Summary
The public dashboard KPI strip labels a number as “Frontier decode” but prefers Qwen3.6 512-context throughput (q36ctx512.tps) over the scored competition frontier (q36.frontier_tps, which matches the 128-tok row and the SOTA card).
Against live dashboard/data.js:
| source |
tok/s |
qwen36.frontier_tps / ctx 128 |
476.54 |
ctx 512 (what KPI strip shows) |
480.91 |
So the page publishes two disagreeing “frontier” decode numbers, and the vs-llama.cpp % is computed on the wrong context.
Expected
Bind the KPI strip to q36.frontier_tps / q36.ref_tps (128-tok competition frontier), matching the SOTA card and SN74 scored frontier identity.
Scope
dashboard/index.html only. No GPU. Distinct from #639/#640 (bar colors) and #641/#642 (prefill labels).
Summary
The public dashboard KPI strip labels a number as “Frontier decode” but prefers Qwen3.6 512-context throughput (
q36ctx512.tps) over the scored competition frontier (q36.frontier_tps, which matches the 128-tok row and the SOTA card).Against live
dashboard/data.js:qwen36.frontier_tps/ ctx128512(what KPI strip shows)So the page publishes two disagreeing “frontier” decode numbers, and the vs-llama.cpp % is computed on the wrong context.
Expected
Bind the KPI strip to
q36.frontier_tps/q36.ref_tps(128-tok competition frontier), matching the SOTA card and SN74 scored frontier identity.Scope
dashboard/index.htmlonly. No GPU. Distinct from #639/#640 (bar colors) and #641/#642 (prefill labels).