Pull requests / #2520
#2520 bench: measure resident lookahead on RX 7900 XTX
open · @KeyanHu-git · 0 comentarios · En GitHub
Descripción
## Title bench: measure resident lookahead on RX 7900 XTX Issue: Refs #1922. Measurements for #2476. ## Summary This adds the full performance report for the resident-expert lookahead policy in #2476. On a Ryzen 5 7600X / RX 7900 XTX with IQ1_M, 11 matched English pairs improved decode throughput by **1.67%** (approximate 95% interval: **0.80–2.58%**) and reduced request CPU time by **6.10%** (**5.26–7.06%**). The run used 12 balanced, randomly ordered ON/AUTO pairs and two ON/ON controls: 28 engine starts, 84 requests and 50,400 generated tokens. Each arm used the same executable, a 600-token warmup, and 600-token English and code requests. All observations are retained, including unmatched work and the startup rejected for partial residency. ## What changed - A [report](https://github.com/KeyanHu-git/Strata/blob/e4821fe7edb330668c139460b0deb1fe2d50a7f1/bench/results/2026-10-15-rx7900xtx-resident-lookahead/README.md) with hardware, build identity, configuration, method, all-run medians/ranges and each paired result. - All request timings, output IDs, MTP and expert-tier counters, startup excerpts, fixtures, schedule and exclusion reasons in JSON/CSV. - Figures in PNG, SVG and PDF, plus plotting, analysis, verification and rerun scripts. - The earlier five-pair check, kept separately because it used a shorter warmup. ## Extra Notes The included English pairs have identical output, MTP work and expert-tier counts; nine were faster and all eleven used less CPU time. Code has four matching pairs and remains exploratory under the predeclared minimum of eight. Validation: the published data reproduces every eligibility decision and statistic; the portable runner's schedule and failure cleanup pass two offline tests. Figures were rendered and inspected. Machine paths and unrelated service metadata were removed. This report changes no engine source. 
En el sitio
Enlaces a install, modelos, releases.