PyTorch with CUDA 11 compatibility

22

So you mean binary compatible not suitable for Ampere GPU? we need CUDA toolkit >=11.0?

I find binary compatible here: 1. Preface — CUDA C++ Best Practices Guide 12.3 documentation

image
image1722×830 145 KB

23

You are checking the compatibility between the driver and CUDA. It’s unrelated to the fact that your device needs CUDA 11, as it has a compute capability of 8.6.

24

ok, I got it.
Many thanks for your prompt reply!

25

Hi @ptrblck, I’m working on an Azure Data Science VM that has CUDA 11.4 preinstalled (Nvidia driver version 470.82.00). I’m having an issue using Pytorch (1.10.0+cu113) on this machine where I get:

RuntimeError: CUDA error: an illegal memory access was encountered

Maybe it has something to do with the CUDA version mismatch, so I was wondering if a 1.10.0+cu114 build already underway?

26

In case you are seeing “random” CUDA errors, it could point to a setup issue, but usually the error is raised from the code itself. Could you post an executable code snippet, which would reproduce the issue, please?

27

Gents,

It is 2023 now, and Pytorch 2.0 is released with is out. We are stiill having loads of RTX 2080ti around and most of advanced generative require at the least torch 1.13 if not 2.0. Any plan for supporting Turing and the like ?

Thanks,
Steve

28

Turing “and the like” are still supported in all of our released binaries.