# vllm-status-bar
If you run vLLM servers and live in GNOME, this puts your inference cluster in the top bar. vLLM Stats is a GNOME Shell extension that polls each configured server’s /metrics endpoint — every 2 seconds by default — and shows what matters, with status color-coding so you notice trouble before your users do.
The top bar keeps it compact: generation tokens per second, KV cache usage, prefix hit rate and running requests. Click the indicator and you get the full picture in color-coded tiles — requests running, queued for capacity, preemptions, prefix hit rate, generation and prefill tok/s, KV cache percentage and spec-decode acceptance — with status notes under each tile, like an oversubscribed server or a suspiciously low prefix hit rate under load. Model name, endpoint and last update sit in the header, and an unreachable server shows up as an error state instead of silently going stale.
Multiple servers are supported with no fixed limit: each extra endpoint gets its own entry in the popup, and the first nonempty endpoint supplies the top bar. There’s an optional dashboard URL behind an “Open dashboard” action, a configurable refresh interval, and a Test button per endpoint that verifies /metrics and shows the model name. Settings apply to the running indicator immediately, and the same values can be set through GSettings if you script your setup.
Install it with one command from the repo and log out and back in. It targets GNOME Shell 48 through 50 and needs nothing but libsoup3 beyond a vLLM server exposing its Prometheus metrics.
~$