Static frontier status page for SP⚡RKINFER · Powered by SN74 — optimization journey, per-context benchmarks, auto-eval labels, and evaluated PRs. Styled in the org identity (purple #9D7CFF / lime #D4FF12). No build step — index.html + data.js.
Live: gittensor-ai-lab.github.io/sparkinfer/dashboard/
Companion surfaces:
| Surface | What |
|---|---|
| This dashboard | Frontier charts, PR eval history, optimization journey |
| sparkinfer-web | Landing page + SparkInfer Chat UI |
| sparkinfer | Runtime, kernels, bench, SN74 eval loop |
- Target GPUs — RTX Spark, DGX Spark, RTX 5090, RTX PRO 6000 (
sm_120/sm_121) - SOTA spotlight — Qwen3.6-35B-A3B frontier stats at the top
- Optimization journey — tok/s per landed kernel optimization
- vs llama.cpp — same GGUF, per-context decode on RTX 5090
- Evaluated PRs — bot labels, split into Decode vs Prefill tables (by scored metric)
Current frontier (v0.4.1, Qwen3.6 SOTA): 473 tok/s @ 128 ctx · +71% vs llama.cpp.
Open dashboard/index.html in a browser (loads dashboard/data.js).
Canonical data is data.json; data.js is generated from it
(window.SPARKINFER = <data.json>) so the page works on file:// and GitHub Pages.
python3 -c "import json;d=json.load(open('dashboard/data.json'));open('dashboard/data.js','w').write('window.SPARKINFER = '+json.dumps(d,indent=2)+';\n')"The eval bot does this automatically. After each evaluated PR, eval/pr_eval_bot.py upserts the
verdict into prs[], ratchets status.frontier_tps, regenerates data.js, and pushes.
Enable Pages (Settings → Pages → deploy from main, root). Serves at
https://gittensor-ai-lab.github.io/sparkinfer/dashboard/.