#2796 bench: community report - 2x RTX 3090, IQ3_XXS, 262K context, parallel: 2

本文

bench: community report - 2x RTX 3090, IQ3_XXS, 262K, parallel: 2

Adds bench/results/2026-10-16-community-2x-rtx-3090/ and an entry in
docs/COMMUNITY_BENCHMARKS.md.

Measured with the built-in suite (tools/strata_bench.py, suite 1, 3 runs per test,
medians): Strata v0.1.44.1 source build (project 296b59d2), Flash-Next IQ3_XXS,
262,144-token context, --kv int8, vision on, automatic layer split (~99.4% of the
routed expert mass in VRAM) and "parallel": 2.

  short chat decode      98.3 tok/s (81.3-112.5)
  prefill                440.7 / 917.2 / 1,104.9 tok/s at 4K / 16K / 32K
  agent session decode   111.9 tok/s (96.6-149.0), conversation-cache hit 68.7%
  2-client burst         71.4 tok/s total, latency p50 3.574 s, errors 0

Two things the report documents rather than hides:

1. With "parallel": 2 at 262,144 the engine logs its own low-VRAM warning
   (356-446 MiB free, below its 512 MiB threshold, #2065) and names the fix
   (--vram-reserve-mib 768 / --vram-reserve-later-mib 1112, or lower
   --max-context / --batch). No degradation was observed in these runs, but it
   is probably the most transferable finding here for other 2x24 GB owners
   considering batch slots at a long context.

2. Prefill is unchanged between 0.1.39 and 0.1.44.1 on this box for the same
   harness and prose text: 1,221.5 tok/s at 56,755 fresh tokens on 0.1.44.1,
   against 1,220.9-1,246.9 tok/s at 118K-133K in the 0.1.39 log.

The report also notes that the suite caps generation at 24 tokens for its prompt
tests, so its short decode rows understate (a 300-token run on this box measured
120.2 tok/s where the 24-token rows read 69-96).

Files: README.md (report), results.json (per-request data), strata-iq3_xxs.json,
engine-BUILD.json and engine-log-excerpt.txt. No correctness check was run.