CUDA repo issue with nvidia driver on Fedora and EL


There is an issue with the way the nvidia driver is packaged in the nvidia maintained CUDA repository compared to well-known community or even personal repositories.

As a package maintainer experience, it’s very well known problem when two repositories provide the same
set of packages names. Specially when the update method between the two repositories differs.

The problem experienced by some rpm cuda users is that when the driver is updated on the nvidia cuda driver side, it will replace the community based driver. Same for the opposite.
This create some uncertainty on users.

Unfortunately the reason why the switch is a problem is not much because users are changing provider, but rather that:
1/ The nvidia driver from cuda repo often miss the i686 build that is required by some users for some (cuda or non-cuda) related tasks. This lead to packaging break in dependencies from users using the community package.
2/ The kernel module package often miss from cuda repo the latest patch to get compatible with lastest kernel. This more often strike Fedora users, but also EL(1) users using an upgraded kernel.

The only way I’m aware the problem can be fixed was to:
1/ provide the nvidia driver with a separate name not to conflict with the community/personal repo names.
2/ provide the driver in a repository separated from the cuda one on the nvidia side, so the community version will always be picked. (or the driver installed with the installer if users choose that method).

Unfortunately, the solution the cuda-10.0 repo just move the problem without solving it.
I’m looking forward any long term fix for this issue.

Thanks for your understanding.

(1) EL stands for RHEL, CentOS and derivates.

For now Installing the cuda-10.0 repo using the well-known 3rd part RPM Fusion repository for Fedora can be done using as root:

dnf download cuda-drivers-410.48-1 cuda-runtime-10.0
rpm -Uvh cuda-drivers-410.48-1* cuda-runtime-10.0* --nodeps
dnf install cuda-10.0