Package Details: cutlass 4.8.0-1

Git Clone URL: https://aur.archlinux.org/cutlass.git (read-only, click to copy)
Package Base: cutlass
Description: CUDA Templates for Linear Algebra Subroutines
Upstream URL: https://github.com/NVIDIA/cutlass
Keywords: ai gemm nvidia
Licenses: BSD-3-Clause
Submitter: daskol
Maintainer: Smoolak
Last Packager: Smoolak
Votes: 0
Popularity: 0.000000
First Submitted: 2023-06-20 22:59 (UTC)
Last Updated: 2026-09-24 20:17 (UTC)

Latest Comments

arzeth commented on 2026-07-04 16:18 (UTC)

I have a suspicion that -DCUTLASS_NVCC_ARCHS='75;80;86;89;90;90a' should be updated — add 120 there.

Interestingly, in Gentoo (https://gitweb.gentoo.org/repo/gentoo.git/tree/dev-libs/cutlass/cutlass-4.2.1.ebuild), this param is set to all-major (which means 75;80;90;100;110;120 according to /usr/share/cmake/Modules/Internal/CMakeCUDAArchitecturesAll.cmake line 14), but that's not very friendly to RTX 3000, 4000, yet friendly to the rare NVIDIA Jetson.

I myself would prefer native (this package takes long to compile), but other CUDA-related packages on AUR don't use native. BTW, there's also all value possible (but that's overkill).

lahwaacz commented on 2023-11-30 16:51 (UTC)

Now pybind11 is missing in makedepends... Also please don't add the nvidia group - there is already an nvidia package in the official repositories, which exists in the same "namespace" as package groups.

lahwaacz commented on 2023-11-30 08:51 (UTC)

Also python-setuptools is missing in makedepends.

lahwaacz commented on 2023-11-30 08:43 (UTC)

The hardcoded -j 48 flag is not nice. You might add ninja to makedepends, -G Ninja to the configure command and remove -j 48 from the build command. Ninja uses all available cores by default.

The build directory can be simply build rather than $pkgname-$pkgver/build.

Also the prefix should not be $pkgdir/usr, the right way to do this is with -DCMAKE_INSTALL_PREFIX=/usr at configure step and DESTDIR="$pkgdir" cmake --install <build_directory> as the install command.