Installing PyTorch Compatible with CUDA 11.8

[isakqwer]

I’ve converted version of packages.

JetPack : 5.1.2
CUDA : 11.4 → 11.8
cuDNN : 8.6.0. → 8.8.0
OpenCV : 4.5.2
Python : 3.8 → 3.9

1. pytorch release/2.0 + torchvision release/0.15
2. pytorch release/2.1 + torchvision release/0.16

Connecting to CUDA is possible with both versions, but when running the cifar10 CNN network model, the following error occurs:

“RuntimeError: GET was unable to find an engine to execute this computation”

It’s said that this issue arises because the CUDA version is not compatible with PyTorch. Can you guide me on how to resolve it?

[AastaLLL]

Hi,

Do you use our prebuilt package or building from the source?
We don’t have Pytorch prebuilt for the Python 3.9 and CUDA 11.8.

Thanks.

[isakqwer]

Hello,
I am using the source code for v2.0.0 and v2.1.0 from the PyTorch Github, and I’ve confirmed that it’s linked with CUDA

However, I am encountering an error in CIFAR-10 with CNN.

Is it a library linkage issue or a version issuer?

[AastaLLL]

Hi,

Do you build PyTorch with cuDNN enabled?

Since we don’t have a cuDNN package of CUDA 11.8 for Jetson.
You might link cuDNN which is built with CUDA 11.4 and cause some compatibility issues.

Thanks.