I installed cuda toolkit cudnn with debian And it is clearly installed. But When I commanded ‘jtop’, The cudnn version shows ‘1.0.0’
My cuda version is 11.8 and cudnn is `8.9.7’ installed
I don’t know why that happens… Is it related to the cuda version compatibility? And is there issue like me?
Also when i ran the example code Installation Guide :: NVIDIA cuDNN Documentation
The error occured.
Hi,
On JetPack 5, only CUDA is upgradable.
So if you use pre-installed cuDNN with CUDA 11.8, it will cause some error since the package is built for CUDA 11.4.
Thanks.
It is not pre-installed, but I installed manually cuda, cudnn and I downgraded cuda 11.4 and cudnn 8.6.0 it shows
cudnn 1.0 version in jtop and jetson_release. But When I checked it the file installed in the local, 8.6.0is correctlly installed .
Hi,
Could you share the detailed link and commands you use for manual installation with us?
Thanks.
Hi,
I just upgraded to Jetpack 6 on my Orin NX 16 GB version. After flashing, I am installing CUDA, cuDNN, TensorRT manually, one-by-one (by using commands available on respective NVIDIA’s official downloads pages). I have installed CUDA-12.4 and cuDNN-9.4.0. When trying to verify in the jtop utility, the CUDA version is detected correctly, whereas the cuDNN version shows up as 1.0. More than that, I have installed from Pytorch 2.3.0 for CUDA 12.4 from pip wheel, following the instructions on https://forums.developer.nvidia.com/t/pytorch-for-jetson/72048. When trying to import this, it is actually looking for libcudnn.so.8. So, it would be great if you could guide us on the compatibility of the three things: Pytorch2.3 or higher, CUDA and cuDNN versions. TIA!
