I know this was a topic before but I think it becomes relevant again with the CUDA-PYTHON facilities and the progression of CUDA into providing more and more device functionality.
Is there a possibility of using nvrtc with CuRand ? As I understand the cuLinkAddData could be used as long as we provide with the PTX for it.
Of course curand has both host and device APIs. I assume you are referring to the device API. For that, if it were me, I would use jitify. For example, AFAIK cupy uses this method (jitify) to enable support for curand device API in cupy user-defined kernels.