Version '8.9.5.29-1+cuda12.1' for 'libcudnn8' was not found

[FlatMapIO]

It looks like the cudnn install script needs to be changed to accommodate different version combinations.

Here is the discussion: https://forums.developer.nvidia.com/t/e-version-8-3-1-22-1-cuda10-2-for-libcudnn8-was-not-found/200801/30

config file:

{
  "$schema": "https://raw.githubusercontent.com/devcontainers/spec/main/schemas/devContainer.base.schema.json",
  "name": "pytorch-devcontainer",
  "image": "mcr.microsoft.com/devcontainers/base:ubuntu-22.04",
  "features": {
    "ghcr.io/devcontainers/features/nvidia-cuda:1": {
      "cudaVersion": "12.1",


      "installCudnn": true,
      // https://developer.nvidia.com/rdp/cudnn-download
      "cudnnVersion": "8.9.5.29"
    },
}

[FlatMapIO]

I think it is unnecessary to specify the cudnn version externally. I use this method to find the highest matching version:

DISTRO=ubuntu2204
ARCH=x86_64

CUDA_VERSION=${CUDAVERSION:-12.1}

CUDNN_VERSION=$(apt-cache policy libcudnn8 | grep "$CUDA_VERSION" | grep -Eo '^[^-1+]*' | sort -V | tail -n1)

[samruddhikhandale]

Hi 👋

Currently, the nvidia-cuda Feature expects cudaVersion and cudnnVersion to install tools correctly (ensure compatabolity). However, I agree that we should update it to automatically find a compatible version.

@FlatMapIO would you be interested in contributing a PR with the enhancement changes? We'd appreciate that, thanks!

[prathameshzarkar9]

Hi @samruddhikhandale ,
Created PR #1126