diff --git a/paddle/fluid/operators/split_selected_rows_op.h b/paddle/fluid/operators/split_selected_rows_op.h
index af64607fafc6544047714e731846a2440be219b8..1fef2b3d378c96d087118d0136885e7e29aa237c 100644
--- a/paddle/fluid/operators/split_selected_rows_op.h
+++ b/paddle/fluid/operators/split_selected_rows_op.h
@@ -72,10 +72,11 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel<T> {
     for (size_t i = 0; i < outs_rows_idx.size(); ++i) {
       auto rows_idx = outs_rows_idx[i];
       outs[i]->set_height(height_sections[i]);
+      auto dims = x->GetCompleteDims();
+      dims[0] = rows_idx.size();
+      outs[i]->mutable_value()->mutable_data<T>(dims, x->place());
+      outs[i]->mutable_rows()->clear();
       if (rows_idx.size() > 0) {
-        auto dims = x->GetCompleteDims();
-        dims[0] = rows_idx.size();
-        outs[i]->mutable_value()->mutable_data<T>(dims, x->place());
         for (auto idx : rows_idx) {
           outs[i]->mutable_rows()->push_back(idx - abs_sections[i]);
         }
@@ -98,6 +99,8 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel<T> {
           }
         }
       }
+      PADDLE_ENFORCE_EQ(rows_idx.size(), outs[i]->rows().size(),
+                        "rows should has the same size with tensor dim 0");
     }
   }
 };