diff --git a/paddle/operators/recv_op.cc b/paddle/operators/recv_op.cc
index 731e5e4756a5d6656e001d93e34105d8b81f6d2f..9c3e8953bb781079c5fbca611182d083350c2ea3 100644
--- a/paddle/operators/recv_op.cc
+++ b/paddle/operators/recv_op.cc
@@ -98,7 +98,8 @@ class RecvOp : public framework::OperatorBase {
         auto *merged_grad = recv_scope.FindVar(grad_var_name);
         if (merged_grad == nullptr) {
           // create output of merged var.
-          recv_scope.Var(grad_var_name);
+          auto merged_var = recv_scope.Var(grad_var_name);
+          merged_var->GetMutable<framework::LoDTensor>();
         }
 
         if (trainer_count > 1) {
diff --git a/python/paddle/v2/fluid/distribute_transpiler.py b/python/paddle/v2/fluid/distribute_transpiler.py
index 4919dce20d560e46358e59c716d305b8356e6168..13006bfd137b5cbb0109925b74fa7f405abba01c 100644
--- a/python/paddle/v2/fluid/distribute_transpiler.py
+++ b/python/paddle/v2/fluid/distribute_transpiler.py
@@ -66,6 +66,24 @@ class DistributeTranspiler:
             Use different methods to split trainable varialbles to different
             parameter servers.
 
+            Example to run:
+
+            exe = fluid.Executor(place)
+            t = fluid.DistributeTranspiler()
+            t.transpile(optimize_ops, params_grads, pservers="127.0.0.1:6174", trainers=1)
+
+            pserver_endpoint = os.getenv("PSERVER")
+            if pserver_endpoint:
+                pserver_prog = t.get_pserver_program(pserver_endpoint, optimize_ops)
+                exe.run(fluid.default_startup_program())
+                exe.run(pserver_prog)
+            else:
+                feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
+                exe.run(fluid.default_startup_program())
+
+                for pass_id in range(PASS_NUM):
+                    ...
+
             :param optimize_ops: op list of optimization, should be the
                                  return value of Optimizer.minimize
             :type optimize_ops: list