Got tired of juggling nvidia-smi, htop, and vLLM's raw /metrics endpoint in three terminals, so I built one dashboard that pulls it all together.
- GPU / CPU / unified memory / disk / network at 1s polling
- vLLM stats: tok/s, TTFT, queue time, KV cache %, prefix cache hit rate
- Auto-discovers running engines via process scan + Docker (multi-engine tabs if you're running more than one)
- 15-min rolling history, thermal throttle / power brake detection
- Rust backend, React frontend, streams over WebSocket
- One-liner install: cargo install spark-dashboard && sudo spark-dashboard service install
MIT, no telemetry, no cloud all runs entirely on the Spark.
If you've got a Spark sitting on your desk, give it a try and let me know what's missing or broken. Especially curious what other metrics you'd want surfaced.
https://github.com/niklasfrick/spark-dashboard
submitted by