#3598 [Benchmark] RX 6700 XT (gfx1031, Windows 10) - Flash-Next IQ3_XXS self-built HIP engine: prefill 293-315 tok/s, decode 33-38 tok/s

Beschreibung

### What did you measure, or what went wrong

Rerun of the community benchmark on a native-Windows self-built gfx1031 engine, posted here per the 0.2.0-beta tracker reset (originally a comment on #1993, which was closed in the reset).

**Hardware and software**

- GPU: AMD Radeon RX 6700 XT 12 GB (gfx1031 / Navi 22)
- CPU / RAM: Core i5-12490F (6P+4E) / 64 GB DDR4-3200
- OS / driver: Windows 10 Pro 22H2 (19045) / Adrenalin 32.0.21045.11001 (2026-09-25)
- Strata: v0.1.45, engine self-built with `tools\hip\build_windows.bat` (gfx1031 only); BUILD.json: ROCm 10.2.0a20260930, hipBLASLt 100500, vision none; llama.cpp snapshot 3cf03257
- Model: Qwen3.8-Flash-Next IQ3_XXS (ModelScope, SHA256-verified), MTP draft layer on, 64K context (`--max-context 65536 --kv int8 --kv-resident 32768`; the speed table below was measured at 32K)

**Method**: the community harness from `bench/results/2026-10-03-community-r9700-windows/benchmark.py` (synthetic code filler + unique nonce, greedy, 256-token cap, `reasoning_effort:"none"`), 3 reps at 4K, 1 at 16K; numbers below are the engine's own log lines.

| prompt tokens | prefill | decode (with MTP) | TTFT |
|---|---|---|---|
| 4,196 x3 | **292.7 tok/s** (292.6-292.8) | **33.9-38.1** (median 35.5) | 14.4 s |
| 16,680 x1 | **314.7 tok/s** | 35.7 | 53.0 s |

- Expert load: 39.97 GiB at 1.96 GiB/s (42 s); decode expert-cache hit 74.9-75.9% (~10% of routed experts over PCIe)
- MTP drafts accepted ~60% (e.g. 145/239, 137/248, 153/221)
- Confirms the ~33 tok/s reported for this card in #1993; for reference, the archived Linux run (`2026-10-05-community-rx-6700-xt`, IQ2_XS with `HSA_OVERRIDE` to gfx1030) measured 32.1 tok/s decode at 4K - this native Windows build, two quants higher, is on par or faster
- Poor-results note: nothing failed at 4K/16K; a separate intermittent crash during long-prompt prefill is filed as a [Engine] issue

### Reproduction status

Measured locally (every number above is from my machine; logs and bench-output JSON available on request).