Easily selecting compatible PyTorch version for python version and GPU

1

Hi, I made page PyTorch install command picker which shows you all compatible pytorch versions for the python version and GPU you have, it takes compute capability into account, and it’s based on reading wheel filenames and parsing the compute capability from github.
I hope you’ll find it useful.

2

Really cool, thanks for sharing. It would be nice to have a link to this website from https://pytorch.org/get-started/locally/ in my opinion.

4

Thanks for sharing this! Generally a good idea but unfortunately the default wheels installable via pip install torch are already showing wrong information:

torch 2.13.0 CUDA wheels are built for NVIDIA compute capabilities 5.0 (Maxwell) to 12.0 (Blackwell). GPUs older than CC 5.0 have no supported wheel for this release;

CUDA 13.0 dropped architectures up to sm_75 (Turing) so sm_50 is not and can not be supported in these builds.

GPUs newer than CC 12.0 need a later torch release.

This is also wrong since e.g. DGXSpark is using sm_121 and is binary compatible with sm_120 and works today.

5

Thanks for pointing it out, I fixed it and now the information should be correct.