Jetson AGX Xavier Pytorch Wheel files for latest Python 3.8/3.9 versions with CUDA 10.2 support

NVIDIA forums had wheel files for Python 3.6 versions for installing pytorch with cuda support and they worked fine, but slowfast that I’m using in the project requires python versions 3.8/3.9 hence the versions are not compatible. Could someone please provide wheel files for torch/pytorch version 1.8 with cuda 10.2 support for Jetson AGX Xavier with python 3.8/9 version? Tht’d be helpful.

I have tried the suggestions below for Jetson Nano, but it didn’t work for me.

I tried the wheel from the Google link, however here is an error? have anyone met this problem? Looking forward to your reply!

The error indicates that PyTorch wheel was built with support for Nano/TX1 GPU (sm_53) but not the Xavier GPU (sm_72), so it would need rebuilt to include support for sm_72.

You can find the instructions for building PyTorch wheels in the first post in this thread (this is tested against Python 3.6)

Since you are using AGX Xavier, that is the fastest Jetson to build these wheels on. FYI here is another topic about building PyTorch for Python 3.8:

@dusty_nv Thank you for your quick answer! I have checked it carefully. I find that the pytorch version I need is 1.8 or later rather than 1.7. Could you please offer the pytorch 3.8 wheel file for Python3.8 with CUDA10.2 for AGX XAVIER as it’s difficult for me to build from source.

How can I rebuid the wheels to include support for sm_72? I used to type “pip install torch-…whl” to install pytorch. Do you mean buliding from source?

Looking forward to your kind reply!

I’m sorry, I don’t personally have the bandwidth to build and support the PyTorch wheels for multiple versions of Python. When Xavier moves to Ubuntu 20.04 in JetPack 5, I will build the wheels for Python 3.8 because that is the default version of Python for Ubuntu 20.04 (the default version of Python for Ubuntu 18.04 is Python 3.6). In the meantime, perhaps someone from the community has Python 3.8 wheels to share for PyTorch.

To build PyTorch wheel with support for sm_72, follow the instructions under Build from Source in this post (except you will want to invoke python3.8 in the build step instead) https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-10-now-available/72048

The one environment variable that is set export TORCH_CUDA_ARCH_LIST="5.3;6.2;7.2" will make it so it’s built for Xavier/Nano/TX1/TX2.

1 Like

Thank you very much! I finally successfully build pytorch and torchvision from source.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.