It should be possible to install the nvidia kernel modules _only_. This is useful in the case of HPC machines where the card is actually only used in conjunction with, for example, dev-util/nvidia-cuda-sdk. I have successfully installed and used both dev-util/nvidia-cuda-sdk and x11-drivers/nvidia-drivers by using --nodeps. I would suggest the addition or re-use of a USE flag specific to this situation such as X (-X), cuda, cuda-nox, etc. Note that this is not a requirement for the CUDA SDK to work but rather a request to keep the install base of HPC machines streamlined ans slimmed down. Installing the x11-drivers/nvidia-drivers would currently pull in close to 57 (useless) packages. Reproducible: Always
If you code it up and it works both ways, I'll commit it. Otherwise this is a very small edge case feature request without a patch. read: I'll have to be really bored with nothing higher on my priority list to knock this one out. I'm not being rude, just being honest.
This actually doesn't seem hard, I just tried with "-K" option on Nvidia's script and it doesn't even error out since it doesn't look at any X dependencies. Will have a look at the ebuild an try to make it work fine with a -X flag that uses the -K parameter.