metrics.
• Request inspector - per-request wall clock, prefill/generation timing, and token counts.

Install

1. Register the native host (Windows; no admin required):

   cd host
   install_host.bat

   The host wraps nvidia-smi and the server log. A prebuilt smi_bridge.exe ships with the repo; rebuild it with sh build.sh (or sh build.sh --all for Windows/Linux/macOS).
2. In Chrome: chrome://extensions -> enable Developer mode -> Load unpacked -> this folder.
3. Optional: enable Allow access to file URLs on the extension card for log reading.

Server settings

Only two flags matter:

Flag	Purpose
--metrics	exposes /metrics so the speed gauges populate
--log-file <path>	enables the per-request inspector (plain-text log; do not use --log-jsonl)

/slots and /v1/models are on by default, so model name and context fill need no flag. Unsloth Studio needs none of these - it writes its log automatically and the monitor auto-detects the port and API key.

License

MIT. See LICENSE.