Is it OK to run PyTorch 2.0 docker image on a host with driver 460?

1

Our GPU cluster is shipped with nvidia driver version 460.67 and CUDA 11.2 now and, for some reasons, we are not able to upgrade the underlying driver version.

I want to try out PyTorch 2.0 in such a condition, but saw that the PyTorch 2.0 docker image ghcr.io/pytorch/pytorch:2.0.0-devel was built with CUDA 11.7. I’ve tried to run it and it works well, but I am not sure if there will be implicit problems concerning the CUDA version mismatch.

Could somebody explain is it OK to do so? Or should I build PyTorch from source with CUDA 11.2?

2

Driver 460.67 will allow you to run CUDA 11.x applications in “Minor Version Compatibility” mode as explained here.

I don’t know which user-mode driver the PyTorch 2.0.0 docker image uses, but would guess a newer one.

3

Thanks for reply! So are you suggesting there could be issues when running PyTorch 2.0.0 in such case?

4

No, I would not expect to see any issues running the PyTorch 2.0.0 binaries in such a container.