[nsbradley88]

Results-only community report (no engine changes), following docs/COMMUNITY_BENCHMARKS.md.

• Hardware: HP ZBook Ultra G1a laptop, Ryzen AI Max+ 395 / Radeon 8060S (gfx1151), 128 GB unified memory, 70 W package limit under load.
• Software: Strata v0.1.41 (fb58e0d) source build, ROCm 7.14.1, table 100401; docs/STRIX_HALO.md §5 fast config plus --spec 5 --mtp-window 8192.
• Model: Unsloth UD-IQ4_XS, all 24,576 experts on the GPU.
• Results (medians of 4–6 fresh-process runs at 8,192 / 32,768 / 131,072 prompt tokens): prompt 969 / 1,067 / 1,031 tok/s; decode 59.2 / 57.8 / 42.0 tok/s. Needle 15/15, agentic coding suite 5/5.
• Limitations: prompt throughput is ~20% below the §6 desktop table, which we attribute to the 70 W limit (config, pack, ROCm runtime and kernel args were A/B'd); TTFT not measured separately; tools/needle_bench.py not run.
• Also includes same-session comparisons (--spec 4/5/6, the §6 flags, ROCm 10.2 vs 7.14.1, batch slots) and a failure note: setup's --expert-cache auto caused a global OOM here ([Strix Halo, 128 GB unified memory] setup's --expert-cache auto (no --resident-budget-gib) causes a global OOM with UD-IQ4_XS #1715).

Write-up and configs: https://github.com/nsbradley88/halo-tuning

🤖 Generated with Claude Code

[Niko1221]

Thank you, in v0.1.42 with your commit, listed in docs/COMMUNITY_BENCHMARKS.md and in bench/results/COMMUNITY.md.