Proprietary package cuda-drivers from NVIDIA comes with a fixed rich dependency (kmod-nvidia-latest-dkms = $epoch:$version or kmod-nvidia = $epoch:$version). Since we can’t change that requirement and our OpenRM package is named kmod-nvidia-open, it wouldn’t match the kmod-nvidia = $epoch:$version on its own. To work around this, we added Provides: kmod-nvidia = $epoch:$version to kmod-nvidia-open so the DNF solver would accept it as satisfying the kmod-nvidia alternative.

That workaround is now breaking our intended behavior: kmod-nvidia-open is designed to be parallel-installable across versions, but it also conflicts with kmod-nvidia. In RPM terms, conflicts are checked against capabilities, and after adding the virtual provide, every installed kmod-nvidia-open version also “is” a provider of kmod-nvidia. This makes different versions of kmod-nvidia-open appear to conflict with each other via the shared provided capability, so multiple versions can no longer be installed together.

What needs to change is to remove the Provides: kmod-nvidia from kmod-nvidia-open. That restores parallel-installability by ensuring kmod-nvidia-open versions don’t all provide the same conflicting capability. The tradeoff is that kmod-nvidia-open will no longer satisfy the kmod-nvidia = $version branch of cuda-drivers by name alone, so compatibility with cuda-drivers must be handled some other way.