diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec
index 2ce73df0248bee244665b033edddcea70407546d..46e56981ea57722bbc064304761e7ab7b7aee141 100644
--- a/paddle/fluid/API.spec
+++ b/paddle/fluid/API.spec
@@ -6,7 +6,7 @@ paddle.fluid.Program.create_block ArgSpec(args=['self', 'parent_idx'], varargs=N
 paddle.fluid.Program.current_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Program.get_desc ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Program.global_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.Program.inference_optimize ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Program.inference_optimize ArgSpec(args=['self', 'export_for_deployment'], varargs=None, keywords=None, defaults=(True,))
 paddle.fluid.Program.list_vars ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Program.optimized_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
 paddle.fluid.Program.parse_from_string ArgSpec(args=['binary_str'], varargs=None, keywords=None, defaults=None)
@@ -18,6 +18,9 @@ paddle.fluid.Operator.all_attrs ArgSpec(args=['self'], varargs=None, keywords=No
 paddle.fluid.Operator.attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.attr_type ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.block_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Operator.block_attr_id ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Operator.blocks_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Operator.blocks_attr_ids ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.has_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.has_kernel ArgSpec(args=['self', 'op_type'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.input ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
@@ -74,7 +77,7 @@ paddle.fluid.io.save_persistables ArgSpec(args=['executor', 'dirname', 'main_pro
 paddle.fluid.io.load_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.io.load_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.io.load_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
-paddle.fluid.io.save_inference_model ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename'], varargs=None, keywords=None, defaults=(None, None, None))
+paddle.fluid.io.save_inference_model ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename', 'export_for_deployment'], varargs=None, keywords=None, defaults=(None, None, None, True))
 paddle.fluid.io.load_inference_model ArgSpec(args=['dirname', 'executor', 'model_filename', 'params_filename'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.io.get_inference_program ArgSpec(args=['target_vars', 'main_program'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.initializer.ConstantInitializer.__init__ ArgSpec(args=['self', 'value', 'force_cpu'], varargs=None, keywords=None, defaults=(0.0, False))
diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc
index af26cf287205833ebdfe1c3971b760adbf5c0a68..122dc161b41246e5f08bd0ae8b763489e9ee22f9 100644
--- a/paddle/fluid/framework/op_desc.cc
+++ b/paddle/fluid/framework/op_desc.cc
@@ -297,7 +297,20 @@ Attribute OpDesc::GetNullableAttr(const std::string &name) const {
   }
 }
 
-int OpDesc::GetBlockAttr(const std::string &name) const {
+std::vector<int> OpDesc::GetBlocksAttrIds(const std::string &name) const {
+  auto it = attrs_.find(name);
+  PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);
+  auto blocks = boost::get<std::vector<BlockDesc *>>(it->second);
+
+  std::vector<int> ids;
+  for (auto n : blocks) {
+    ids.push_back(n->ID());
+  }
+
+  return ids;
+}
+
+int OpDesc::GetBlockAttrId(const std::string &name) const {
   auto it = attrs_.find(name);
   PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);
   return boost::get<BlockDesc *>(it->second)->ID();
diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h
index 6805d25934b42a5752cbb54174d3017cf63e4b23..2422392e24d864dc3e7973ab35e038ecf2c0392a 100644
--- a/paddle/fluid/framework/op_desc.h
+++ b/paddle/fluid/framework/op_desc.h
@@ -85,7 +85,9 @@ class OpDesc {
 
   Attribute GetNullableAttr(const std::string &name) const;
 
-  int GetBlockAttr(const std::string &name) const;
+  int GetBlockAttrId(const std::string &name) const;
+
+  std::vector<int> GetBlocksAttrIds(const std::string &name) const;
 
   void Rename(const std::string &old_name, const std::string &new_name);
 
diff --git a/paddle/fluid/framework/program_desc.cc b/paddle/fluid/framework/program_desc.cc
index 1e01a6e900404990e16674755367d2fc6d832725..20bdc7830f32564448a69e9cd76c02585b7a1aca 100644
--- a/paddle/fluid/framework/program_desc.cc
+++ b/paddle/fluid/framework/program_desc.cc
@@ -58,7 +58,7 @@ ProgramDesc::ProgramDesc(const ProgramDesc &o) {
       for (const std::string &attr_name : op->AttrNames()) {
         if (op->GetAttrType(attr_name) == proto::AttrType::BLOCK) {
           int sub_block_id =
-              o.Block(block_id).Op(op_id)->GetBlockAttr(attr_name);
+              o.Block(block_id).Op(op_id)->GetBlockAttrId(attr_name);
           op->SetBlockAttr(attr_name, MutableBlock(sub_block_id));
         }
       }
diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc
index d0286719b9ea1aa671294f519051ac1e269c4e93..652a6ec7a4e2e823b28f39b449570cd375e88e18 100644
--- a/paddle/fluid/platform/profiler.cc
+++ b/paddle/fluid/platform/profiler.cc
@@ -270,12 +270,13 @@ struct EventItem {
   double min_time;
   double max_time;
   double ave_time;
+  float ratio;
 };
 
 // Print results
 void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
                    const std::string& sorted_domain, const size_t name_width,
-                   const size_t data_width) {
+                   const size_t data_width, double total) {
   // Output header information
   std::cout << "\n------------------------->"
             << "     Profiling Report     "
@@ -300,7 +301,8 @@ void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
   std::cout << std::setw(name_width) << "Event" << std::setw(data_width)
             << "Calls" << std::setw(data_width) << "Total"
             << std::setw(data_width) << "Min." << std::setw(data_width)
-            << "Max." << std::setw(data_width) << "Ave." << std::endl;
+            << "Max." << std::setw(data_width) << "Ave."
+            << std::setw(data_width) << "Ratio." << std::endl;
   for (size_t i = 0; i < events_table.size(); ++i) {
     for (size_t j = 0; j < events_table[i].size(); ++j) {
       const EventItem& event_item = events_table[i][j];
@@ -309,7 +311,9 @@ void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
                 << std::setw(data_width) << event_item.total_time
                 << std::setw(data_width) << event_item.min_time
                 << std::setw(data_width) << event_item.max_time
-                << std::setw(data_width) << event_item.ave_time << std::endl;
+                << std::setw(data_width) << event_item.ave_time
+                << std::setw(data_width) << event_item.total_time / total
+                << std::endl;
     }
   }
   std::cout << std::endl;
@@ -359,6 +363,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
 
   std::vector<std::vector<EventItem>> events_table;
   size_t max_name_width = 0;
+  double total = 0.;  // the total time
   for (size_t i = 0; i < events.size(); i++) {
     std::list<Event> pushed_events;
     std::vector<EventItem> event_items;
@@ -379,6 +384,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
                                g_state == ProfilerState::kAll)
                                   ? rit->CudaElapsedMs(events[i][j])
                                   : rit->CpuElapsedMs(events[i][j]);
+          total += event_time;
 
           std::string event_name =
               "thread" + std::to_string(rit->thread_id()) + "::" + rit->name();
@@ -387,7 +393,8 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
           if (event_idx.find(event_name) == event_idx.end()) {
             event_idx[event_name] = event_items.size();
             EventItem event_item = {event_name, 1,          event_time,
-                                    event_time, event_time, event_time};
+                                    event_time, event_time, event_time,
+                                    0.};
             event_items.push_back(event_item);
           } else {
             int index = event_idx[event_name];
@@ -431,7 +438,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
   }
 
   // Print report
-  PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12);
+  PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12, total);
 }
 
 void DisableProfiler(EventSortingKey sorted_key,
diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc
index 7f06f7a9d72de909da1507d62054da4905084354..c2137ec6d7df24251432a4dfb8fffc3d3f77194e 100644
--- a/paddle/fluid/pybind/protobuf.cc
+++ b/paddle/fluid/pybind/protobuf.cc
@@ -296,7 +296,8 @@ void BindOpDesc(pybind11::module *m) {
              std::string ser(seriralized);
              self.SetAttr(name, ser);
            })
-      .def("block_attr", &pd::OpDesc::GetBlockAttr)
+      .def("block_attr_id", &pd::OpDesc::GetBlockAttrId)
+      .def("blocks_attr_ids", &pd::OpDesc::GetBlocksAttrIds)
       .def("check_attrs", &pd::OpDesc::CheckAttrs)
       .def("infer_shape", &pd::OpDesc::InferShape)
       .def("infer_var_type", &pd::OpDesc::InferVarType)
diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py
index aa62d344911cfdee82bf9b868ebed3e38eef80fb..3824b21ec298de6e42d91a128d9abd0cff69d0bc 100644
--- a/python/paddle/fluid/backward.py
+++ b/python/paddle/fluid/backward.py
@@ -344,7 +344,7 @@ def _append_backward_ops_(block,
         grad_sub_block_list = []
         # If the op has its own sub-block, deal with the sub-block first
         if op.has_attr("sub_block"):
-            sub_block = program.block(op.block_attr("sub_block"))
+            sub_block = program.block(op.block_attr_id("sub_block"))
             grad_sub_block = program.create_block()
             grad_sub_block._set_forward_block_idx(sub_block.idx)
             cb = _callback_lookup_(op)
@@ -406,7 +406,7 @@ def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
     for op_idx in range(start_op_idx, block.desc.op_size()):
         op_desc = block.desc.op(op_idx)
         if op_desc.has_attr("sub_block"):
-            sub_block = block.program.block(op_desc.block_attr("sub_block"))
+            sub_block = block.program.block(op_desc.block_attr_id("sub_block"))
             _append_backward_vars_(sub_block, 0, grad_to_var, grad_info_map)
         new_vars = set()
         # create new gradient variables
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index fbeb0e5940ee1f87935a4e4815c907efdf68a1ae..d413f96e95d1613577f24b4b32a3aa0239b6cc1e 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -477,23 +477,25 @@ class Operator(object):
                  attrs=None):
         self.block = block
         self.desc = desc
-        self.attrs = attrs
-        if self.attrs is None:
-            self.attrs = dict()
+        # note: not add self.attrs here:
+        # https://github.com/PaddlePaddle/Paddle/pull/12583#pullrequestreview-145093173
+        op_attrs = attrs
+        if op_attrs is None:
+            op_attrs = dict()
         del attrs
 
         op_maker = core.op_proto_and_checker_maker
 
-        if op_maker.kOpRoleAttrName() not in self.attrs:
-            self.attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role
+        if op_maker.kOpRoleAttrName() not in op_attrs:
+            op_attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role
 
         role_var_name = op_maker.kOpRoleVarAttrName()
         if len(self.block.program.
-               op_role_var) != 0 and role_var_name not in self.attrs:
-            self.attrs[role_var_name] = self.block.program.op_role_var
+               op_role_var) != 0 and role_var_name not in op_attrs:
+            op_attrs[role_var_name] = self.block.program.op_role_var
 
-        if role_var_name in self.attrs and len(self.attrs[role_var_name]) == 0:
-            del self.attrs[role_var_name]
+        if role_var_name in op_attrs and len(op_attrs[role_var_name]) == 0:
+            del op_attrs[role_var_name]
 
         if len(self.desc.type()) != 0:
             return
@@ -563,15 +565,14 @@ class Operator(object):
                     arg.op = self
                 self.desc.set_output(out_proto.name, out_arg_names)
 
-        if self.attrs is not None:
-            if not isinstance(self.attrs, dict):
+        if op_attrs is not None:
+            if not isinstance(op_attrs, dict):
                 raise TypeError("'attrs' should be a dict.")
             for attr in proto.attrs:
                 attr_name = attr.name
-                if (attr_name not in self.attrs) or (
-                        self.attrs[attr_name] is None):
+                if (attr_name not in op_attrs) or (op_attrs[attr_name] is None):
                     continue
-                attr_val = self.attrs[attr_name]
+                attr_val = op_attrs[attr_name]
                 self._update_desc_attr(attr_name, attr_val)
 
         self.desc.check_attrs()
@@ -719,7 +720,6 @@ class Operator(object):
         Raises:
             ValueError: If the type of value doesn't match with desc.attr_type(name).
         """
-        self.attrs[name] = val
         self._update_desc_attr(name, val)
 
     def _update_desc_attr(self, name, val):
@@ -761,9 +761,9 @@ class Operator(object):
         """
         return self.desc.attr(name)
 
-    def block_attr(self, name):
+    def block_attr_id(self, name):
         """
-        Get the block attribute by name.
+        Get the block attribute's id by name.
 
         Args:
             name(str): the attribute name.
@@ -771,22 +771,74 @@ class Operator(object):
         Returns:
             int: the block index.
         """
-        return self.desc.block_attr(name)
+        return self.desc.block_attr_id(name)
+
+    def block_attr(self, name):
+        """
+        Get the block attribute  by name.
+
+        Args:
+            name(str): the attribute name.
+
+        Returns:
+            block: the block attribute.
+        """
+
+        id = self.block_attr_id(name)
+        assert (id >= 0 and id < len(self.block.program.blocks))
+        return self.block.program.blocks[id]
+
+    def blocks_attr(self, name):
+        """
+        Get the blocks attribute  by name.
+
+        Args:
+            name(str): the attribute name.
+
+        Returns:
+            list: list of the blocks attribute.
+        """
+        attrs = []
+        for i in self.blocks_attr_ids(name):
+            assert (i >= 0 and i < len(self.block.program.blocks))
+            attrs.append(self.block.program.blocks[i])
+
+        return attrs
+
+    def blocks_attr_ids(self, name):
+        """
+        Get the blocks attribute's ids by name.
+
+        Args:
+            name(str): the attribute name.
+
+        Returns:
+            list: list of the blocks ids.
+        """
+
+        return self.desc.blocks_attr_ids(name)
 
     def all_attrs(self):
         """
         Get the attribute dict.
 
         Returns:
-            dict: The Operator's attribute dict.
+            dict: The Operator's attribute dict, name->attr.
         """
         attr_names = self.attr_names
         attr_map = {}
         for n in attr_names:
-            if n == 'sub_block':
+            attr_type = self.desc.attr_type(n)
+            if attr_type == core.AttrType.BLOCK:
                 attr_map[n] = self.block_attr(n)
-            else:
-                attr_map[n] = self.attr(n)
+                continue
+
+            if attr_type == core.AttrType.BLOCKS:
+                attr_map[n] = self.blocks_attr(n)
+                continue
+
+            attr_map[n] = self.attr(n)
+
         return attr_map
 
 
@@ -1507,13 +1559,19 @@ class Program(object):
             The two code snippets above will generate same programs.
         """
         if for_test:
-            p = self.inference_optimize()
+            p = self.inference_optimize(export_for_deployment=False)
         else:
             p = Program()
+            p.current_block_idx = self.current_block_idx
+            p._seed = self._seed
             p.desc = core.ProgramDesc(self.desc)
             p.blocks = [
                 Block(p, i) for i in six.moves.range(self.desc.num_blocks())
             ]
+
+            p._current_role = self._current_role
+            p._op_role_var = self._op_role_var
+
             p._sync_with_cpp()
 
         p._copy_param_info_from(self)
@@ -1571,7 +1629,7 @@ class Program(object):
         res._sync_with_cpp()
         return res
 
-    def inference_optimize(self):
+    def inference_optimize(self, export_for_deployment=True):
         """
         This method will create a new program and do following adjustments on it:
         1. Remove all reader variables and their creator ops if exist.
@@ -1582,6 +1640,10 @@ class Program(object):
         attribute of operators to :code:`True`. All the :code:`Parameter`
         information will be lost.
 
+        Args:
+            export_for_deployment(bool): remove the read ops that are added by py_reader
+                                        for cpp inference library
+
         Notes: This API is a very low level API. Use
         :code:`Program.clone(for_test=True)` instead.
 
@@ -1596,16 +1658,17 @@ class Program(object):
         # remove all readers and the read_op if exist
         read_op_idx = 0
         root_block = res.desc.block(0)
-        while True:
-            if read_op_idx >= root_block.op_size() or root_block.op(
-                    read_op_idx).type() == 'read':
-                break
-            read_op_idx += 1
-        if read_op_idx < root_block.op_size():
-            root_block._remove_op(0, read_op_idx + 1)
-        for var in root_block.all_vars():
-            if var.type() == core.VarDesc.VarType.READER:
-                root_block._remove_var(var.name())
+        if export_for_deployment:
+            while True:
+                if read_op_idx >= root_block.op_size() or root_block.op(
+                        read_op_idx).type() == 'read':
+                    break
+                read_op_idx += 1
+            if read_op_idx < root_block.op_size():
+                root_block._remove_op(0, read_op_idx + 1)
+            for var in root_block.all_vars():
+                if var.type() == core.VarDesc.VarType.READER:
+                    root_block._remove_var(var.name())
 
         # change all `is_test` attributes to True
         for i in six.moves.range(res.desc.num_blocks()):
diff --git a/python/paddle/fluid/initializer.py b/python/paddle/fluid/initializer.py
index 83290ac60839b855a0348696ad6898af7335e2fc..3f740dd7c539973c5bb2b260120729e9563eedd2 100644
--- a/python/paddle/fluid/initializer.py
+++ b/python/paddle/fluid/initializer.py
@@ -264,7 +264,8 @@ class NormalInitializer(Initializer):
                 "dtype": int(var.dtype),
                 "mean": self._mean,
                 "std": self._std_dev,
-                "seed": self._seed
+                "seed": self._seed,
+                "use_mkldnn": False
             })
         var.op = op
         return op
diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py
index 78e5ef30ccc891039e78e029fef9c792ad96661f..e277c85021c971ad7b3deda6990391c600a6ebaa 100644
--- a/python/paddle/fluid/io.py
+++ b/python/paddle/fluid/io.py
@@ -555,7 +555,8 @@ def save_inference_model(dirname,
                          executor,
                          main_program=None,
                          model_filename=None,
-                         params_filename=None):
+                         params_filename=None,
+                         export_for_deployment=True):
     """
     Prune the given `main_program` to build a new program especially for inference,
     and then save it and all related parameters to given `dirname` by the `executor`.
@@ -577,6 +578,8 @@ def save_inference_model(dirname,
         params_filename(str|None): The name of file to save all related parameters.
                                    If it is setted None, parameters will be saved
                                    in separate files .
+        export_for_deployment(bool): remove the read ops that are added by py_reader
+                                    for cpp inference lib. Default True
 
     Returns:
         None
@@ -633,7 +636,8 @@ def save_inference_model(dirname,
     copy_program.desc.flush()
 
     pruned_program = copy_program.prune(targets=target_vars)
-    inference_program = pruned_program.inference_optimize()
+    inference_program = pruned_program.inference_optimize(
+        export_for_deployment=export_for_deployment)
     fetch_var_names = [v.name for v in target_vars]
 
     prepend_feed_ops(inference_program, feeded_var_names)
diff --git a/python/paddle/fluid/tests/unittests/test_desc_clone.py b/python/paddle/fluid/tests/unittests/test_desc_clone.py
new file mode 100644
index 0000000000000000000000000000000000000000..8603d3a5b3b5d368fe87b8dcf9dc7363f95caf86
--- /dev/null
+++ b/python/paddle/fluid/tests/unittests/test_desc_clone.py
@@ -0,0 +1,196 @@
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import argparse
+import time
+import math
+
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.profiler as profiler
+from paddle.fluid import core
+import unittest
+from multiprocessing import Process
+import os
+import signal
+import collections
+
+SEED = 1
+DTYPE = "float32"
+paddle.dataset.mnist.fetch()
+
+
+# random seed must set before configuring the network.
+# fluid.default_startup_program().random_seed = SEED
+def cnn_model(data):
+    conv_pool_1 = fluid.nets.simple_img_conv_pool(
+        input=data,
+        filter_size=5,
+        num_filters=20,
+        pool_size=2,
+        pool_stride=2,
+        act="relu")
+    conv_pool_2 = fluid.nets.simple_img_conv_pool(
+        input=conv_pool_1,
+        filter_size=5,
+        num_filters=50,
+        pool_size=2,
+        pool_stride=2,
+        act="relu")
+
+    # TODO(dzhwinter) : refine the initializer and random seed settting
+    SIZE = 10
+    input_shape = conv_pool_2.shape
+    param_shape = [reduce(lambda a, b: a * b, input_shape[1:], 1)] + [SIZE]
+    scale = (2.0 / (param_shape[0]**2 * SIZE))**0.5
+
+    predict = fluid.layers.fc(
+        input=conv_pool_2,
+        size=SIZE,
+        act="softmax",
+        param_attr=fluid.param_attr.ParamAttr(
+            initializer=fluid.initializer.NormalInitializer(
+                loc=0.0, scale=scale)))
+    return predict
+
+
+def get_model(batch_size):
+    # Input data
+    images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE)
+    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+
+    # Train program
+    predict = cnn_model(images)
+    cost = fluid.layers.cross_entropy(input=predict, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+
+    # Evaluator
+    batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
+    batch_acc = fluid.layers.accuracy(
+        input=predict, label=label, total=batch_size_tensor)
+
+    inference_program = fluid.default_main_program().clone()
+    # Optimization
+    opt = fluid.optimizer.AdamOptimizer(
+        learning_rate=0.001, beta1=0.9, beta2=0.999)
+
+    # Reader
+    train_reader = paddle.batch(
+        paddle.dataset.mnist.train(), batch_size=batch_size)
+    test_reader = paddle.batch(
+        paddle.dataset.mnist.test(), batch_size=batch_size)
+    opt.minimize(avg_cost)
+    return inference_program, avg_cost, train_reader, test_reader, batch_acc, predict
+
+
+def get_transpiler(trainer_id, main_program, pserver_endpoints, trainers):
+    t = fluid.DistributeTranspiler()
+    t.transpile(
+        trainer_id=trainer_id,
+        program=main_program,
+        pservers=pserver_endpoints,
+        trainers=trainers)
+    return t
+
+
+def operator_equal(a, b):
+    for k, v in a.__dict__.iteritems():
+        if isinstance(v, fluid.framework.Program) or \
+                isinstance(v, fluid.framework.Block):
+            continue
+
+        elif isinstance(v, core.OpDesc):
+            if v.serialize_to_string() != b.__dict__[k].serialize_to_string():
+                raise ValueError("In operator_equal not equal:{0}\n".format(k))
+
+        elif isinstance(v, collections.OrderedDict):
+            v0 = sorted(v.iteritems(), key=lambda x: x[0])
+            v1 = sorted(b.__dict__[k].iteritems(), key=lambda x: x[0])
+
+            if v0 != v1:
+                raise ValueError("In operator_equal not equal:{0}\n".format(k))
+
+        elif (v != b.__dict__[k]):
+            raise ValueError("In operator_equal not equal:{0}\n".format(k))
+
+    return True
+
+
+def block_equal(a, b):
+    for k, v in a.__dict__.iteritems():
+        if isinstance(v, core.ProgramDesc) or isinstance(
+                v, fluid.framework.Program) or isinstance(v, core.BlockDesc):
+            continue
+
+        elif k == "ops":
+            for i in range(0, len(a.ops)):
+                if not operator_equal(a.ops[i], b.ops[i]):
+                    raise ValueError("In block_equal not equal:{0}\n".format(k))
+            assert (len(a.ops) == len(b.ops))
+
+        elif isinstance(v, collections.OrderedDict):
+            v0 = sorted(v.iteritems(), key=lambda x: x[0])
+            v1 = sorted(b.__dict__[k].iteritems(), key=lambda x: x[0])
+
+            if v0 != v1:
+                raise ValueError("In block_equal not equal:{0}\n".format(k))
+
+        elif (v != b.__dict__[k]):
+            raise ValueError("In block_equal not equal:{0}\n".format(k))
+
+    return True
+
+
+def program_equal(a, b):
+    for k, v in a.__dict__.iteritems():
+        if isinstance(v, core.ProgramDesc):
+            continue
+
+        elif k == 'blocks':
+            for i in range(0, len(a.blocks)):
+                if not block_equal(a.blocks[i], b.blocks[i]):
+                    raise ValueError("In operator_equal not equal:{0}\n".format(
+                        k))
+                    return False
+            assert (len(a.blocks) == len(b.blocks))
+
+        elif (v != b.__dict__[k]):
+            raise ValueError("In program_equal not equal:{0}\n".format(k))
+
+    return True
+
+
+class TestDistMnist(unittest.TestCase):
+    def test_desc_clone(self):
+        get_model(batch_size=20)
+
+        pserver_endpoints = "127.0.0.1:9123"
+        trainers = 1
+        current_endpoint = "127.0.0.1:9123"
+        t = get_transpiler(0,
+                           fluid.default_main_program(), pserver_endpoints,
+                           trainers)
+
+        pserver_prog = t.get_pserver_program(current_endpoint)
+        startup_prog = t.get_startup_program(current_endpoint, pserver_prog)
+        main = pserver_prog.clone()
+        startup = startup_prog.clone()
+
+        self.assertTrue(program_equal(main, pserver_prog))
+        self.assertTrue(program_equal(startup, startup_prog))
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/test_dist_base.py b/python/paddle/fluid/tests/unittests/test_dist_base.py
index e059f2cd2ab994d580b3a494f6f5556938be1fa1..ab028dd36faf1fe0fc6fdaf3160bfd5b44d6ccdd 100644
--- a/python/paddle/fluid/tests/unittests/test_dist_base.py
+++ b/python/paddle/fluid/tests/unittests/test_dist_base.py
@@ -134,7 +134,7 @@ class TestDistBase(unittest.TestCase):
         self._ps_endpoints = "127.0.0.1:9123,127.0.0.1:9124"
         self._python_interp = "python"
 
-    def start_pserver(self, model_file):
+    def start_pserver(self, model_file, check_error_log):
         ps0_ep, ps1_ep = self._ps_endpoints.split(",")
         ps0_cmd = "%s %s pserver %s 0 %s %d TRUE" % \
             (self._python_interp, model_file, self._ps_endpoints, ps0_ep,
@@ -143,11 +143,23 @@ class TestDistBase(unittest.TestCase):
             (self._python_interp, model_file, self._ps_endpoints, ps1_ep,
              self._trainers)
 
+        ps0_pipe = subprocess.PIPE
+        ps1_pipe = subprocess.PIPE
+        if check_error_log:
+            print("ps0_cmd:", ps0_cmd)
+            print("ps1_cmd:", ps1_cmd)
+            ps0_pipe = open("/tmp/ps0_err.log", "wb")
+            ps1_pipe = open("/tmp/ps1_err.log", "wb")
+
         ps0_proc = subprocess.Popen(
-            ps0_cmd.split(" "), stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+            ps0_cmd.split(" "), stdout=subprocess.PIPE, stderr=ps0_pipe)
         ps1_proc = subprocess.Popen(
-            ps1_cmd.split(" "), stdout=subprocess.PIPE, stderr=subprocess.PIPE)
-        return ps0_proc, ps1_proc
+            ps1_cmd.split(" "), stdout=subprocess.PIPE, stderr=ps1_pipe)
+
+        if not check_error_log:
+            return ps0_proc, ps1_proc, None, None
+        else:
+            return ps0_proc, ps1_proc, ps0_pipe, ps1_pipe
 
     def _wait_ps_ready(self, pid):
         retry_times = 50
@@ -164,7 +176,7 @@ class TestDistBase(unittest.TestCase):
                                  (e, retry_times))
                 retry_times -= 1
 
-    def check_with_place(self, model_file, delta=1e-3):
+    def check_with_place(self, model_file, delta=1e-3, check_error_log=False):
         # *ATTENTION* THIS TEST NEEDS AT LEAST 2GPUS TO RUN
         required_envs = {
             "PATH": os.getenv("PATH"),
@@ -173,17 +185,32 @@ class TestDistBase(unittest.TestCase):
             "FLAGS_fraction_of_gpu_memory_to_use": "0.15",
             "FLAGS_cudnn_deterministic": "1"
         }
+
+        if check_error_log:
+            required_envs["GLOG_v"] = "7"
+            required_envs["GLOG_logtostderr"] = "1"
+
         # Run local to get a base line
         env_local = {"CUDA_VISIBLE_DEVICES": "0"}
         env_local.update(required_envs)
         local_cmd = "%s %s trainer %s 0 %s %d FLASE" % \
             (self._python_interp, model_file,
              "127.0.0.1:1234", "127.0.0.1:1234", 1)
-        local_proc = subprocess.Popen(
-            local_cmd.split(" "),
-            stdout=subprocess.PIPE,
-            stderr=subprocess.PIPE,
-            env=env_local)
+        if not check_error_log:
+            local_proc = subprocess.Popen(
+                local_cmd.split(" "),
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                env=env_local)
+        else:
+            print("trainer cmd:", local_cmd)
+            err_log = open("/tmp/trainer.err.log", "wb")
+            local_proc = subprocess.Popen(
+                local_cmd.split(" "),
+                stdout=subprocess.PIPE,
+                stderr=err_log,
+                env=env_local)
+
         local_proc.wait()
         out, err = local_proc.communicate()
         local_ret = cpt.to_text(out)
@@ -191,7 +218,8 @@ class TestDistBase(unittest.TestCase):
         sys.stderr.write('local_stderr: %s\n' % err)
 
         # Run dist train to compare with local results
-        ps0, ps1 = self.start_pserver(model_file)
+        ps0, ps1, ps0_pipe, ps1_pipe = self.start_pserver(model_file,
+                                                          check_error_log)
         self._wait_ps_ready(ps0.pid)
         self._wait_ps_ready(ps1.pid)
 
@@ -209,15 +237,23 @@ class TestDistBase(unittest.TestCase):
         env1.update(required_envs)
         FNULL = open(os.devnull, 'w')
 
+        tr0_pipe = subprocess.PIPE
+        tr1_pipe = subprocess.PIPE
+        if check_error_log:
+            print("tr0_cmd:", tr0_cmd)
+            print("tr1_cmd:", tr1_cmd)
+            tr0_pipe = open("/tmp/tr0_err.log", "wb")
+            tr1_pipe = open("/tmp/tr1_err.log", "wb")
+
         tr0_proc = subprocess.Popen(
             tr0_cmd.split(" "),
             stdout=subprocess.PIPE,
-            stderr=subprocess.PIPE,
+            stderr=tr0_pipe,
             env=env0)
         tr1_proc = subprocess.Popen(
             tr1_cmd.split(" "),
             stdout=subprocess.PIPE,
-            stderr=subprocess.PIPE,
+            stderr=tr1_pipe,
             env=env1)
 
         tr0_proc.wait()
@@ -234,6 +270,13 @@ class TestDistBase(unittest.TestCase):
         local_first_loss = eval(local_lines[0])[0]
         local_last_loss = eval(local_lines[1])[0]
 
+        # close trainer file
+        if check_error_log:
+            tr0_pipe.close()
+            tr1_pipe.close()
+
+            ps0_pipe.close()
+            ps1_pipe.close()
         # FIXME: use terminate() instead of sigkill.
         os.kill(ps0.pid, signal.SIGKILL)
         os.kill(ps1.pid, signal.SIGKILL)
diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py
index ac1719020bf6b6883619d613c8f34ae43c4cb224..3a91b34e6946118b963595a6222d23656e7af782 100644
--- a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py
+++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py
@@ -260,7 +260,7 @@ class TestLRDecayConditional(TranspilerTest):
         serv_op = pserver.blocks[0].ops[0]
         sub_blocks = []
         optimize_blocks = []
-        for b in serv_op.attrs["optimize_blocks"]:
+        for b in serv_op.all_attrs()["optimize_blocks"]:
             optimize_blocks.append(b.idx)
         for b in pserver.blocks:
             if b.idx not in optimize_blocks:
diff --git a/python/paddle/fluid/tests/unittests/test_program.py b/python/paddle/fluid/tests/unittests/test_program.py
index c51a48239330621d8e008415f81361616467cabf..0997afc97a97333c914a3027103ec48733b410dc 100644
--- a/python/paddle/fluid/tests/unittests/test_program.py
+++ b/python/paddle/fluid/tests/unittests/test_program.py
@@ -17,6 +17,7 @@ import unittest
 
 from paddle.fluid.framework import Program, default_main_program, program_guard, grad_var_name
 import paddle.fluid.layers as layers
+import paddle.fluid as fluid
 
 main_program = default_main_program()
 
@@ -98,6 +99,39 @@ class TestProgram(unittest.TestCase):
         new_program = main_program.clone()
         self.assertNotEqual(0, len(new_program.blocks[0].all_parameters()))
 
+    def test_program_inference_optimize(self):
+        def net():
+            reader = fluid.layers.py_reader(
+                capacity=10,
+                shapes=[[-1, 10], [-1, 1]],
+                lod_levels=[0, 0],
+                dtypes=['float32', 'int64'],
+                use_double_buffer=True)
+            in_data, label = fluid.layers.read_file(reader)
+            predict_label = fluid.layers.fc(in_data, size=2, act='softmax')
+            loss = fluid.layers.mean(
+                fluid.layers.cross_entropy(
+                    input=predict_label, label=label))
+
+            optimizer = fluid.optimizer.Adam()
+            optimizer.minimize(loss)
+
+        startup_program = fluid.Program()
+        main_program = fluid.Program()
+        with fluid.program_guard(main_program, startup_program):
+            net()
+        no_read_program = main_program.inference_optimize()
+        keep_read_program = main_program.inference_optimize(
+            export_for_deployment=False)
+        no_read_ops = no_read_program.global_block().ops
+        keep_read_ops = keep_read_program.global_block().ops
+        self.assertEqual(len(keep_read_ops) - len(no_read_ops), 2)
+        self.assertEqual(keep_read_ops[0].type, 'create_double_buffer_reader')
+        self.assertEqual(keep_read_ops[1].type, 'read')
+
+        for i in range(len(no_read_ops)):
+            self.assertEqual(no_read_ops[i].type, keep_read_ops[i + 2].type)
+
 
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py
index 37de79211413b757dccdd3ed68dfc3c38a02becb..f7087299cfb94a3d9413fa10ae8b192e2bf172d3 100644
--- a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py
+++ b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py
@@ -69,7 +69,7 @@ class TestOpDesc(unittest.TestCase):
         self.assertEqual(8, len(op.attr_names()))
 
         op.set_block_attr("block_attr", program_desc.block(0))
-        self.assertEqual(0, op.block_attr("block_attr"))
+        self.assertEqual(0, op.block_attr_id("block_attr"))
 
         mul_op = block.append_op()
         mul_op.set_type("mul")
diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py
index 466e603bdcc0fb710f91e8de1f7325ab3bec9c61..6d94d5745a44fe320cddaab30bc6ed8073c4126e 100644
--- a/python/paddle/fluid/transpiler/distribute_transpiler.py
+++ b/python/paddle/fluid/transpiler/distribute_transpiler.py
@@ -586,12 +586,12 @@ class DistributeTranspiler(object):
                 if op.type in [
                         "gaussian_random", "fill_constant", "uniform_random"
                 ]:
-                    op.attrs["shape"] = new_outputs["Out"].shape
+                    op.set_attr("shape", list(new_outputs["Out"].shape))
                 s_prog.global_block().append_op(
                     type=op.type,
                     inputs=new_inputs,
                     outputs=new_outputs,
-                    attrs=op.attrs)
+                    attrs=op.all_attrs())
         return s_prog
 
     # ====================== private transpiler functions =====================
@@ -605,7 +605,7 @@ class DistributeTranspiler(object):
         self.table_name = None
         for op in self.origin_program.global_block().ops:
             if op.type == LOOKUP_TABLE_TYPE:
-                if op.attrs['is_distributed'] is True:
+                if op.attr('is_distributed') is True:
                     if self.table_name is None:
                         self.table_name = op.input("W")[0]
                     if self.table_name != op.input("W")[0]:
@@ -1265,7 +1265,7 @@ class DistributeTranspiler(object):
             type=opt_op.type,
             inputs=new_inputs,
             outputs=outputs,
-            attrs=opt_op.attrs)
+            attrs=opt_op.all_attrs())
 
     def _is_splited_grad_var(self, var, var_dict):
         grad_block = None
@@ -1296,7 +1296,7 @@ class DistributeTranspiler(object):
                     block._clone_variable(var)
 
         return block.append_op(
-            type=op.type, inputs=inputs, outputs=outputs, attrs=op.attrs)
+            type=op.type, inputs=inputs, outputs=outputs, attrs=op.all_attrs())
 
     def _append_pserver_non_opt_ops(self, optimize_block, opt_op):
         program = optimize_block.program
@@ -1337,7 +1337,7 @@ class DistributeTranspiler(object):
             type=opt_op.type,
             inputs=inputs,
             outputs=outputs,
-            attrs=opt_op.attrs)
+            attrs=opt_op.all_attrs())
 
     def _is_op_connected(self, op1, op2):
         # If one op's input is another op's output or
@@ -1442,8 +1442,8 @@ class DistributeTranspiler(object):
         # optimize
         op_maker = core.op_proto_and_checker_maker
         optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize
-        if op_maker.kOpRoleAttrName() in op.attrs and \
-            int(op.attrs[op_maker.kOpRoleAttrName()]) == int(optimize_role):
+        if op_maker.kOpRoleAttrName() in op.attr_names and \
+                int(op.all_attrs()[op_maker.kOpRoleAttrName()]) == int(optimize_role):
             return True
         return False
 
@@ -1466,8 +1466,8 @@ class DistributeTranspiler(object):
                 # and op_role_var to get the pair.
                 for input_name in op.input_arg_names:
                     if input_name.find("@GRAD") != -1 and \
-                        op.attrs[RPC_OP_ROLE_ATTR_NAME]:
-                        param_name = op.attrs[OP_ROLE_VAR_ATTR_NAME][0]
+                        op.attr(RPC_OP_ROLE_ATTR_NAME):
+                        param_name = op.attr(OP_ROLE_VAR_ATTR_NAME)[0]
                         params_grads.append([
                             origin_var_dict[param_name],
                             origin_var_dict[input_name]