diff --git a/paddle/platform/for_range.h b/paddle/platform/for_range.h
index 5427aa28238d6b46eb72d1fb49303dce3d871d7d..694a66d9ac4eb6ad02daf1931806fa1287de7cab 100644
--- a/paddle/platform/for_range.h
+++ b/paddle/platform/for_range.h
@@ -62,7 +62,7 @@ struct ForRange<CUDADeviceContext> {
 
   template <typename Function>
   inline void operator()(Function func) const {
-    constexpr size_t num_threads = 1024;
+    constexpr int num_threads = 1024;
     int block_size = limit_ <= num_threads ? limit_ : num_threads;
     int grid_size = (limit_ + num_threads - 1) / num_threads;