Package Details: python-llama-cpp-cuda-git 0.3.16.r2070.8dcdbd2-1

Git Clone URL: https://aur.archlinux.org/python-llama-cpp-cuda-git.git (read-only, click to copy)
Package Base: python-llama-cpp-cuda-git
Description: Python bindings for llama.cpp (git version with CUDA support)
Upstream URL: https://github.com/inference-sh/llama-cpp-python
Licenses: MIT
Conflicts: python-llama-cpp, python-llama-cpp-cuda
Provides: python-llama-cpp, python-llama-cpp-cuda
Submitter: Premik
Maintainer: Premik
Last Packager: Premik
Votes: 0
Popularity: 0.000000
First Submitted: 2025-10-19 07:24 (UTC)
Last Updated: 2025-10-19 07:24 (UTC)

Latest Comments

throstur commented on 2025-12-18 18:18 (UTC)

If anyone has problems building this package with cuda v. 13+, it's probably because the current PKGBUILD attempts to build "for the current architecture", which will probably be on your CPU, so you get an error like:

nvcc fatal   : Unsupported gpu architecture 'compute_50'

The offending argument: -DCMAKE_CUDA_ARCHITECTURES=native" needs to be corrected.

For example, an RTX 3050 is the sm_86 architecture, so the argument can be changed to: -DCMAKE_CUDA_ARCHITECTURES=86" to compile CUDA just for that one card.