Cudnn 7 with Jetpack 4.4

I just flashed my Xavier dev kit with the latest Jetpack 4.4, but I noticed that it comes with CUDNN 8.0 preview edition. The preview is not compatible with our codebase at the moment, what is the proper way to fall back to CUDNN 7.1? I would still like to keep Jetpack 4.4 because we do depend on the newer Tensor RT library.

Production version was recently released: JetPack 4.4 - L4T R32.4.3 production release - #11 by suhash

Yeah, but I’m still stuck requiring CUDNN 7.1 regardless :(

Sorry, misread your post and thought you were looking for CuDNN 8 production release. In case you need CuDNN 7 I think your only option is to go back to JetPack 4.3

Hi,

It’s recommended to use cuDNN v8.0 to make sure the compatibility for other package.
Would you mind to share the error you meet so we can share some information with you?

If cuDNN v7.x is essential, please use our previous JetPack package instead.

Thanks.

Yeah, we are going to use the latest jetpack 4.4 and just make it work with CUDNN 8.0. We were trying to compile pytorch, but it doesn’t work out of the box with CUDNN 8.

Hi @jakepoz, the latest PyTorch 1.6.0 branch (currently rc2) does build with cuDNN 8.0. I posted the wheel here:

https://forums.developer.nvidia.com/t/pytorch-for-jetson-nano-version-1-5-0-now-available/72048/392