Remove obsolete script to install CUDA SDK under host_injections

[trz42]

After #772 we don't need the script /cvmfs/software.eessi.io/versions/2023.06/scripts/gpu_support/nvidia/install_cuda_host_injections.sh any longer. It is replaced with /cvmfs/software.eessi.io/versions/2023.06/scripts/gpu_support/nvidia/install_cuda_and_libraries.sh. Thus, it should be removed and the documentation for GPU support should be updated accordingly (using the new script). See also documentation issue EESSI/docs#327

[ocaisa]

Rather than a removal, maybe we just make a wrapper that calls the other script with the appropriate options to get the same end result

[TopRichard]

#965

[adammccartney]

The old script seems to offer an option to specify the CUDA version with "-c". The new script seems to tie itself to CUDA 12.1.1 and there is no obvious way to specify the version. Just wondering why the version option was removed?

[ocaisa]

That option may not do what you think it does. The applications that EESSI ships are built against a specific CUDA version, using a different version does not change what the applications are linked against. If you want a specific CUDA version that is not available in EESSI, you don't need this script to have that (though there is perhaps a bug right now with an Lmod hook right now that makes this harder than it needs to be).

[adammccartney]

Thanks for the info and the clarification, it makes sense. I think we were maybe getting confused as we were approaching with the goal of building software with eessi-extend. I guess anyway running the script would be a necessary prerequisite for us if we intend to provide the software that eessi ships. In any case, it currently throws an error when it gets run, the error message prints a message about the relocation of all CUDA installations and module files being relocated to GPU-specific. I saw the workaround mentioned elsewhere, this worked fine for the manual install of 12.6.0. I guess the script could be copied and patched with the workaround also. I don't really follow exactly what the implications of the relocation are, but it seems like using the workaround installs using the path.

We're learning a lot about eessi, and easybuild at the moment. It seems like we have a few interesting use cases that will probably school us a bit in how it all works. We've just been playing around with various types of build environment to familiarize ourselves with what's going on. Was able to follow a couple of the issue threads related to the workaround for the bug that you mentioned when installing a specific cuda version. Anyway, thanks for the info and keep up the good work!