GPU not detected Win10 WSL2/Ubuntu24.04 Tensorflow2.13.1 CUDA 12.3 cuDNN 8.9.7. Has anyone got this to work?

[cormac.garvey1]

I am trying to follow this nvidia guide but cannot. (1. NVIDIA GPU Accelerated Computing on WSL 2 — CUDA on WSL 12.8 documentation)
I am trying to follow this tensorflow guide (Install TensorFlow with pip)

It all falls apart setting up specific older archived versions of cuda, cudnn on specific versions of ubuntu running on latest version of WSL (Nvidia GPU on WSL). Does anyone know of a more detailed guide out there? The nvidia article is not helpful as it does not take into account the specific versions of cuda and cudnn that must be used to get tensorflow V2.13.1 (running on windows/visual studio code) to work.
Has anyone got Win 10 WSL2 Tensorflow latest version to recognise the Nvidia GPU?
I thought the idea was to put all the Cuda cudnn libraries on ubuntu in WSL2, and the Windows nvidia graphics driver could somehow use them there (instead of installing cuda/cudnn natively on win10 which is no longer supported for latest tensorflow)

After some package support file hacking (nvidia - Installing CUDA on Ubuntu 23.10 - libt5info not installable - Ask Ubuntu), it appears that i got CUDA 12.3 installed on ubuntu 24.04 running in WSL2 on Windows 10 latest version. (I couldnt get anything to install correctly on ubuntu 22.04)
But i have not been able to make the following nvidia cuDNN install lines work for v 8.9.7 (I modified them but does anyone have a working version of this, for 8.9.7?). I dont know if it will even work on ubuntu24, as there is only a package available for ubuntu22.Is there a way around this or is this a hard stop? The objective is the use the GPU for tensorflow 2.13.1 onWindows 10. But Im starting to think this is simply not possible and i need a dedicated linux box. What are other folks doing? Any help most appreciated!!
Many thanks

`wget https://developer.download.nvidia.com/compute/cudnn/9.8.0/local_installers/cudnn-local-repo-ubuntu2404-9.8.0_1.0-1_amd64.debsudo dpkg -i cudnn-local-repo-ubuntu2404-9.8.0_1.0-1_amd64.debsudo cp /var/cudnn-local-repo-ubuntu2404-9.8.0/cudnn-*-keyring.gpg /usr/share/keyrings/sudo apt-get updatesudo apt-get -y install cudnn`

WSL version:

Linux version 5.15.167.4-microsoft-standard-WSL2 (root@f9c826d3017f) (gcc (GCC) 11.2.0, GNU ld (GNU Binutils) 2.37) #1 SMP Tue Nov 5 00:21:55 UTC 2024

[cormac.garvey1]

One thing to add. After a lot of messing around, i have GPU detected on WSL2/ Ubuntu but no GPU detected on the same PC, but in Windows10. I know that there is no native windows support after Tensorflow2.10 (which requires an older version of python than i use so cannot test). However there is supposed to be some paravirtualization protocol magic (Announcing CUDA on Windows Subsystem for Linux 2 | NVIDIA Technical Blog) that allows the WSL Linux GPU to be used from the Windows 10 VS Code Python IDE.
My ubuntu:

My windows10 VS Code/Python 3.11.6 IDE on the same PC:

Can anyone tell me if this is a normal setup ?! I would have expected that the GPU would be detected by tensorflow in Windows (maybe with a ‘CUDA GPU / paravrtualization’ detected or soemthing like that.

I want to develop in windows, not in ubuntu. But it looks like Windows WSL / Ubuntu is not a solution, despite all the hype about the paravirtualization protocol.

As a first timer coming to this, i would say its almost impossible to get the correct combination of windows 10, wsl2, windows nvidia driver, python version, windows tensorflow version, ubuntu on WSL2 version,cuda version , cudnn version, conda venv on ubuntu, python ubuntu version,python ubuntu tensorflow version etc etc to work together, and is basically not worth the hassle. Glad i didnt pay for a high end nvidia card! Because they are no longer of any value for doing machine learning on windows, at least with Tensorflow.