AFAIK, nvidia-smi shows the max supported CUDA version.
cuda-toolkit will not install a newer driver. Other then a complete CUDA installation (which is not recommended due to non distro packaged driver).
Just for the record and for the use of coming-by readers of this thread:
I installed 11.2.0 via the runfile installer, avoiding the package manager. It seems to work flawlessly, I built ‘deviceQuery’ and ‘fluidsGL’ from the samples and they ran ok. Also my application asxp (Algebraic Surface eXPlorer) (available on GitHub), which was the ultimate purpose of installing CUDA, compiled and ran perfectly.
The runfile, upon executed, displayed a simple ncurses (text based pseudo GUI) interface where it was possible to unselect the option to install the drivers.
So I installed everything but the drivers and therefore could do all selection in my X terminal, no need to reboot and switch to runlevel 3. This is only necessary if one intends to install also the drivers contained in the runfile. (By the way, they had a smaller version number than my currently installed ones).
I think using the runfile is for me the better option, because
If one installs via the package manager, one gets update recommendations automatically, which one has to decline explicitly, if one does not want the local CUDA installation to be a “moving target”.
If one follows the official path of install via deb local (or net) and concludes with ‘apt install cuda’ one does, what I did not want, namely replace the drivers too (so one has, what I wanted to avoid “a complete CUDA installation”).
It seems there is clean way to uninstall everything if necessary very simple by executing an uninstall script. If one wants to deinstall using the package manager there is always the danger to deinstall too much or too less, if one works with wildcards in ‘apt remove’.
If you want to see a preview of the install process, look at
(the author of the video tries first to install via a deb-package but finally uses the runfile method).