Missing `torchvision` for JP60 images running CUDA 12.2

[IamShubhamGupto]

Hey Dusty,

I am in the process of building my own container and have used your build.sh to create a container for Jetson Orin nano running JP60 preview. However inside the container I have pytorch version 2.1.0 however it did not install torchvision. When I try to manually pip install it, it is incompatible as pytorch hasn't released support for CUDA 12.2.

I verified I am running cuda 12.2 in the container using nvcc --version.

Installing torch and torchvision with cuda 12.1 pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121 gives torch.cuda.is_available() as False

The nightly builds fails to resolve a version pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121

Is there a way to downgrade cuda to 12.1 in the container or provide support for torchvision with cuda 12.2 ?

Thank you!

[dusty-nv]

Hi @IamShubhamGupto, you can use my l4t-pytorch container which includes PyTorch, torchvision, and torchaudio and has container image for it up on DockerHub:

• https://github.com/dusty-nv/jetson-containers/tree/master/packages/l4t/l4t-pytorch
• dustynv/l4t-pytorch:r36.2.0

Or if you are building your own container using build.sh, just add torchvision to your list of packages. Any aarch64 wheels from PyPi or pytorch.org won't be built with CUDA enabled, which is why I build them from source.

[IamShubhamGupto]

hey @dusty-nv thanks for the quick response, I will give that a try and post an update. I wasn't aware I had to compile torchvision manually.

Thanks

[IamShubhamGupto]

update, adding torchvision to the required packages solved the issue