Install pytorch 2.6 and torchvision 0.21

Robotics & Edge Computing Jetson & Embedded Systems Jetson AGX Xavier
,
1

Hello.
I need to install pytorch 2.6 and torchvision 0.21 on my Jetson AGX Xavier.
The latest Jetpack is 5.15 with Cuda 11.4.
I tried to build pytorch but the build failed because cuda 11.4 do not support bfloat16 and the mininal version of cuda should be 11.8.
I found that I can install cuda 11.8 on Xavier (CUDA Toolkit 11.8 Downloads | NVIDIA Developer)
by run the next commands:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/arm64/cuda-ubuntu2004.pin
sudo mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda-tegra-repo-ubuntu2004-11-8-local_11.8.0-1_arm64.deb
sudo dpkg -i cuda-tegra-repo-ubuntu2004-11-8-local_11.8.0-1_arm64.deb
sudo cp /var/cuda-tegra-repo-ubuntu2004-11-8-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get -y install cuda

and can install cudnn. The last version of cudnn 9.8.0 does not support cuda 11 for jetson platform:

So what version of cudnn I should choose?

Another option is to install cuda 12.2:

But when I try to install cudnn it installs cudnn for last cuda (12.8) Is there workaround to fix this?

One more thing I found that there is prebuild version of pytorch 2.2 for cuda 12.2 from jetpack 6.0 (PyTorch for Jetson)
Will it work if I succesfully install cuda 12.2 with cudnn?

3

Hi,

You can find the latest Pytorch version we provided for JetPack 5 below. It is v2.2:

For other versions, you will need to build it on your own.
On JetPack 5, only CUDA is upgradable and the latest version is 12.2 as you mentioned.
There are no cuDNN and TensorRT packages for aarch64-jetson in Ubuntu 20.04 available.

Thanks.

4

Thanks for your fast reply.
so if I need pytorch 2.6 I’ll need to build it from source. As I mentioned I can not build it with Cuda 11.4 because of errors

You mean that for cuda 12.2 there is no cuDNN and TensorRT?
I’ve checked cuDNN. and for jetson there is 9.8.0:

Can I use it for cuda 12.2?

Another question:
Can I build and install new versions of Ubuntu like 22.04 or 24.04 for Jetson AGX Xavier

5

Hi

Sorry for the non clear statement.

On JetPack 5, there is no ugpradable cuDNN and TensorRT.
Which means the available cuDNN and TensorRT comes from JetPack and built with the default CUDA 11.4.
The link you shared is a known issue and didn’t work:

We don’t test PyTorch 2.6 with CUDA 12.2 before.
But you can give it a try to see if it works.

Unfortunately, to use different Ubuntu versions on Xavier is not officially supported.
Thanks.

7

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