diff --git a/paddle/fluid/memory/allocation/allocator_facade.cc b/paddle/fluid/memory/allocation/allocator_facade.cc
index 44cd915b168153343d9d1e3a7896cf6d2309b978..3e9c2ba269fe2f0d3bf8f87da308883687f9b427 100644
--- a/paddle/fluid/memory/allocation/allocator_facade.cc
+++ b/paddle/fluid/memory/allocation/allocator_facade.cc
@@ -311,6 +311,10 @@ class AllocatorFacadePrivate {
 
   void RecordStream(std::shared_ptr<Allocation> allocation,
                     const gpuStream_t& stream) {
+    if (allocation->size() == 0) {
+      return;
+    }
+
     StreamSafeCUDAAllocation* stream_safe_cuda_allocation =
         dynamic_cast<StreamSafeCUDAAllocation*>(allocation.get());
     PADDLE_ENFORCE_NOT_NULL(stream_safe_cuda_allocation,
diff --git a/paddle/fluid/memory/stream_safe_cuda_alloc_test.cu b/paddle/fluid/memory/stream_safe_cuda_alloc_test.cu
index 286dcdba8f22fecc55ce487857fc1b434b87b546..52c3825053ca226de6a6058197dffaa9c23944f8 100644
--- a/paddle/fluid/memory/stream_safe_cuda_alloc_test.cu
+++ b/paddle/fluid/memory/stream_safe_cuda_alloc_test.cu
@@ -243,6 +243,27 @@ TEST(StreamSafeCUDAAllocInterfaceTest, GetAllocatorInterfaceTest) {
   CheckMemLeak(place);
 }
 
+TEST(StreamSafeCUDAAllocInterfaceTest, ZeroSizeRecordStreamTest) {
+  platform::CUDAPlace place = platform::CUDAPlace();
+  std::shared_ptr<Allocation> zero_size_allocation = AllocShared(place, 0);
+  EXPECT_EQ(zero_size_allocation->ptr(), nullptr);
+
+  gpuStream_t stream;
+#ifdef PADDLE_WITH_CUDA
+  PADDLE_ENFORCE_GPU_SUCCESS(cudaStreamCreate(&stream));
+#else
+  PADDLE_ENFORCE_GPU_SUCCESS(hipStreamCreate(&stream));
+#endif
+
+  EXPECT_NO_THROW(RecordStream(zero_size_allocation, stream));
+
+#ifdef PADDLE_WITH_CUDA
+  PADDLE_ENFORCE_GPU_SUCCESS(cudaStreamDestroy(stream));
+#else
+  PADDLE_ENFORCE_GPU_SUCCESS(hipStreamDestroy(stream));
+#endif
+}
+
 TEST(StreamSafeCUDAAllocInterfaceTest, GetStreamInterfaceTest) {
   platform::CUDAPlace place = platform::CUDAPlace();
   size_t alloc_size = 256;