Nvidia, CUDA and Fedora for ML

I know this has been asked previously and I apologize for repetition, but can someone resolve this issue.

I have installed fresh installation of Fedora and installed akmod-nvidia and then disabled nvidia-drivers to install cuda as given in this rpm fusion guide.

till this everything was normal until update for cuda that is cuda version 12.6 came. Now I am not able to update cuda, the output is shown below.

it also shows modular filetering issue. I have gone through the guide to understand the modular filterning issue.

But the thing is I dont want to disable cuda-fedora39 repo because there is one package that i want to install from that repo “nvidia-fs-dkms” which is important for “nvidia-gds” which is necessary for my project.

Also when i try to install nvidia-fs-dkms, i get error

skip broken also doesn’t work.

Can someone please help me with this issue.

Unless you specifically need the cuda toolkit, the instructions in that “guide” is out of date.

I have always been successful with using cuda by following the instructions here which is part of the nvidia How-To.

The modular filtering can be identified with dnf module list enabled and if it shows a modular item enabled (often nvidia-driver) that can be disabled with sudo dnf module disable nvidia-driver (or other module name that might interfere).

Also note that the guide you followed is for fedora 39. If you are using fedora 40 the proper repo would be obtained by replacing 39 with 40 everywhere that guide shows the version number.

Thank you Jeff for your help. Upon disabling “cuda-fedora39” repo, everything clears away and the system goes to normal. But the issue is, that repo has some important rpms that I need for my project like nvidia-gds and nvidia-fs-dkms as shown below

Now, I also tried downloading the cuda rpm from nvidia’s website and then installing the same. It again gave me the same error of modular filtering as shown below.

Can this be resolved?

As far as I can understand, I need “nvidia-driver” package to install the other packages I need but I have disabled nvidia-driver in order to install akmod-nvidia. I dont know what can i do in this case. I dont think that RPM Fusion nonfree has nvidia-fs-dkms or nvidia-gds packages in its repo.

and if you are wondering why I need those packages, well because nvidia-gds is the alternative to microsoft direct storage which I need for my ML programs.

Added cuda, f40, nvidia-solution

Your issue appears to be that you are installing from different repos for different purposes and there are conflicting packages and installation methods.

Thus you must decide what is most important to you and work within the restrictions that imposes.

I cannot assist if you must have the cuda-fedora39 repo and packages installed from there. The modular requirement does impact installing the akmod-nvidia driver from rpmfusion.

I would suggest that you find an alternative to the nvidia-gds package if at all possible – though I understand if not.

Just as a note:
Please copy & paste text from your screen using the preformatted text button </> on the toolbar instead of using images. The images cannot be searched and thus do not allow others to find the data included when they are searching for answers for their own questions.

Images also often truncate the output displayed.

1 Like

Thanks for you help.