Build Pytorch 2.1 with Cuda 12.2 and Cudnn 8.9.5

[johnnynunez]

Hello,
I saw that with jetpack 5.1.2 i’ts easier to upgrade cuda. With jetpack 6.0 will come with cuda 12.1 I think to do compatibility with pytorch 2.1.0.
In cuda page I can see jetson repository to upgrade, but with cudnn and tensorrt not appear, so I need also cudnn to compile pytorch.

I’m doing my tutorial: Start with Jetson AGX Orin Developer Kit. - HackMD

All compile and install good in miniconda with python 3.11 but driver GPU is very old.

torch.cuda.is_available()
/home/jetson/miniconda3/envs/prueba/lib/python3.11/site-packages/torch/cuda/init.py:138: UserWarning: CUDA initialization: The NVIDIA driver on your system is too old (found version 11040). Please update your GPU driver by downloading and installing a new version from the URL: Official Drivers | NVIDIA Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver. (Triggered internally at /home/jetson/Projects/pytorch/c10/cuda/CUDAFunctions.cpp:108.)
return torch._C._cuda_getDeviceCount() > 0

Solved with:

export PATH=/usr/local/cuda-12.2/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-12.2/compat${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

How to build cudnn and tensorrt for cuda 12.2?

[AastaLLL]

Hi,

Only CUDA can be upgraded through the website.
For other libraries, ex. cuDNN or TensorRT, please wait for the next JetPack release.

Thanks.

[johnnynunez]

It would be interesting that in both TensorRT and Cudnn pages we have the Jetson section as with CUDA, like x86 or armbsa. That said, it is because some recent models are not backwards compatible. Although the Jetson is an industrial device, it would be interesting to receive the same frequency of updates as the GPUs.

Ubuntu 24.04 is coming out in April next year. I expect a Jetpack 7.0 with this version in mid 2024.

[AastaLLL]

Hi,

To follow the future release, please check the below page:

Jetson Roadmap

Thanks.