trying global seg offset type

cub/cub/device/dispatch/dispatch_segmented_sort.cuh

@@ -78,7 +78,7 @@ namespace segmented_sort
 // Type used to index within segments within a single invocation
 using per_invocation_segment_offset_t =  int;
 // Type used for total number of segments and to index within segments globally
-using global_segment_offset_t = ::cuda::std::int64_t;
+using global_segment_offset_t = int;
 
 template <typename Iterator, typename OffsetItT>
 class OffsetIteratorT