2

I tried to install NVIDIA cuDNN 8.3.2 following the official guide in Dockerfile:

ENV OS=ubuntu2004

RUN wget https://developer.download.nvidia.com/compute/cuda/repos/${OS}/x86_64/cuda-${OS}.pin

RUN mv cuda-${OS}.pin /etc/apt/preferences.d/cuda-repository-pin-600
RUN apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/${OS}/x86_64/7fa2af80.pub
RUN add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/${OS}/x86_64/ /"
RUN apt-get update

ENV cudnn_version=8.3.2.1
ENV cuda_version=cuda11.5

RUN apt-get install libcudnn8=${cudnn_version}-1+${cuda_version}
RUN apt-get install libcudnn8-dev=${cudnn_version}-1+${cuda_version}

and I am getting the error: E: Version '8.3.2.1-1+cuda11.5' for 'libcudnn8' was not found The command '/bin/bash -o pipefail -c apt-get install libcudnn8=${cudnn_version}-1+${cuda_version}' returned a non-zero code: 100

I visited recommended support matrix for versions compatibility, but unfortunately didn't find appropriate instructions. Does anybody know which cudnn and cuda versions are needed here? Is there a link to REPOs where I can see available packages, like this one?

Bar
  • 2,736
  • 3
  • 33
  • 41
Alexander Borochkin
  • 4,249
  • 7
  • 38
  • 53

2 Answers2

4

I found necessary repos in Internet on official sites for developers:

cuda repos for different OS: https://developer.download.nvidia.com/compute/cuda/repos/

libcudnn8 repos for different cuda versions: https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/

onnx execution providers compatibility matrix for cuda and cuDNN: https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html

So, in my case the answer was:

ENV cudnn_version=8.2.4.15
ENV cuda_version=cuda11.4
Alexander Borochkin
  • 4,249
  • 7
  • 38
  • 53
1

Well, using @Alexander provided link for cuda - cudnn compatibility:

if you search for libcudnn you can see that for your case (and taking into consideration the choices you made about cuda and libcudnn versions) are 4:
enter image description here

So, the following combination should work:

ENV cudnn_version=8.3.2.44
ENV cuda_version=cuda11.5

I placed the closer to the version you provided but I guess other options will work also.

Eypros
  • 5,370
  • 6
  • 42
  • 75