Installing PyTorch with older CUDA version

1

I am trying to install PyTorch on a cluster that has NVidia driver version 390.87 and CUDA toolkit version 9.1 installed.

It seems that to install PyTorch 1.4 we require CUDA 9.2 or 10.1, which I cannot install because I don’t have root privileges on the cluster and the admins are reluctant to upgrade the drivers for fear of breaking things.

So what are my options? Is there a way to get install PyTorch 1.4 with CUDA toolkit 9.1? Or should I install an older version of PyTorch? If so, which one and how?

2

If you use a pre-built binary (check https://pytorch.org/get-started/locally/) you don’t need to install CUDA on the machine.

3

But when I install pytorch via conda install pytorch torchvision cudatoolkit=9.2 -c pytorch, I find that torch.cuda.is_available() returns false. I guess because the driver version (390) does not support CUDA toolkit 9.2 which is required by pytorch 1.4?

4

yeah… You are right… Check this answer:

You could install a previous version of PyTorch that supports CUDA-9.0…

1 Like
5

Thanks. That seems like my issue.

Do you if it is possible to build pytorch 1.4 (or master) from source using an older CUDA version?

6

It should be possible and you’ll find the instructions here.
Let us know, if you encounter any issues.