diff --git a/paddle/fluid/operators/controlflow/fetch_v2_op.cc b/paddle/fluid/operators/controlflow/fetch_v2_op.cc
index 5a99dd695c02ba41a2dd3357d4caf0daa47adb23..81d23939f5a259460b07b57a5d2f1597d717c175 100644
--- a/paddle/fluid/operators/controlflow/fetch_v2_op.cc
+++ b/paddle/fluid/operators/controlflow/fetch_v2_op.cc
@@ -34,7 +34,7 @@ namespace operators {
 static void DeepCopy(const phi::DenseTensor &src_item,
                      const std::string &fetch_var_name,
                      phi::DenseTensor *dst_item) {
-  if (src_item.IsInitialized() && src_item.numel() > 0) {
+  if (src_item.IsInitialized()) {
 #ifdef PADDLE_WITH_MKLDNN
     // Conversion from MKL-DNN to Paddle
     if (src_item.layout() == phi::DataLayout::ONEDNN) {
@@ -58,9 +58,7 @@ static void DeepCopy(const phi::DenseTensor &src_item,
     paddle::framework::TensorCopySync(src_item, platform::CPUPlace(), dst_item);
 #endif
   } else {
-    // Not copy, if the src tensor is empty.
-    dst_item->clear();
-    dst_item->Resize({0});
+    VLOG(4) << "No copy";
   }
   dst_item->set_lod(src_item.lod());
 }
diff --git a/python/paddle/fluid/tests/unittests/standalone_executor/test_standalone_executor.py b/python/paddle/fluid/tests/unittests/standalone_executor/test_standalone_executor.py
index cfce6784ad52705e95a9042b94078bc9f2a70118..86060d6bf2de12797f1728d9abd77d38488639f9 100644
--- a/python/paddle/fluid/tests/unittests/standalone_executor/test_standalone_executor.py
+++ b/python/paddle/fluid/tests/unittests/standalone_executor/test_standalone_executor.py
@@ -378,6 +378,19 @@ class TestException(unittest.TestCase):
         )
 
 
+class TestFetchEmptyTensor(unittest.TestCase):
+    def test_fetch(self):
+        places = [paddle.CPUPlace()]
+        if paddle.fluid.core.is_compiled_with_cuda():
+            places.append(paddle.CUDAPlace(0))
+        for place in places:
+            with paddle.static.program_guard(paddle.static.Program()):
+                out = paddle.empty([3, 0])
+                exe = paddle.static.Executor(place)
+                res = exe.run(fetch_list=[out])
+            self.assertEqual(res[0].shape, (3, 0))
+
+
 class TestInplaceApiWithDataTransform(unittest.TestCase):
     def test_increment(self):
         if paddle.fluid.core.is_compiled_with_cuda():
diff --git a/python/paddle/fluid/tests/unittests/test_matrix_nms_op.py b/python/paddle/fluid/tests/unittests/test_matrix_nms_op.py
index 0d60e0edeaabf60873b3d2de767a3bf0d76e6982..24ac8253d4384acbb980e4cb13023c89b4b564fa 100644
--- a/python/paddle/fluid/tests/unittests/test_matrix_nms_op.py
+++ b/python/paddle/fluid/tests/unittests/test_matrix_nms_op.py
@@ -270,7 +270,11 @@ class TestMatrixNMSOp(OpTest):
         )
 
         empty = len(det_outs) == 0
-        det_outs = np.array([], dtype=np.float32) if empty else det_outs
+        det_outs = (
+            np.array([], dtype=np.float32).reshape([0, BOX_SIZE + 2])
+            if empty
+            else det_outs
+        )
         index_outs = np.array([], dtype=np.float32) if empty else index_outs
         nmsed_outs = det_outs.astype('float32')
 
diff --git a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py
index 44fe9939688b4bcc50bb600b8f0336220d3305ae..f6a1bac1a63563a186f5720aecf5908c6d7da2b5 100644
--- a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py
+++ b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py
@@ -626,7 +626,9 @@ class TestMulticlassNMS2Op(TestMulticlassNMSOp):
         det_outs = np.array(det_outs)
 
         nmsed_outs = (
-            det_outs[:, :-1].astype('float32') if len(det_outs) else det_outs
+            det_outs[:, :-1].astype('float32')
+            if len(det_outs)
+            else np.array([], dtype=np.float32).reshape([0, BOX_SIZE + 2])
         )
         index_outs = (
             det_outs[:, -1:].astype('int') if len(det_outs) else det_outs
@@ -696,7 +698,9 @@ class TestMulticlassNMS2LoDInput(TestMulticlassNMSLoDInput):
 
         det_outs = np.array(det_outs)
         nmsed_outs = (
-            det_outs[:, :-1].astype('float32') if len(det_outs) else det_outs
+            det_outs[:, :-1].astype('float32')
+            if len(det_outs)
+            else np.array([], dtype=np.float32).reshape([0, BOX_SIZE + 2])
         )
         index_outs = (
             det_outs[:, -1:].astype('int') if len(det_outs) else det_outs
@@ -768,7 +772,9 @@ class TestMulticlassNMS3Op(TestMulticlassNMS2Op):
         det_outs = np.array(det_outs)
 
         nmsed_outs = (
-            det_outs[:, :-1].astype('float32') if len(det_outs) else det_outs
+            det_outs[:, :-1].astype('float32')
+            if len(det_outs)
+            else np.array([], dtype=np.float32).reshape([0, BOX_SIZE + 2])
         )
         index_outs = (
             det_outs[:, -1:].astype('int') if len(det_outs) else det_outs