To allow compiled in virtuals we need to only build for actual sm archs.

CUDA currently doesn't support building for compute_ and having compiled in virtuals ( using separable compilation ). So we need to transition everything over to sm_