I installed “deb (local)” cuda-repo-ubuntu1404-7-5-local_7.5-18_amd64.deb package on Kubuntu 14.04. The CUDA toolkit related packages allows as a dependency only nvidia-352 driver package.
So users who install CUDA from nvidia download site are stuck to nvidia-352 driver version and they can not upgrade the driver from repositories or PPA’s while maintaining CUDA installed. For example from here:
If you try to install any driver from that PPA, then CUDA related packages will be uninstalled.
Please rebuild CUDA packages you offer to download, allowing higher driver versions than nvidia-352 (or higher versions than the driver version that you choose include when creating packages).
I second that. Could not install CUDA 7.5 om OpenSUSE because I was using 361.42 of the G04 driver. They should at least provide CUDA compiled against their Long Lived Branch version of the driver.
Any nvidia responsible for creating packages read the forums? Or should I make my request somewhere else?
I see that is about to be released CUDA 8. Please you consider that I pose here when you creating CUDA 8 packages for Linux.
Thank you.
My suggestion would be to file a bug. Give an exact command sequence that you are following, and show the indication that things are not proceeding as desired.
And while I haven’t studied your report carefully, I will say that if you are using any repo other than what NVIDIA provides, it’s unlikely that you’ll get any response on your concern.
You should be using only the defined repos as indicated in the CUDA linux installation guide. That guide and method presupposes that you will want to make upgrades in the future:
I don’t think there’s any intention to allow people to install drivers from unknown repos. The NVIDIA driver can be packaged in a variety of different ways, with or without certain modules and libraries. The driver to support CUDA needs to be packaged with specific components.
From what I see in this guide it only talks about official repositories for each distribution, and there are no official package repositories provided by nvidia. And I can not find any official nvidia repository containing each new version of the graphic driver.
All that you say would be valid if nvidia would launch friendly graphics driver installers (deb, rpm) for each new version of the driver, just as it does for Windows. But for Linux nvidia launches only unfriendly .run file where users often damage the system when trying to install the graphic driver. So, ‘thankfully’ third party repositories exist (official from each distribution, and PPA’s in Ubuntu).
Anyway all this is deviating the real point, because the problem I mentioned also would happen with the official repositories of each distribution. The point is that for local/network CUDA installer packages, nvidia is creating wrong package dependencies, allowing only one version of the driver. If we want to be more exhaustive, for CUDA for developers packages nVidia does not even obliged to include the graphic driver in the local/network installer packages. Simply provide CUDA for developers packages in local installer with the correct dependencies so that the user can install any driver version from the repositories.