run-0113
performance
citable URL: https://halobench.com/records/run-0113/ — this address never moves; the anchor /records/#run-0113 keeps resolving
date 2026-08-12 · window 2026-08-12T20:27:50Z → 2026-08-12T20:28:58Z
suite llama-bench@1 · config cfg-0032
cell depth 0 tokens (machine-written trailer)
Metrics
| metric | value |
|---|---|
| decode_tps | 51.056208 |
N=3 · stddev 0.11947 t/s — machine-written trailer, parsed at build time
Guard chain
waiver — No capability guard exists for Qwen3.6-35B on this backend/build/KV combination (run-0004's guard is a different suite and build; run-0005 guards the 122B, not this model). llama-bench measures throughput only and serves no endpoint for a guard to attach to. clm-0050/0051's methodology note records the closest available correctness check for these arms — 41/41 layers on GPU, CPU utilisation identical to stock, no stride bug — which is evidence against a gross correctness failure but is not a capability guard. Recorded as a gap.