Pull requests / #1849
#1849 bench: RTX 4090 Laptop GPU follow-up, engine 0.1.41, defaults and --calibrate settings (IQ3_S, 131K and 262K)
open · @30crows · 0 コメント · GitHub で見る
BenchmarksSetup & installNVIDIA / CUDAModels & quantsDocumentation
本文
## Summary Follow-up to the [2026-10-04 RTX 4090 Laptop report](https://github.com/Niko1221/Strata/blob/main/bench/results/2026-10-04-community-rtx-4090-laptop/README.md): the same Lenovo Legion 82WQ (RTX 4090 Laptop GPU at 175 W, i9-13900HX, 64 GB), now on Gentoo with Strata v0.1.41 built from source and NVIDIA 615.78.08. Same model (Flash-Next IQ3_S), scripts and workload at both context limits of that report, once with the defaults and once with the settings `tools/calibrate.py` (setup's `--calibrate`) picked on this laptop: - `--pcie-frac 0.20` (default 0.55), `--spec-min-p 0.30` at 131,072 tokens and `0.70` at 262,144, and `--adapt-every 1 --adapt-swaps 80 --adapt-decay 0.97`. - Against the same day's default runs, decode rose **25-33% at 131,072 tokens** (75.9 / 81.9 / 76.6 tok/s at 4K / 32K / 128K prompt tokens) and **32-39% at 262,144** (74.6 / 71.8 / 73.4 / 70.6 tok/s at 4K / 32K / 128K / 250K). Every calibrated decode run was faster than every default run at the same length. Prompt throughput stayed within 1.2%. - All needle checks passed (6/6 and 9/9 per session); the GPU ran at its 175 W limit with no thermal slowdown. ## What this changes - New folder `bench/results/2026-10-10-community-rtx-4090-laptop-engine-0.1.41-calibrated/`: README, per-run JSON, configs, power logs, engine timing lines, both calibration outputs and the session script. - A short follow-up pointer at the top of `bench/results/2026-10-04-community-rtx-4090-laptop/README.md`; its measurements are unchanged. - One row in `bench/results/COMMUNITY.md`. Results only, no code changes. ## Limitations - One session per setting; the calibrated sessions shared a boot with other measurements (listed in the README). - The comparison with 2026-10-04 is not controlled: OS, kernel, driver, engine and filesystem all changed. - The calibration measures on its own short prompts; its draft-floor picks rest on small differences. - Greedy text differed between sessions, also between two default sessions with the same driver and config; output quality was not evaluated beyond the needle checks. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
関連リンク
インストール・モデル・リリースへの站内リンク。