Repository navigation
benchmark report is wrong about active GPUs. #1985
Description
I have two GPUs (5060 TI 16 GB VRAM + 3060 12 GB VRAM) but am only able to use one with Strata (See #112)
The report created by bench.sh still lists both GPUs as if they are being used. I doubt this will be a very useful benchmark to share.
> head bench/results/2026-10-11-community-2x-rtx-5060-ti-qwen3-8-flash-next-q2-0/README.md
# Community benchmark on NVIDIA GeForce RTX 5060 Ti + NVIDIA GeForce RTX 3060
Measured on 2026-10-11 with `tools/strata_bench.py` (suite 1, 3 runs per test, medians with ranges). Speeds are the engine's own timings; first-token times are measured by the client over HTTP. Speed only: no correctness check was run.
## Hardware and software
- GPU and VRAM: NVIDIA GeForce RTX 5060 Ti + NVIDIA GeForce RTX 3060, 28 GiB VRAM
Activity
bench: report only the GPUs the engine uses (#1985)
Niko1221 commentedon Oct 12, 2026
Thank you for the report, and you are right: the report listed every card nvidia-smi sees even when the engine ran on one.
The fix (0.1.43.1, out now): tools/strata_bench.py now lists only the GPUs the engine uses: the bench's --gpu if it starts the server, otherwise the config's "gpu" (one card or a layer-split list), otherwise a --gpu in its args. If nothing selects a card it names the one card the server reports, not all of them, and the VRAM line sums only those cards. Covered by a test with two fake cards.
Meanwhile: edit the first lines of the generated README.md and results.json (machine.gpu, gpu_vram_mib) by hand, or run the bench with --gpu N so the selection is explicit. Using both of your cards together is a separate question (#112).
Merge upstream main (a0116b2, engine 0.1.43.1): server and bench fixes
tools/strata_bench.py: 報告にはエンジンが使う GPU だけを書く（Niko1221#1985、upstream c…
bench: report only the GPUs the engine uses (Niko1221#1985)