NVRTC is fully supported in CUDA 7.5, which we're targeting. We should provide a way for users to generate kernels directly from Python via NVRTC.

See for example, @szagoruyko's RTC bindings for Lua Torch:
https://github.com/szagoruyko/nvrtc.torch
https://github.com/szagoruyko/cutorch-rtc