←Work

sparkDash

A fix for the SGLang monitoring path in sparkDash, the multi-Spark monitoring dashboard.

Year
2026
Status
Merged Aug 19, 2026
Role
Contributor
Link
PR #53

Overview

sparkDash is an open-source monitoring dashboard for NVIDIA Spark systems: a server plus a web dashboard, packaged for Docker, that watches multiple machines at once. I run local AI hardware daily, and this is the tool that keeps the lab honest.

Pull request

SGLang engine tile and live tok/s under load

#53 · Opened Aug 19, 2026 · Merged Aug 19, 2026  Merged

On a live SGLang server the LLM panel showed the backend badge and model id, but the Engine tile stayed blank and Generation tok/s dropped to zero while requests were still running. The root cause: the Engine state was driven by a sleep gauge that only exists on vLLM, and the panel fell back to a sticky throughput reading that expired after about six seconds.

The fix keeps it SGLang-only: probe /v1/loads (falling back to /get_load) for running and waiting request counts, treat a reachable SGLang process as Active rather than Sleeping, and keep throughput live while requests are in flight. No invented totals: Total Generated still requires real counters.

Status

Merged into MiaAI-Lab/sparkDash on Aug 19, 2026. The same machines that serve Kearuga are the ones being monitored.