Xavier agx torch

[bsowislo]

I have Jetpack 5.13 but I do not find the correct torch whl.
I need Cuda 11 + Python 3.10. Any tries to compile + link from source (version 2.1.0) fails at 87% completed Are there any wheels or patches to solve this problem?

BR

[AastaLLL]

Hi,

Please follow the building steps and apply the corresponding patch in the below topic:

PyTorch for Jetson

Below are pre-built PyTorch pip wheel installers for Jetson Nano, TX1/TX2, Xavier, and Orin with JetPack 4.2 and newer. Download one of the PyTorch binaries from below for your version of JetPack, and see the installation instructions to run on your Jetson. These pip wheels are built for ARM aarch64 architecture, so run these commands on your Jetson (not on a host PC). You can also use the containers from jetson-containers. PyTorch pip wheels JetPack 6

PyTorch v2.2.0

JetPack 6.0 DP (L4T R3…

Thanks.

[bsowislo]

I have seen the patches for SDK4. I need help for Xavier SDK 5.13.
The SDK 6.0 whl is not supported for Xavier, the 2.2.0 Pytorch whl supports Cuda 12 and creates errors when it is invoked with Python3 import

[AastaLLL]

Hi,

Do you want to build PyTorch 2.2.0 on JetPack 5.1.3 with the default CUDA 11.4 and Python 3.10?

The combination should work if building PyTorch from the source.
Is there any error when building time?
What kind of error when importing?

Thanks.

[bsowislo]

Hi,
ok,
but as I said I have no valid torchvision version for 2.2.0

BR
Bernd

[AastaLLL]

Hi

Based on the below link, it should be 0.17.0.
Did you meet any error when building it?

GitHub - pytorch/vision: Datasets, Transforms and Models specific to Computer...

Thanks.

[bsowislo]

Hi,
I thing I simply incremented torchvision as a guess to v0.17.1 for 2.2.0 and it did not work on Orin AGX but I am not sure after all my tries to compile a stalel version. Did you try?

BR
Bernd Sowislo

[bsowislo]

Hi,
I checked torchvision 0.17.1, it works with 2.2.0 and the installation tests (python3 import torch…) succeed

the problem python3.10 + cuda 11.4 on SDK 5.13 is not solved yet.

BR
Bernd

[bsowislo]

Hi,
torch 2.20 + torchvision 0.17.1 are installable but ex. yolov8 does not run

RuntimeError: Couldn’t load custom C++ ops. This can happen if your PyTorch and torchvision versions are incompatible, or if you had errors while compiling torchvision from source. For further information on the compatible versions, check GitHub - pytorch/vision: Datasets, Transforms and Models specific to Computer Vision for the compatibility matrix. Please check your PyTorch version with torch.version and your torchvision version with torchvision.version and verify if they are compatible, and if not please reinstall torchvision so that it matches your PyTorch install.

BR

Bernd Sowislo

[AastaLLL]

Hi,

Could you share more about the status?
Is the success of PyTorch+TorchVision built with Python 3.8?
Could you share the error when you built the package for Python 3.10?

If Python 3.8 is an option for you.
We have some containers that have Pytorch and TorchVision installed already:

NVIDIA L4T PyTorch | NVIDIA NGC

Thanks.

[bsowislo]

Hi,
I use an Xavier with newest sdkmanager image. The current python version in this distribution is 3.10, so with the current sdkmanager image for Xavier torchvision is impossible to use. Error is simply that the wheel image is denied due to wrong python version. Perhaps it would be possible to downgrade to python 3.8 (update-alternatives does not work) Python 3.10 must be totally removed and restored after installation of torchvision. But to create a wheel for 3.10 should be easier for nvidia

BR
Bernd

[AastaLLL]

Hi,

The latest JetPacks support Xavier is 5.1.3.
JetPack 5.1.3 is a Ubuntu 20.04 system with default Python 3.8.

We do provide PyTorch for Python 3.10 but it is only available for the Orin series with JetPack 6.
Thanks.