Jetson AGX Xavier Jetpack 5 upgraded to CUDA 11.8 can't use torch :(

Hello,

Please please may I have a python wheel for pytorch using Jetpack 5 with CUDA 11.8 for python 3.8 on Jetson AGX Xavier arch64.

I’m trying to use pytorch and torch audio but torch won’t work as there isn’t a version of pytorch out there compiled for CUDA 1.8 with jetpack 5. I would upgrade Jetpack but it can’t go past 5.

Thanks

Hi,

Jetpack 5 includes CUDA 11.4, and the wheels are compiled for the default CUDA version. If you need a custom CUDA version, you will need to compile Torch from source yourself. You can check the wheels provided by NVIDIA here:

1 Like

Hi @miguel.taylor ,

Thanks so much for your quick reply! I need CUDA 11.8 for installing the CUDA ONNX runtime Execution Provider. I have been able to upgrade CUDA to 11.8 using this blog. However building the onnxruntime wheel from source repo is tricky as it requires manually changes to the code for the patch to be applied as detailed here under the ‘build from source’ dropdown.

Would anyone be able to build this wheel for me with the patch applied?

P.s. I forgot to specify the torch and torchaudio version, 2.1.0 would be great

Thanks

Hi,

We have ONNXRuntime prebuilt for JetPack.
Have you tried it?

https://elinux.org/Jetson_Zoo#ONNX_Runtime

Thanks.

Hi @AsataLLL,

I have but no luck there either. I could install it ok but couldn’t use the CUDA Execution provider.

Thanks for the suggestion though.