How to install Pytorch 1.7 with cuDNN 10.2?

Autonomous Machines Jetson & Embedded Systems Jetson Xavier NX
1

Description

Hello,

What are the commands needed to install pytorch 1.7 with torchvision 0.8.1 for cuDNN 10.2 in Jetson Xavier NX?

I read in this forum post PyTorch for Jetson that 1.7 is available but from past experiences without matching the correct pytorch version with torchvision and cuDNN, running a computer vision model would generate the the Segmentation Fault (core dumped) issue.

At the moment I am running Pytorch 1.6 with torchvision 0.7 but I am quite sure there is a compatibility issue as after executing the CNN model in pytorch, the program does not close correctly and shows a Segmentation Fault (core dumped) error. I read in a different post that, this issue could be caused by Pytorch not coupled with the correct torchvision library and/or cuDNN version.

As I am relatively new to using the ubuntu environment and using jetson nx, I could not make sure if the commands listed above forum would install pytorch 1.7 with the correct version of torchvision and will work in cuDNN 10.2

With thanks,
Azmyin

Environment

** Jetson Xavier NX running Jetpack 4.4.1
** OS - Ubuntu 18.04 LTS

3

Hi,

The PyTorch in the topic you shared is built from the source on a specific JetPack version.
So if you install the corresponding package, you won’t meet a compatible issue between PyTorch and CUDA.

For JetPack 4.4.1, please install PyTorch v1.7 with torchvision v0.8.1.
You can also use our NGC container directly: l4t-pytorch or l4t-ml.

Thanks.

4

@AastaLLL Thank you very much for the suggestion. I was able to install both of them successfully. Now the segmentation error fault no longer causes an issue.

With best,
Azmyin

5

Good to know this!
Thanks for the feedback : )