#2047 Community benchmark: Radeon AI PRO R9700 + Ryzen 7 5700X (AM4), IQ3_S, HIP + WMMA, engine 0.1.43

Descrição

Results-only submission per docs/COMMUNITY_BENCHMARKS.md.

**Hardware:** Radeon AI PRO R9700 32 GB (gfx1201) on an AM4 board, Ryzen 7 5700X (8C/16T, no AVX-512), 64 GB DDR4-3600 dual channel, Ubuntu 26.04, dedicated inference server (no desktop/background load).

**Model/config:** Qwen3.8-Flash-Next GSQ-RCO IQ3_S, 262K context, `--kv int8 --kv-resident 32768`, MTP `--spec 4`, expert-cache/prefill auto, `STRATA_HIP_WMMA=1` set (env var, not visible in the launch line). Engine 0.1.43, HIP backend from setup's TheRock wheels.

**Headline (medians of 3, suite with `--long`):** 16K fresh prompt 2,820 tok/s (TTFT 5.73 s), 32K 2,785 tok/s (TTFT 11.55 s), 4K 2,009 tok/s, chat decode 75.5 tok/s, agent-session decode 82.9 tok/s with 68.7% cache hit and ~1.8 s per-turn first token.

**Notes:** 16K prefill slightly above the RDNA4 numbers in docs/AMD_HIP.md (2,700 with WMMA) on an older AM4 platform. Two correctness spot checks (tool call + executed coding task) appended to the report; both passed. No engine changes in this PR.