#3124 Request: benchmark config behind 0.1.45 UD-Q4_K_XL decode 117.8 tok/s (R9700)

Descripción

Hi, the 0.1.45 release notes say:

> R9700 32 GB, UD-Q4_K_XL: decode 83.2 to 117.8 tok/s (1.42x, 5 of 5 pairs)

I've been trying to reproduce the 117.8 number on the same hardware (R9700 32GB, Windows, gfx1201) running 0.1.45 with UD-Q4_K_XL (Unsloth), and I'm consistently well below it. Before digging further, could you share the exact benchmark setup behind that number?

My best results, ~4K prompt, single request, 3 runs each (first discarded as warmup), decode tok/s from the server log:

| config | best decode |
|---|---|
| --spec 6, --pcie-frac 0.2, --kv int8, --pool-workers 5, --prefill 8192, --resident-budget-gib 60 | 79.5 |
| --spec 2 (same otherwise) | 68.5 |
| --pcie-frac 0, --spec 6 (same otherwise) | 74.9 |

Full args for the 79.5 run:
--pack strata-data/packs/ud-q4_k_xl --native <UD-Q4_K_XL gguf> --resident-budget-gib 60 --vram-reserve-mib 1024 --expert-profile strata/data/expert-profile.bin --expert-cache auto --pcie-frac 0.2 --pool-workers 5 --prefill 8192 --spec 6 --spec-min-p 0.7 --mtp strata-data/mtp/rt --kv int8 --kv-resident 32768 --max-context 262144
env: STRATA_HIP_WMMA=1, STRATA_SELECT_WMMA=1, STRATA_MMQ_KQ=1, STRATA_ROUTE_RESIDENT=0.25

What I'd like to know for the 117.8 run:
1. Full command line / config (especially --spec/--mtp, --pcie-frac, --kv, --resident-budget-gib, --pool-workers)
2. Prompt length and generated tokens per run
3. How decode tok/s was measured (server log "done" line? averaged over how many runs?)

This would help me (and other R9700 users) tell whether the gap is config or something else. Thanks!