diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc
index 0ec18de5b8a0e7cebdb91c30d2b45596b02dfa51..2c13ddd8d01a9ff99c60ede5074f2a5209037e07 100644
--- a/paddle/framework/backward.cc
+++ b/paddle/framework/backward.cc
@@ -141,9 +141,27 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
         net->ops_[op_offset]->Rename(name, dup_outputs.back());
       }
       // collect all the offset to append `add` op for each alias
-      insert_position.push_back(
-          {dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}},
-                                               {{"Out", {name}}}, {})});
+      //
+      // one variable is shared between multiple operators.
+      // insert add operator one by one, then add it to output
+      for (size_t output_idx = 0; output_idx < dup_outputs.size() - 1;
+           ++output_idx) {
+        auto insert_add_x = dup_outputs[output_idx];
+        auto insert_add_y = dup_outputs[output_idx];
+        auto insert_add_out = name + "@SHARED@" + std::to_string(output_idx);
+        // first add op inserted
+        if (output_idx == dup_outputs.size() - 2) {
+          insert_add_out = name;
+        }
+        if (output_idx != 0) {
+          insert_add_y = name + "@SHARED@" + std::to_string(output_idx - 1);
+        }
+        insert_position.push_back(
+            {dup_op.back(),
+             OpRegistry::CreateOp(
+                 "sum", {{"X", {insert_add_x}}, {"X", {insert_add_y}}},
+                 {{"Out", {insert_add_out}}}, {})});
+      }
     }
 
     // make sure the inserted `add` ops follow the BFS order.
diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc
index 6932f5b989a3e21ebc44ec4fec9f5223f2547d7a..a36e7bde8c61c6a8bfc6eefd46ce5e09907476ef 100644
--- a/paddle/framework/backward_test.cc
+++ b/paddle/framework/backward_test.cc
@@ -133,15 +133,18 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker {
   }
 };
 
-class AddOpMaker : public OpProtoAndCheckerMaker {
+class SumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+  SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "x").AsDuplicable();
-    AddOutput("Out", "out");
+    AddInput("X", "the input tensors of sum operator.")
+        .AsDuplicable()
+        .NotInGradient();
+    AddOutput("Out", "the output tensor of sum operator.").NotInGradient();
     AddComment("");
   }
 };
+
 }  // namespace framework
 }  // namespace paddle
 
@@ -154,7 +157,7 @@ REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
 REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
 REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
-REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
+REGISTER_OP(sum, f::NOP, f::SumOpMaker, sum_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
 REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
             f::NOP);
@@ -283,7 +286,7 @@ TEST(Backward, net_shared_weight) {
   ASSERT_TRUE(bwd->IsNetOp());
   auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
   ASSERT_EQ(3UL, bwd_net->ops_.size());
-  ASSERT_EQ("add", bwd_net->ops_[2]->Type());
+  ASSERT_EQ("sum", bwd_net->ops_[2]->Type());
 }
 
 TEST(Backward, op_register_grad_not_for_network) {