Skip to content

Linux engine 0.1.20: sm_80, CUDA 12.8, AVX2

Choose a tag to compare

Strata engine 0.1.20 for Linux x64, compiled for sm_80 (A100) with CUDA 12.8, ggml's CPU backend at the AVX2 baseline (STRATA_PORTABLE; Strata's own AVX-512 kernels are still picked at run time on CPUs that have them). No image encoder.

CUDA 12.8 runs on NVIDIA drivers from 525 up, which is what cloud GPUs such as Colab's tend to have (CUDA 13 needs 580). It is used through this fork's setup.py (branch linux-prebuilt), which adds STRATA_CUDA=12: CUDA 12's pip libraries, driver 525 as the minimum, and a check that the ready-made engine was built with the same CUDA major.

STRATA_CUDA=12 STRATA_PREBUILT_URL=https://github.com/architectds/Strata/releases/download/v0.1.20-linux-sm80-cuda12.8/ ./setup.sh

Built from source by the linux-prebuilt workflow (Ubuntu 22.04, no GPU needed to compile): https://github.com/architectds/Strata/actions/runs/36506157966

SHA-256 9389ec88f32681131e186f5d09cf3c7b2906175ed16e1548bf974ff8cdb44583 strata-linux-x64.zip

Run on a Colab A100-SXM4-40GB (6-core Xeon at 2.2 GHz, 83.5 GiB RAM, no AVX-512 VBMI) on 2026-09-29 with IQ3_S at a 262K context: 17,037 of 24,576 experts in VRAM, decode 42-52 t/s (expert hit rate ~0.97), prefill 870 t/s at 32K and 1,558 t/s at 86K.