#2395 bench: community report, Tesla V100-PCIE-32GB + Tesla PG500-216 (2 x sm_70, layer split), Flash-Next GSQ-RCO IQ3_XXS at 256K: --pipeline-windows 2 on two stages, decode +5.3% to +17.9%, prompt unchanged, replies bit-identical

Descrição

Results-only submission, no engine changes. Adds `bench/results/2026-10-14-community-v100-pg500-pipeline-windows/` (README, three CSVs, the engine's timing lines, the two probe scripts) and one index line in docs/COMMUNITY_BENCHMARKS.md, following its report template.

`--pipeline-windows 2` against off on a two-stage layer split of two sm_70 cards; the flag is the only difference between the arms. docs/MULTI_GPU.md turns it on by default for two or more AMD gfx12 stages (two stages since 0.1.44, with the draft layer moved to the first card) and keeps NVIDIA builds opt-in; on CUDA builds `mtp_first_default_on()` in src/program/generate.cpp is false, so the draft layer stays on the last card unless `STRATA_MTP_STAGE0=1`. Here, on the current main with the draft layer on the last card, decode rose 5.3% to 17.9% (engine timings, medians), prompt reads stayed within +0.0% to +0.6%, and the replies were bit-identical in all 72 requests. With the flag on, `STRATA_MTP_STAGE0=1` (the draft layer on the first card, as on gfx12) was slower here: decode -2.9% to -6.7% against the draft layer on the last card, same replies.

Other two-stage numbers in the repository were taken under other conditions: docs/MULTI_GPU.md has RTX 4060 Ti + RTX 5080, decode +16% (no faster with `--mmap-experts` on that 32 GB PC), and a 2x RTX 3090 table (#1447) whose `--pipeline-windows 2` row (decode 136.7, prefill 2,097) used layer split 29 and no `--remote-expert-opt`, against setup's automatic split in the baseline row (137.0 / 136.9, 1,798). Here every expert is resident in VRAM on both cards. We did not investigate the reasons for the differences. The open PR #1747 (engine 0.1.41 with patches, four GP100 stages, +14.6% to +28.8%) points the same way.

🤖 Generated with [Claude Code](https://claude.com/claude-code)