run-0025
performance
citable URL: https://halobench.com/records/run-0025/ — this address never moves; the anchor /records/#run-0025 keeps resolving
date 2026-08-08
suite llama-bench@1 · config cfg-0010
cell depth 32,768 tokens (machine-written trailer)
Metrics
| metric | value |
|---|---|
| decode_tps | 15.256436 |
N=3 · stddev 0.008009 t/s — machine-written trailer, parsed at build time
Guard chain
waiver — Guarded at run-0005 for this model+build+backend at --parallel 1. llama-bench cells vary KV-quant (lossy) and flash-attn (binary), across which capability cannot be inherited, and llama-bench serves no endpoint to guard directly. Recorded as a known gap rather than an implied pass.
Energy
metered window: eng-0004
Cited by — computed at build time, never stored
model pages qwen35-122b