pytorch

bb0c15a7 - cuda function correctly handles nan and passes correct initi vals for floating and integer types, also added tests