diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc
index a070b8efb86088532daf304cd80ff7b8c608ac42..70ec6e90a4d0106b7f838e51b8357798daa4b10d 100644
--- a/paddle/fluid/framework/executor.cc
+++ b/paddle/fluid/framework/executor.cc
@@ -66,7 +66,7 @@ void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope*>>();
+    var->GetMutable<std::vector<framework::Scope>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
diff --git a/paddle/fluid/framework/feed_fetch_method.cc b/paddle/fluid/framework/feed_fetch_method.cc
index 3e9353f5cf67d8de62c5551f12ea786e49190549..8e1f93c5ebd448903d70f9668539e077875836e4 100644
--- a/paddle/fluid/framework/feed_fetch_method.cc
+++ b/paddle/fluid/framework/feed_fetch_method.cc
@@ -27,7 +27,8 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
   // be created.
   VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
   Variable* g_feed_value = scope->Var(var_name);
-  auto& feed_inputs = *(g_feed_value->GetMutable<FeedFetchList>());
+  auto& feed_inputs =
+      *(g_feed_value->GetMutable<std::vector<paddle::framework::LoDTensor>>());
   if (index >= feed_inputs.size()) {
     feed_inputs.resize(index + 1);
   }
diff --git a/paddle/fluid/framework/naive_executor.cc b/paddle/fluid/framework/naive_executor.cc
index 2840d503f1454271afb309efdd435225ab077dc0..ba10687d65cfbbac89cfc76879c8b202ebd03229 100644
--- a/paddle/fluid/framework/naive_executor.cc
+++ b/paddle/fluid/framework/naive_executor.cc
@@ -37,7 +37,7 @@ static void InitializeVariable(Variable *var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope *>>();
+    var->GetMutable<std::vector<framework::Scope>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
diff --git a/paddle/fluid/framework/variable.h b/paddle/fluid/framework/variable.h
index 873e1b20a584df3ba90cf5c1a62a3879bf98ce5c..067e0c2b8389f88639fd9b95bd680702517efee1 100644
--- a/paddle/fluid/framework/variable.h
+++ b/paddle/fluid/framework/variable.h
@@ -38,12 +38,8 @@ class Variable {
 
   template <typename T>
   T* GetMutable() {
-    if (!holder_) {
+    if (!IsType<T>()) {
       holder_.reset(new PlaceholderImpl<T>(new T()));
-    } else {
-      PADDLE_ENFORCE(IsType<T>(),
-                     "Variable must be type %s, the holding type is %s",
-                     typeid(T).name(), holder_->Type().name());
     }
     return static_cast<T*>(holder_->Ptr());
   }
diff --git a/paddle/fluid/framework/variable_test.cc b/paddle/fluid/framework/variable_test.cc
index 003dcfd3dfe5ecfd563a686bb72b061aff602f73..c5c1d215f4a6affae0a3bdafacec40a2aee2ca19 100644
--- a/paddle/fluid/framework/variable_test.cc
+++ b/paddle/fluid/framework/variable_test.cc
@@ -33,10 +33,9 @@ TEST(Variable, GetMutable) {
   const Tensor& tt = v->Get<Tensor>();
   EXPECT_EQ(1234, tt.content_);
 
-  try {
-    v->GetMutable<std::string>();
-  } catch (std::exception& e) {
-    return;
-  }
-  EXPECT_TRUE(false);
+  std::string* s = v->GetMutable<std::string>();
+  *s = "hello";
+
+  const std::string& ss = v->Get<std::string>();
+  EXPECT_EQ("hello", ss);
 }
diff --git a/python/paddle/fluid/tests/book/test_word2vec.py b/python/paddle/fluid/tests/book/test_word2vec.py
index 1f3a230048f0bee8b8aa0d6384f0c78b27973cab..9191f0fc2037d32159c66ca64911864efb34ae30 100644
--- a/python/paddle/fluid/tests/book/test_word2vec.py
+++ b/python/paddle/fluid/tests/book/test_word2vec.py
@@ -17,6 +17,7 @@ from __future__ import print_function
 import paddle
 import paddle.fluid as fluid
 from paddle.fluid.layers.device import get_places
+from paddle.fluid.layers.control_flow import ParallelDo
 import unittest
 import os
 import numpy as np
@@ -83,7 +84,18 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True):
         avg_cost, predict_word = __network__(
             [first_word, second_word, third_word, forth_word, next_word])
     else:
-        raise ValueError('is_parallel=True not implemented')
+        places = get_places()
+        pd = ParallelDo(places)
+        with pd.do():
+            avg_cost, predict_word = __network__(
+                list(
+                    map(pd.read_input, [
+                        first_word, second_word, third_word, forth_word,
+                        next_word
+                    ])))
+            pd.write_output(avg_cost)
+
+        avg_cost = fluid.layers.mean(pd())
 
     sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
     sgd_optimizer.minimize(avg_cost)
@@ -250,7 +262,7 @@ def inject_test_method(use_cuda, is_sparse, is_parallel):
 
 for use_cuda in (False, True):
     for is_sparse in (False, True):
-        for is_parallel in (False, ):  # TODO(paddle-dev): Add parallel test.
+        for is_parallel in (False, True):
             inject_test_method(use_cuda, is_sparse, is_parallel)
 
 if __name__ == '__main__':