#2530 Community benchmark: RTX 4090 48 GiB, unsloth UD-IQ4_XS, 131072 ctx

Description

## Summary

Results-only submission (no engine changes): a `tools/strata_bench.py` suite 1 report at `bench/results/2026-10-15-community-rtx-4090/`, run twice (defaults, then `--clients 4 --long`) against the HTTP API of a running server, 3 runs per test with ranges, per-run JSON included (`results.json` plus `results-earlier-run.json` as a cross-check - the two agree).

## Hardware

- NVIDIA GeForce RTX 4090, 48 GiB variant (49140 MiB), PCIe Gen4 x16
- AMD Ryzen 9 9950X (16c/32t), 123.4 GiB RAM
- Ubuntu 22.04.5, driver 580.126.18, CUDA 13.0; engine 0.1.44.1 compiled from source (no ready-made Linux engine at this version)
- Model files on NVMe (WD PC SN810 1 TB), read in place

## Model and configuration

- unsloth/Qwen3.8-Flash-Next-GGUF @ 38bb39ee, UD-IQ4_XS (93.7 GB, 3 shards)
- Context 131072, KV int8 (resident 32768), expert-cache auto (all 55.4 GiB of experts resident in RAM), prefill auto, MTP `--spec 4 --spec-min-p 0.5`, parallel slots 1, images off
- Headline numbers: short-chat decode 152.3 tok/s, 32K fresh prompt 5370 tok/s prefill, agent-session cache hit 68.5% with ~1.06 s first token on follow-up turns, 4 concurrent clients 114.8 tok/s total with 0 errors

## Limitations

- Speed only: no correctness checks were run
- Text only: vision path off
- Single parallel slot: the 4-client concurrency row measures queueing, not batched decode
- Background workloads and power limits not recorded