[Vop3p]

Data only: one folder under bench/results/ and its line in docs/COMMUNITY_BENCHMARKS.md. Re-filed from #927 (closed with the history rewrite), measured again on 0.1.40.1 instead of 0.1.39.

2x RX 6900 XT (gfx1030, PCIe 4.0 x8 each), Ryzen 5 5600X, 128 GB, Ubuntu 26.04 / Linux 7.0 / ROCm 10.0.0, Qwen3.8-Flash-Next GSQ-RCO IQ3_S at a 131,072-token context, the repository's benchmark.py (4,096 / 32,768 / 128,000 tokens x 3) and needle_bench.py (six needles), one cold server per configuration, adaptive swaps on. Nine configurations: one card, the layer split and the expert-helper mode, each as stock 82f46a8, with the two switches 0.1.40 ships for gfx103x (STRATA_HIP_PROMPT_F16=1 STRATA_SH_STREAM=1), and with #1150 + #1151 + #1167 on top.

	one card	layer split	helper
prompt tok/s at 4K / 32K / 128K, stock	458 / 473 / 458	446 / 723 / 821	459 / 474 / 459
the two switches	811 / 973 / 913	791 / 1,443 / 1,631	816 / 975 / 913
+ #1150 #1151 #1167	969 / 1,152 / 1,126	968 / 1,685 / 2,003	980 / 1,151 / 1,127
decode tok/s (range over the three lengths), stock	47-48	59-67	64-71
the two switches	49-51	65-73	69-75
+ the PRs	50-52	70-75	66-72

All 54 needles found. Two things the README says in full: the two-card configurations were measured with GFXOFF disabled, because at the kernel's default the stock and the switches-only layer split both stalled on their first 128,000-token request with the adaptive swaps on (#884; the stalled runs' 4K and 32K rows are kept and match); and the decode gain over the 0.1.39 report from this machine (8-18%) is the swaps and the 0.1.40 changes together.

Developed with an AI coding assistant; every number above was measured on 2x RX 6900 XT (gfx1030, PCIe 4.0 x8 each) / Ryzen 5 5600X, ROCm 10.0.

[Niko1221]

Thank you, xjc10. The 2x RX 6900 XT report is in 0.1.40.2 as a bench/results folder with your authorship kept. Raw dumps and logs were trimmed (TRIMMED.md in the folder lists them); the README, summaries and configs are unchanged. The row is in bench/results/COMMUNITY.md.