#3665 [Benchmark] Community Report: RX 7900 XTX / Qwen3.8-Flash-Next / IQ2_XS / 0.1.45

Description

Adds `bench/results/2026-10-21-community-rx-7900-xtx-qwen3-8-flash-next-iq2-xs/` (results only, no engine changes).

- Hardware: RX 7900 XTX 24 GB (PCIe 4.0 x16, 291 W cap), Ryzen 9 7950X, 30 GB DDR5, Kingston NV2 NVMe; Linux (CachyOS), ROCm 7.2.4
- Software: Strata 0.1.45 (477ac13c plus the local changes listed in the README), local HIP source build for gfx1100
- Model: original Qwen3.8-Flash-Next, ISTA-DASLab GSQ-RCO IQ2_XS at revision ed59f920 (setup's pin), 64K context, `--kv int8`, resident experts
- `bench.sh --long`, 3 runs each: chat decode 87.3 tok/s (code 117-123, prose/list 83-89), agent decode 99.7; prompt 802 / 1,280 / 1,462 tok/s at 4K / 16K / 32K; needle recall 9/9 at 8K / 32K / 56K
- Configuration: tuned, not the defaults. It turns on the gfx12 decode and prompt switch sets and route tail skip on gfx1100; tail skip changes answers (teacher-forced KL 0.036, measured). Every switch is listed with its measured effect in the README.
- Limitations: 30 GB of RAM (decode dips when the memory scope's swap fills); a community `bench.sh` report, not a standard-protocol (`strata bench run`) result; the tool read the GPU only as "AMD Radeon", so machine facts were added by hand (marked in the README).

Suggested row for bench/results/COMMUNITY.md:

| [2026-10-21-community-rx-7900-xtx-qwen3-8-flash-next-iq2-xs](2026-10-21-community-rx-7900-xtx-qwen3-8-flash-next-iq2-xs/) | RX 7900 XTX 24 GB, Ryzen 9 7950X, 30 GB | IQ2_XS, 64K | 802 / 1,462 (4K / 32K) | 87.3 chat, 99.7 agent | 0.1.45 + local patches | |

🤖 Generated with [Claude Code](https://claude.com/claude-code)