#2206 bench: community report, 2x RTX 3090 24 GB (NVLink), IQ3_S, 204K, eng…

本文

Community bench report: 2x RTX 3090 24 GB (NVLink), Linux CUDA.

- CPU: Xeon E5-2696 v3 (18c/36t), 126 GiB RAM, driver 580.159.03
- Model: Swift-Qwen3.8-Flash-Next IQ3_S, 204,800-token context, layer split auto
- Engine 0.1.43.1 (commit a0116b27), suite 1, 3 runs per test

Decode 118.6 (93.5-151.9) short chat / 123.9 (92.6-140.6) agent follow-ups;
prompt 1,458.7 (4K) / 2,392.2 (16K) tok/s; TTFT 2.78 s (4K) / 6.74 s (16K);
agent cache hit 68.7 %.

STRATA_ADAPT_OVERLAP=1 is set explicitly in the launch script (kept on from
the 0.1.43 CUDA default; 661ff293 turned it opt-in).

The server's allowed_hosts (a personal domain) was removed from results.json.