PyTorch + CUDA 11.4

Hi all!

Does Pytorch supports CUDA 11.4 toolkit? Or could I build pytorch with cuda 11.4 support?

Yes, you can build PyTorch from source using all released CUDA versions between 10.2 and 11.5.

2 Likes

For older verion pytorch, such as 1.1.0, is it possible to build from soure with CUDA 11.4?

We are not backporting new library support for old releases. The already released PyTorch versions are supporting the CUDA toolkits which were supported at that time.
With that being said, if no changes e.g. between CUDA 11.3 and 11.4 were needed, you might be able to use the newer CUDA toolkit, but there is no guarantee.

Pytorch for CUDA 11.4 can’t be build because MAGMA-CUDA114 is needed from pytorch :: Anaconda.org but it does not exist. Is there any solution? I’m working in a VM with vGPU 13.1 that supports CUDA 11.4 and I can’t change the drivers because I’m not not admin.

MAGMA is not a requirement and PyTorch can be built with CUDA11.4.
You could try to either install MAGMA with CUDA11.3, 11.5, or build it from source, if needed.

I successfully accessed GPU on CUDA 11.4 and NVIDIA GeForce RTX 2080 Ti

pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html

Your locally installed CUDA toolkit won’t be used unless you build PyTorch from source or a custom CUDA extension. PyTorch binaries ship with their own CUDA dependencies and in your case 11.1 will be used.