We should figure out a way to support users calling Base.sin instead of CUDAnative.sin. One possible option would be to hook inference, overriding method calls when operating under @cuda. The same mechanism could be used to dispatch on the hardware generation.