Repository navigation
[Feature]: Cuda 13 support #1360
Description
Background & Description
Building binary files for cuda 13.
API & Usage
No response
How to implement
No response
{{ message }}
Building binary files for cuda 13.
No response
No response
Activity
[-][Feature]: Support CUDA 13[/-][+][Feature]: Cuda 13 support[/+]aropb commentedon Apr 9, 2026
@martindevans Is it possible to do this?
martindevans commentedon Apr 10, 2026
If llama.cpp supports CUDA 13 then it can probably be added to the list here to produce the binaries. That would be the first step.
However, CUDA builds are very slow, so I'm hesitant to add a new one. What does CUDA 13 add over our current CUDA 12 binaries?
aropb commentedon Apr 10, 2026
This is the current version of CUDA, the 12th version is already archived.
https://developer.nvidia.com/cuda-downloads
You don't even have to make nuget packages, Llamasharp Binaries are enough.
The latest version is CUDA 12 -> 12.9
https://developer.nvidia.com/cuda-12-9-1-download-archive
manutiedra commentedon Jul 4, 2026
CUDA 13 is a big improvement:
https://www.linkedin.com/pulse/whats-new-cuda-13-why-matters-bojan-tunguz-ph-d--e4ele
Some apps I had improvements with CUDA 13 so It will be great to have CUDA 13 supported in LlamaSharp
aropb commentedon Aug 29, 2026
Is it planned to add a build for CUDA13?
I think cuda11 can be removed.
martindevans commentedon Sep 1, 2026
This would be a good issue for a new contributor to take on. The steps involved would be:
All our builds are made with this GitHub action. First up remove CUDA11 from there and replace it with CUDA13.
LL# works out with DLLs to load at runtime, replace CUDA11 with CUDA13 there. Start here.
Create a nuspec file for CUDA13 (see CUDA12 nuspec), so that a nuget package gets published.
(these can be tackled as individual PRs, by different people if necessary)
modelpath-dev commentedon Sep 16, 2026
I'll take this.
compile.ymlon master is already CUDA 12.4.0 only, so rather than swapping 12 out I added a CUDA 13.0.0 job next to it and wired the loader plusLLamaSharp.Backend.Cuda13packages.PR: #1439
The compile workflow still has to produce the
cu13.0.0binaries (Update Binaries) before those packages have anything to ship. Happy to split this if you'd rather land the loader first.aropb commentedon Oct 17, 2026
Thank you very much, it works about 10 % faster than CUDA 12.