#2629 bench: community report, 2x Radeon AI PRO R9700 (gfx1201), IQ3_S layer split, v0.1.43

Descrição

## Summary

A community report: 2x AMD Radeon AI PRO R9700 32 GB (gfx1201) on Strata v0.1.43, with a layer split that keeps every expert in VRAM. It compares stock v0.1.43 with the gfx12 decode switches of #2628, and with `--batch 2`. No code changes.

Setup: Ryzen 9 5900X, PCIe 4.0 x8 per card, 125 GiB DDR4, Ubuntu 24.04, ROCm 7.2.4, v0.1.43 source build. Flash-Next IQ3_S at 262,144 tokens, `--layer-split 26`, all 24,576 experts resident, spec 4 + MTP, int8 KV.

All runs in one session, a fresh server each, in this order:

| Run | Greedy tok/s | Temp 1.0 tok/s | Prompt tok/s at 4.6K / 19K / 69K |
|---|---:|---:|---|
| B: stock | 108.9 | 101.5 | 1,367 / 1,584 / 2,483 |
| E: + the nine decode switches | 115.0 | 106.2 | 1,515 / 1,756 / 2,761 |
| C: E + `HIP_WMMA` + `VERIFY_QDEDUP` | 121.1 | 109.1 | 1,759 / 2,188 / 3,628 |
| D: C + `--batch 2` | 117.1 | 107.8 | 1,750 / 2,213 / 3,502 |
| B2: stock again | 108.5 | 101.0 | 1,368 / 1,552 / 2,441 |

Decode is the geometric mean of four prompts (3 runs each). Prompt reads are one request per size.

## What changed

One folder: `bench/results/2026-10-12-community-2x-r9700-iq3_s-0.1.43/`. It has the README, per-run JSON, configs, engine log excerpts, the benchmark scripts with an offline self-test, and run E's engine source as a patch against v0.1.43.

## Extra Notes

- E against the mean of B and B2: +5.8% greedy, +4.9% sampled, +11-12% prompt. E's greedy output matches stock on all four prompts in all three runs. 11 of 12 seeded sampled runs match; stock itself varies on some seeded requests from run to run.
- C is faster, but its greedy output differs on three of four prompts. `STRATA_HIP_WMMA` alone causes that.
- `--batch 2` with the same requests at every level: 99.5 / 103.3 / 106.9 tok/s in total at 1 / 2 / 4 requests. The batch slots run without MTP drafts.
- Not tested: one R9700, other packs, Windows, quality or recall.
- The README has a proposed `bench/results/COMMUNITY.md` row and `docs/COMMUNITY_BENCHMARKS.md` entry.