Cudnn error while running custom Yolov5 model on Xavier AGX using Pytorch+GPU

[heerah1493]

Hi

I have set the Nvidia Jetson Xavier agx device with default cuda 11.4, cudnn 8.4, tensorrt 8.4 and flowed the instructions in the below page to make pytorch run on GPU
https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html
and also verified torch.cuda.is_available(), it says TRUE

when I try to run my custom Yolov5l model, it starts to run the model on GPU but not successfully running. I am getting following error

RuntimeError: Unable to find a valid cuDNN algorithm to run convolution

[AastaLLL]

Hi,

Which PyTorch package do you install?
It looks like you are using JetPack 5.0.x.

So please find the compatible package below:
https://developer.download.nvidia.cn/compute/redist/jp/v502/pytorch/

Thanks.

[heerah1493]

Yes I have installed the Jetpack 5.0.2 compatible PyTorch package. But still the same issue

[AastaLLL]

Hi,

Not sure if PyTorch cannot find a suitable algorithm since Jetson’s memory is much limited.
Could you try to reduce the batchsize to see if it helps?

If the error still occurs, please share the complete log with us.
Thanks.