#2280 bench: community report, RTX PRO 6000 Blackwell Workstation Edition 96 GB (sm_120), IQ3_S fully resident

Descripción

Adds one community benchmark folder from an RTX PRO 6000 Blackwell Workstation Edition 96 GB (reported sm_120 by the engine archs), Core Ultra 9 285K, 253.4 GiB RAM, Windows 11 build 26200.

What was measured: official release Strata 0.1.43.1, Flash-Next IQ3_S GSQ-RCO native weights + iq3_s pack + RT MTP, 262,144-token context, every expert resident in VRAM (----kv int8 --kv-resident 0 --spec 4 --mtp rt --lookup-chain 4 --batch-mtp'""' "''", the configuration discussed in #2261). One full `--long` suite of tools/strata_bench.py at three runs each, quiet evening regime: prompt tok/s 4,135.6 / 4,937.6 / 5,049.1 at 4,019 / 16,058 / 32,049 tokens; short-chat decode median 191.3 tok/s; agent follow-up turns 180.8 tok/s with 68.8 % conversation-cache hit. README.md and results.json come from the tool unchanged (already placeholder-sanitized); no correctness check ran in this run, said so in the report.

What it changes: the new folder plus one row in bench/results/COMMUNITY.md and one entry in docs/COMMUNITY_BENCHMARKS.md. Nothing else touched.

How checked: numbers read straight from the suite output; engine manifest BUILD.json archs [75,86,89,120] + ptx on CUDA 13.0; expert residency 100% with zero doorbell observed during the runs (#2261 setup).