diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt
index 42af482f852a1f32fa0fb5595b7c533d612f7d02..43e1bc6b2efecd24d2d5bade5b6a7727bfb0a607 100644
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@@ -31,9 +31,7 @@ function(windows_symbolic TARGET)
 endfunction()
 
 add_subdirectory(ir)
-if (NOT WIN32)
 add_subdirectory(details)
-endif (NOT WIN32)
 # ddim lib
 proto_library(framework_proto SRCS framework.proto)
 
@@ -118,13 +116,8 @@ cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 
-if (NOT WIN32)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
     shape_inference data_transform lod_tensor profiler)
-else()
-cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
-    shape_inference data_transform lod_tensor)
-endif(NOT WIN32)
 
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
 
@@ -179,12 +172,10 @@ else()
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
 
-if (NOT WIN32)
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
         threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
         graph build_strategy
         fast_threaded_ssa_graph_executor)
-endif() # NOT WIN32
 
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)
diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h
index 949616f02d5168e6abab932d608e4b20ee64304a..c3a8b85423403992e3a12ceb0a1acbae82d25dfa 100644
--- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h
+++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h
@@ -13,9 +13,9 @@
 // limitations under the License.
 
 #pragma once
+#include <ThreadPool.h>
 #include <string>
 #include <vector>
-#include "ThreadPool.h"
 #include "paddle/fluid/framework/blocking_queue.h"
 #include "paddle/fluid/framework/details/exception_holder.h"
 #include "paddle/fluid/framework/details/execution_strategy.h"
diff --git a/paddle/fluid/memory/allocation/cpu_allocator.h b/paddle/fluid/memory/allocation/cpu_allocator.h
index 165f11cd3b0015db92951bdc5de34f53dfdc36f6..26d3643f4edff1f2d71b1c761e915a6dacb485ad 100644
--- a/paddle/fluid/memory/allocation/cpu_allocator.h
+++ b/paddle/fluid/memory/allocation/cpu_allocator.h
@@ -17,7 +17,8 @@
 
 #ifdef _WIN32
 #define posix_memalign_free _aligned_free
-#define posix_memalign(p, a, s) (((*(p)) = _aligned_malloc((s), (a))), *(p) ? 0 : errno)
+#define posix_memalign(p, a, s) \
+  (((*(p)) = _aligned_malloc((s), (a))), *(p) ? 0 : errno)
 #endif
 
 namespace paddle {
diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt
index 0d0613e1a4364e300640b62687c8a045e40b9ca9..93cb5eb2dc0b3480ebd05dcc6b36d8915d057bab 100644
--- a/paddle/fluid/platform/CMakeLists.txt
+++ b/paddle/fluid/platform/CMakeLists.txt
@@ -1,4 +1,3 @@
-if (NOT WIN32)
 proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto)
 py_proto_compile(profiler_py_proto SRCS profiler.proto)
 
@@ -6,11 +5,19 @@ add_custom_target(profiler_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch _
 
 add_dependencies(profiler_py_proto profiler_py_proto_init)
 
+if (NOT WIN32)
 add_custom_command(TARGET profiler_py_proto POST_BUILD
         COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler
         COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler
         COMMENT "Copy generated python proto into directory paddle/fluid/proto/profiler."
         WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+else(NOT WIN32)
+string(REPLACE "/" "\\" proto_dstpath "${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler/")
+add_custom_command(TARGET profiler_py_proto POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler
+        COMMAND copy /Y *.py ${proto_dstpath}
+        COMMENT "Copy generated python proto into directory paddle/fluid/proto/profiler."
+        WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
 endif(NOT WIN32)
 
 if(WITH_GPU)
@@ -60,12 +67,9 @@ cc_test(init_test SRCS init_test.cc DEPS device_context)
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
 nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context)
 
-
-if (NOT WIN32)
 cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto ${GPU_CTX_DEPS})
 cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
-endif(NOT WIN32)
 
 nv_test(float16_gpu_test SRCS float16_test.cu DEPS lod_tensor)
 cc_test(float16_test SRCS float16_test.cc DEPS lod_tensor)
diff --git a/paddle/fluid/platform/device_tracer.h b/paddle/fluid/platform/device_tracer.h
index f59fc40b71699a790978e22fd7e26da8d4d94c5f..eaf047d4744762f69d50bff8d467da8e3b8317cc 100644
--- a/paddle/fluid/platform/device_tracer.h
+++ b/paddle/fluid/platform/device_tracer.h
@@ -13,17 +13,11 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #pragma once
 
-#if !defined(_WIN32)
-#include <sys/time.h>
-#else
-#include <windows.h>
-#endif  // !_WIN32
-
-#include <time.h>
 #include <chrono>  // NOLINT
 #include <string>
 
 #include "paddle/fluid/platform/dynload/cupti.h"
+#include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/platform/profiler.pb.h"
 
 namespace paddle {
@@ -32,15 +26,11 @@ namespace platform {
 ///////////////////////
 // WARN: Under Development. Don't depend on it yet.
 //////////////////////
-#if !defined(_WIN32)
 inline uint64_t PosixInNsec() {
   struct timeval tv;
   gettimeofday(&tv, nullptr);
   return 1000 * (static_cast<uint64_t>(tv.tv_sec) * 1000000 + tv.tv_usec);
 }
-#else
-inline uint64_t PosixInNsec() { return static_cast<uint64_t>(0); }
-#endif  // !_WIN32
 
 // DeviceTracer performs the following tasks:
 // 1. Register cuda callbacks for various events: kernel, memcpy, etc.
diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h
index 3643d2ad15b9e99e579f062c78cef81edb2a818f..31309738a52f07096eca891e745ae31ea6d56924 100644
--- a/paddle/fluid/platform/enforce.h
+++ b/paddle/fluid/platform/enforce.h
@@ -134,7 +134,7 @@ struct EOFException : public std::exception {
 #define LIKELY(condition) __builtin_expect(static_cast<bool>(condition), 1)
 #else
 // there is no equivalent intrinsics in msvc.
-#define LIKELY(condition) !(condition)
+#define LIKELY(condition) (condition)
 #endif
 
 template <typename... Args>
diff --git a/paddle/fluid/platform/port.h b/paddle/fluid/platform/port.h
index a07b993c8a84e26fb079804b6ec3ff9df3efd09e..8be77fe46457511220a26b9733df2c4b8d013823 100644
--- a/paddle/fluid/platform/port.h
+++ b/paddle/fluid/platform/port.h
@@ -17,6 +17,7 @@
 #include <cstdio>
 #include <stdexcept>
 
+#include <time.h>
 #include <memory>
 #include <string>
 
@@ -27,6 +28,7 @@
 #include <dlfcn.h>     //  dladdr
 #include <execinfo.h>  // backtrace
 #include <sys/stat.h>
+#include <sys/time.h>
 #include <algorithm>  // std::accumulate
 #else
 #include <io.h>  // _popen, _pclose
@@ -57,6 +59,25 @@ static void *dlopen(const char *filename, int flag) {
   return reinterpret_cast<void *>(hModule);
 }
 
+static int gettimeofday(struct timeval *tp, void *tzp) {
+  time_t clock;
+  struct tm tm;
+  SYSTEMTIME wtm;
+
+  GetLocalTime(&wtm);
+  tm.tm_year = wtm.wYear - 1900;
+  tm.tm_mon = wtm.wMonth - 1;
+  tm.tm_mday = wtm.wDay;
+  tm.tm_hour = wtm.wHour;
+  tm.tm_min = wtm.wMinute;
+  tm.tm_sec = wtm.wSecond;
+  tm.tm_isdst = -1;
+  clock = mktime(&tm);
+  tp->tv_sec = clock;
+  tp->tv_usec = wtm.wMilliseconds * 1000;
+
+  return (0);
+}
 #endif  // !_WIN32
 
 static void ExecShellCommand(const std::string &cmd, std::string *message) {
diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc
index 56bf9e31a35fdec5b7f04849068ff96ac9776c0e..03c102e24a189ea9e1614ffa6773e0c6e742c107 100644
--- a/paddle/fluid/platform/profiler.cc
+++ b/paddle/fluid/platform/profiler.cc
@@ -13,8 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/platform/profiler.h"
+#include "paddle/fluid/platform/port.h"
 
-#include <sys/time.h>
 #include <algorithm>
 #include <iomanip>
 #include <limits>
@@ -438,10 +438,10 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
             event_items[index].total_time += event_time;
             // min time
             event_items[index].min_time =
-                std::min(event_time, event_items[index].min_time);
+                (std::min)(event_time, event_items[index].min_time);
             // max time
             event_items[index].max_time =
-                std::max(event_time, event_items[index].max_time);
+                (std::max)(event_time, event_items[index].max_time);
           }
 
           // remove the push marker from the list
diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h
index e8eae874afa3d17f0d3374eef457cdbacb3f8424..f5d3490634f3199a23986ec3ae13d9fe3577ac35 100644
--- a/paddle/fluid/platform/profiler.h
+++ b/paddle/fluid/platform/profiler.h
@@ -69,7 +69,6 @@ void PushEvent(const std::string& name, const DeviceContext* dev_ctx);
 
 void PopEvent(const std::string& name, const DeviceContext* dev_ctx);
 
-#if !defined(_WIN32)
 struct RecordEvent {
   // dev_ctx can be set to nullptr if device is cpu.
   RecordEvent(const std::string& name, const DeviceContext* dev_ctx);
@@ -106,15 +105,6 @@ struct RecordBlock {
   std::string name_;
   uint64_t start_ns_;
 };
-#else
-// windows do not support profiler temporarily.
-struct RecordEvent {
-  RecordEvent(const std::string& name, const DeviceContext* dev_ctx) {}
-};
-struct RecordBlock {
-  explicit RecordBlock(int block_id) {}
-};
-#endif
 
 // Return the event list of all threads. Assumed the returned value calls
 // event_lists, event_lists[i][j] represents the j-th Event of i-th thread.
diff --git a/paddle/fluid/platform/stream_callback_manager.h b/paddle/fluid/platform/stream_callback_manager.h
index 0e88a439cf6ca83e3d98725f58875adeeea86be0..11c68f3449ee26b64c121acd081479b37c94fac4 100644
--- a/paddle/fluid/platform/stream_callback_manager.h
+++ b/paddle/fluid/platform/stream_callback_manager.h
@@ -45,16 +45,15 @@ class StreamCallbackManager {
   inline void AddCallback(Callback &&callback) const {
     auto *stream_callback_context =
         new StreamCallbackContext(this, std::forward<Callback>(callback));
-    PADDLE_ENFORCE(
 #if CUDA_VERSION >= 10000
-        cudaLaunchHostFunc(stream_, StreamCallbackManager::StreamCallbackFunc,
-                           stream_callback_context)
+    PADDLE_ENFORCE(cudaLaunchHostFunc(stream_,
+                                      StreamCallbackManager::StreamCallbackFunc,
+                                      stream_callback_context));  // NOLINT
 #else
-        cudaStreamAddCallback(stream_,
-                              StreamCallbackManager::StreamCallbackFunc,
-                              stream_callback_context, 0)
+    PADDLE_ENFORCE(cudaStreamAddCallback(
+        stream_, StreamCallbackManager::StreamCallbackFunc,
+        stream_callback_context, 0));  // NOLINT
 #endif
-            );  // NOLINT
   }
 
   void Wait() const { thread_pool_.reset(new ThreadPool(1)); }
diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt
index 25e919105cba48ea28b6d429542c33ba48515ef5..fb6ee2f4a53925f64b61b9fe03f5a4bd7203ed53 100644
--- a/paddle/fluid/pybind/CMakeLists.txt
+++ b/paddle/fluid/pybind/CMakeLists.txt
@@ -1,9 +1,6 @@
 
-set(PYBIND_DEPS pybind python proto_desc memory executor prune feed_fetch_method pass_builder)
+set(PYBIND_DEPS pybind python proto_desc memory executor prune feed_fetch_method pass_builder parallel_executor profiler)
 set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc)
-if(NOT WIN32)
-  list(APPEND PYBIND_DEPS parallel_executor profiler)
-endif(NOT WIN32)
 if(WITH_PYTHON)
   if(WITH_AMD_GPU)
     hip_library(paddle_pybind SHARED
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 2f040e1c34c7956959dc9fe55b4aaf0b3b6f880c..102fa02adf3fb05e36183f87e95d8178be4e8742 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -36,9 +36,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"
-#ifndef _WIN32
 #include "paddle/fluid/framework/parallel_executor.h"
-#endif
 #include "paddle/fluid/framework/prune.h"
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/selected_rows.h"
@@ -637,7 +635,6 @@ All parameter, weight, gradient are variables in Paddle.
 #endif
 #endif
 
-#ifndef _WIN32
   py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
       .value("kDisabled", platform::ProfilerState::kDisabled)
       .value("kCPU", platform::ProfilerState::kCPU)
@@ -658,7 +655,6 @@ All parameter, weight, gradient are variables in Paddle.
   m.def("disable_profiler", platform::DisableProfiler);
   m.def("is_profiler_enabled", platform::IsProfileEnabled);
   m.def("reset_profiler", platform::ResetProfiler);
-#endif
 
   py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
   pass.def(py::init())
@@ -687,7 +683,6 @@ All parameter, weight, gradient are variables in Paddle.
       .def("remove_pass",
            [](ir::PassBuilder &self, size_t idx) { self.RemovePass(idx); });
 
-#ifndef _WIN32
   // -- python binds for parallel executor.
   py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
   py::class_<ExecutionStrategy> exec_strategy(pe, "ExecutionStrategy", R"DOC(
@@ -913,7 +908,6 @@ All parameter, weight, gradient are variables in Paddle.
         pybind11::gil_scoped_release release;
         self.Run(fetch_tensors, fetched_var_name);
       });
-#endif
 
   BindRecordIOWriter(&m);
   return m.ptr();
diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
index 6a4a5e098fc27e02104baa837d939a62cc2a22a8..543acf2d349c7d02480204699f497536c7a4ca60 100644
--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -47,8 +47,7 @@ from . import profiler
 from . import unique_name
 from . import recordio_writer
 from . import parallel_executor
-if os.name != 'nt':
-    from .parallel_executor import *
+from .parallel_executor import *
 from paddle.fluid.layers.math_op_patch import monkey_patch_variable
 
 Tensor = LoDTensor
diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py
index 0d53f53a9ef6d5f90d188e1ad10785b5558c3e80..3f4dd5eb712e738bbee8f93c062375033b8ab2f6 100644
--- a/python/paddle/fluid/parallel_executor.py
+++ b/python/paddle/fluid/parallel_executor.py
@@ -25,264 +25,263 @@ import os
 
 __all__ = ['ParallelExecutor', 'ExecutionStrategy', 'BuildStrategy']
 
-if os.name != 'nt':
-    ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy
-    BuildStrategy = core.ParallelExecutor.BuildStrategy
-
-    class ParallelExecutor(object):
+ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy
+BuildStrategy = core.ParallelExecutor.BuildStrategy
+
+
+class ParallelExecutor(object):
+    """
+    ParallelExecutor is designed for data parallelism, which focuses on distributing
+    the data across different nodes and every node operates on the data in parallel.
+    If you use ParallelExecutor to run the current program on GPU, the node means GPU
+    device, and ParallelExecutor will get the available GPU device automatically on
+    the current machine. If you use ParallelExecutor to run the current program on CPU,
+    the node means the CPU device, and you can specify the CPU device number by adding
+    'CPU_NUM' environment variable, for example 'CPU_NUM=4', if the environment variable
+    is not found, ParallelExecutor will call `multiprocessing.cpu_count` to get the number
+    of CPUs in the system.
+
+    Args:
+        use_cuda (bool): Whether to use CUDA or not.
+        loss_name (str): The loss name must set in training. Default None.
+        main_program (Program): The program that need to run, if not provided,
+            then default_main_program will be used. Default None.
+        share_vars_from(ParallelExecutor): If provide, it will share variables
+            from the specified ParallelExecutor. Default None.
+        exec_strategy(ExecutionStrategy): exec_strategy is used to control how to run
+            the program in ParallelExecutor, for example how many threads are used to
+            execute the program, how many iterations to clean up the temp variables
+            which is generated during execution. For more information, please refer
+            to fluid.ExecutionStrategy. Default None.
+        build_strategy(BuildStrategy): build_strategy is used to control how to
+            build the SSA Graph in ParallelExecutor by setting the property,
+            for example reduce_strategy, gradient_scale_strategy. For more information,
+            please refer to fluid.BuildStrategy. Default None.
+        num_trainers(int): If greater than 1, NCCL will be initialized with
+            multiple rank of nodes, each node should have same number of GPUs.
+            Distributed training will be enabled then. Default 1.
+        trainer_id(int): Must use together with num_trainers. trainer_id is the
+            "rank" of current node starts from 0. Default 0.
+        scope(Scope): scope to run with, default use fluid.global_scope().
+
+    Returns:
+        ParallelExecutor: The initialized ParallelExecutor object.
+
+    Raises:
+        TypeError: If share_vars_from is provided, but not ParallelExecutor object.
+
+    Examples:
+        .. code-block:: python
+
+          train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name)
+          test_exe = fluid.ParallelExecutor(use_cuda=True,
+                                            main_program=test_program,
+                                            share_vars_from=train_exe)
+
+          train_loss, = train_exe.run([loss.name], feed=feed_dict)
+          test_loss, = test_exe.run([loss.name], feed=feed_dict)
+    """
+
+    def __init__(self,
+                 use_cuda,
+                 loss_name=None,
+                 main_program=None,
+                 share_vars_from=None,
+                 exec_strategy=None,
+                 build_strategy=None,
+                 num_trainers=1,
+                 trainer_id=0,
+                 scope=None):
+        self._places = []
+        self._act_places = []
+        if use_cuda:
+            for i in six.moves.range(core.get_cuda_device_count()):
+                p = core.Place()
+                self._act_places.append(core.CUDAPlace(i))
+                p.set_place(self._act_places[-1])
+                self._places.append(p)
+        else:
+            cpu_num = int(
+                os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+            for i in six.moves.range(cpu_num):
+                p = core.Place()
+                self._act_places.append(core.CPUPlace())
+                p.set_place(self._act_places[-1])
+                self._places.append(p)
+        assert self._places, "no place for execution"
+
+        if exec_strategy is None:
+            exec_strategy = ExecutionStrategy()
+        exec_strategy.use_cuda = use_cuda
+
+        if exec_strategy.num_threads == 0:
+            if use_cuda:
+                # Experiments on se-resnext shows that too many threads hurt
+                # performance. Worth tunning for other models in the future.
+                exec_strategy.num_threads = len(self._places) * 4
+            else:
+                cpu_num = int(
+                    os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+                exec_strategy.num_threads = cpu_num * 2
+
+        # Set 1 thread num under nccl2 distribute 
+        #   env to make sure all gpus run ops in same order.
+        if num_trainers > 1:
+            assert (use_cuda)
+            # FIXME(gongwb): avoid this set.
+            exec_strategy.num_threads = 1
+
+        if build_strategy is None:
+            build_strategy = BuildStrategy()
+
+        main = main_program
+        main = main if main else framework.default_main_program()
+        if scope == None:
+            scope = executor.global_scope()
+
+        if share_vars_from and not isinstance(share_vars_from,
+                                              ParallelExecutor):
+            raise TypeError("share_vars_from must be ParallelExecutor.")
+
+        local_scopes = share_vars_from.executor.local_scopes(
+        ) if share_vars_from else []
+
+        self.persistable_vars = [
+            v.name for v in [
+                var for var in main.list_vars()
+                if var.persistable and var.type != core.VarDesc.VarType.RAW
+            ]
+        ]
+
+        self.executor = core.ParallelExecutor(
+            self._places,
+            set([
+                cpt.to_text(p.name)
+                for p in main.global_block().iter_parameters()
+                if not p.stop_gradient
+            ]),
+            set(cpt.to_text(var) for var in self.persistable_vars), main.desc,
+            cpt.to_text(loss_name)
+            if loss_name else six.u(''), scope, local_scopes, exec_strategy,
+            build_strategy, num_trainers, trainer_id)
+        self.scope = scope
+
+    def run(self, fetch_list, feed=None, feed_dict=None, return_numpy=True):
         """
-        ParallelExecutor is designed for data parallelism, which focuses on distributing
-        the data across different nodes and every node operates on the data in parallel.
-        If you use ParallelExecutor to run the current program on GPU, the node means GPU
-        device, and ParallelExecutor will get the available GPU device automatically on
-        the current machine. If you use ParallelExecutor to run the current program on CPU,
-        the node means the CPU device, and you can specify the CPU device number by adding
-        'CPU_NUM' environment variable, for example 'CPU_NUM=4', if the environment variable
-        is not found, ParallelExecutor will call `multiprocessing.cpu_count` to get the number
-        of CPUs in the system.
+        Run a parallel executor with fetch_list.
+
+        The feed parameter can be a dict or a list. If feed is a dict, the
+        feed data will be split into multiple devices. If feed is a list, we
+        assume the data has been splitted into multiple devices, the each
+        element in the list will be copied to each device directly.
+
+        For example, if the feed is a dict:
+
+        >>> exe = ParallelExecutor()
+        >>> # the image will be splitted into devices. If there is two devices
+        >>> # each device will process an image with shape (24, 1, 28, 28)
+        >>> exe.run(feed={'image': numpy.random.random(size=(48, 1, 28, 28))})
+
+        For example, if the feed is a list:
+
+        >>> exe = ParallelExecutor()
+        >>> # each device will process each element in the list.
+        >>> # the 1st device will process an image with shape (48, 1, 28, 28)
+        >>> # the 2nd device will process an image with shape (32, 1, 28, 28)
+        >>> #
+        >>> # you can use exe.device_count to get the device number.
+        >>> exe.run(feed=[{"image": numpy.random.random(size=(48, 1, 28, 28))},
+        >>>               {"image": numpy.random.random(size=(32, 1, 28, 28))},
+        >>>              ])
 
         Args:
-            use_cuda (bool): Whether to use CUDA or not.
-            loss_name (str): The loss name must set in training. Default None.
-            main_program (Program): The program that need to run, if not provided,
-                then default_main_program will be used. Default None.
-            share_vars_from(ParallelExecutor): If provide, it will share variables
-                from the specified ParallelExecutor. Default None.
-            exec_strategy(ExecutionStrategy): exec_strategy is used to control how to run
-                the program in ParallelExecutor, for example how many threads are used to
-                execute the program, how many iterations to clean up the temp variables
-                which is generated during execution. For more information, please refer
-                to fluid.ExecutionStrategy. Default None.
-            build_strategy(BuildStrategy): build_strategy is used to control how to
-                build the SSA Graph in ParallelExecutor by setting the property,
-                for example reduce_strategy, gradient_scale_strategy. For more information,
-                please refer to fluid.BuildStrategy. Default None.
-            num_trainers(int): If greater than 1, NCCL will be initialized with
-                multiple rank of nodes, each node should have same number of GPUs.
-                Distributed training will be enabled then. Default 1.
-            trainer_id(int): Must use together with num_trainers. trainer_id is the
-                "rank" of current node starts from 0. Default 0.
-            scope(Scope): scope to run with, default use fluid.global_scope().
+            fetch_list(list): The fetched variable names
+            feed(list|dict|None): The feed variables. If the feed is a dict,
+                tensors in that dict will be splitted into each devices. If
+                the feed is a list, each element of the list will be copied
+                to each device. Default None.
+            feed_dict: Alias for feed parameter, for backward compatibility.
+                This parameter has been deprecated. Default None.
+            return_numpy(bool): Whether converts the fetched tensor to numpy.
+                Default: True.
 
         Returns:
-            ParallelExecutor: The initialized ParallelExecutor object.
+            List: The fetched result list.
 
         Raises:
-            TypeError: If share_vars_from is provided, but not ParallelExecutor object.
+            ValueError: If the feed is a list, but its length is not equal the
+                length of active places, or its element's is not dict.
+
+        NOTES:
+            1. If the feed's type is dict, the number of data that feeds to
+               ParallelExecutor must be bigger than active places. Otherwise,
+               it will throw exception from C++ side. Special attention should be
+               paid to check whether the last batch of the dataset is bigger
+               than active places.
+            2. If active places are more than one, the fetch results for each
+               variable is a list, and each element of this list is the variable of
+               respective active place.
 
         Examples:
             .. code-block:: python
 
-              train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name)
-              test_exe = fluid.ParallelExecutor(use_cuda=True,
-                                                main_program=test_program,
-                                                share_vars_from=train_exe)
-
-              train_loss, = train_exe.run([loss.name], feed=feed_dict)
-              test_loss, = test_exe.run([loss.name], feed=feed_dict)
+                pe = fluid.ParallelExecutor(use_cuda=use_cuda,
+                                            loss_name=avg_cost.name,
+                                            main_program=fluid.default_main_program())
+                loss = pe.run(feed=feeder.feed(cur_batch),
+                              fetch_list=[avg_cost.name]))
         """
-
-        def __init__(self,
-                     use_cuda,
-                     loss_name=None,
-                     main_program=None,
-                     share_vars_from=None,
-                     exec_strategy=None,
-                     build_strategy=None,
-                     num_trainers=1,
-                     trainer_id=0,
-                     scope=None):
-            self._places = []
-            self._act_places = []
-            if use_cuda:
-                for i in six.moves.range(core.get_cuda_device_count()):
-                    p = core.Place()
-                    self._act_places.append(core.CUDAPlace(i))
-                    p.set_place(self._act_places[-1])
-                    self._places.append(p)
-            else:
-                cpu_num = int(
-                    os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
-                for i in six.moves.range(cpu_num):
-                    p = core.Place()
-                    self._act_places.append(core.CPUPlace())
-                    p.set_place(self._act_places[-1])
-                    self._places.append(p)
-            assert self._places, "no place for execution"
-
-            if exec_strategy is None:
-                exec_strategy = ExecutionStrategy()
-            exec_strategy.use_cuda = use_cuda
-
-            if exec_strategy.num_threads == 0:
-                if use_cuda:
-                    # Experiments on se-resnext shows that too many threads hurt
-                    # performance. Worth tunning for other models in the future.
-                    exec_strategy.num_threads = len(self._places) * 4
-                else:
-                    cpu_num = int(
-                        os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
-                    exec_strategy.num_threads = cpu_num * 2
-
-            # Set 1 thread num under nccl2 distribute
-            #   env to make sure all gpus run ops in same order.
-            if num_trainers > 1:
-                assert (use_cuda)
-                # FIXME(gongwb): avoid this set.
-                exec_strategy.num_threads = 1
-
-            if build_strategy is None:
-                build_strategy = BuildStrategy()
-
-            main = main_program
-            main = main if main else framework.default_main_program()
-            if scope == None:
-                scope = executor.global_scope()
-
-            if share_vars_from and not isinstance(share_vars_from,
-                                                  ParallelExecutor):
-                raise TypeError("share_vars_from must be ParallelExecutor.")
-
-            local_scopes = share_vars_from.executor.local_scopes(
-            ) if share_vars_from else []
-
-            self.persistable_vars = [
-                v.name for v in [
-                    var for var in main.list_vars()
-                    if var.persistable and var.type != core.VarDesc.VarType.RAW
-                ]
-            ]
-
-            self.executor = core.ParallelExecutor(
-                self._places,
-                set([
-                    cpt.to_text(p.name)
-                    for p in main.global_block().iter_parameters()
-                    if not p.stop_gradient
-                ]),
-                set(cpt.to_text(var)
-                    for var in self.persistable_vars), main.desc,
-                cpt.to_text(loss_name)
-                if loss_name else six.u(''), scope, local_scopes, exec_strategy,
-                build_strategy, num_trainers, trainer_id)
-            self.scope = scope
-
-        def run(self, fetch_list, feed=None, feed_dict=None, return_numpy=True):
-            """
-            Run a parallel executor with fetch_list.
-
-            The feed parameter can be a dict or a list. If feed is a dict, the
-            feed data will be split into multiple devices. If feed is a list, we
-            assume the data has been splitted into multiple devices, the each
-            element in the list will be copied to each device directly.
-
-            For example, if the feed is a dict:
-
-            >>> exe = ParallelExecutor()
-            >>> # the image will be splitted into devices. If there is two devices
-            >>> # each device will process an image with shape (24, 1, 28, 28)
-            >>> exe.run(feed={'image': numpy.random.random(size=(48, 1, 28, 28))})
-
-            For example, if the feed is a list:
-
-            >>> exe = ParallelExecutor()
-            >>> # each device will process each element in the list.
-            >>> # the 1st device will process an image with shape (48, 1, 28, 28)
-            >>> # the 2nd device will process an image with shape (32, 1, 28, 28)
-            >>> #
-            >>> # you can use exe.device_count to get the device number.
-            >>> exe.run(feed=[{"image": numpy.random.random(size=(48, 1, 28, 28))},
-            >>>               {"image": numpy.random.random(size=(32, 1, 28, 28))},
-            >>>              ])
-
-            Args:
-                fetch_list(list): The fetched variable names
-                feed(list|dict|None): The feed variables. If the feed is a dict,
-                    tensors in that dict will be splitted into each devices. If
-                    the feed is a list, each element of the list will be copied
-                    to each device. Default None.
-                feed_dict: Alias for feed parameter, for backward compatibility.
-                    This parameter has been deprecated. Default None.
-                return_numpy(bool): Whether converts the fetched tensor to numpy.
-                    Default: True.
-
-            Returns:
-                List: The fetched result list.
-
-            Raises:
-                ValueError: If the feed is a list, but its length is not equal the
-                    length of active places, or its element's is not dict.
-
-            NOTES:
-                1. If the feed's type is dict, the number of data that feeds to
-                   ParallelExecutor must be bigger than active places. Otherwise,
-                   it will throw exception from C++ side. Special attention should be
-                   paid to check whether the last batch of the dataset is bigger
-                   than active places.
-                2. If active places are more than one, the fetch results for each
-                   variable is a list, and each element of this list is the variable of
-                   respective active place.
-
-            Examples:
-                .. code-block:: python
-
-                    pe = fluid.ParallelExecutor(use_cuda=use_cuda,
-                                                loss_name=avg_cost.name,
-                                                main_program=fluid.default_main_program())
-                    loss = pe.run(feed=feeder.feed(cur_batch),
-                                  fetch_list=[avg_cost.name]))
-            """
-            if feed is None and feed_dict is not None:
-                feed = feed_dict
-                print(
-                    "`feed_dict` is deprecated. Please use `feed=`",
-                    file=sys.stderr)
-
-            if isinstance(feed, dict):
-                feed_tensor_dict = dict()
-                for feed_name in feed:
-                    feed_tensor = feed[feed_name]
-                    if not isinstance(feed_tensor, core.LoDTensor):
-                        feed_tensor = core.LoDTensor()
-                        # always set to CPU place, since the tensor need to be splitted
-                        # it is fast in CPU
-                        feed_tensor.set(feed[feed_name], core.CPUPlace())
-                    feed_tensor_dict[feed_name] = feed_tensor
-
-                self.executor.feed_and_split_tensor_into_local_scopes(
-                    feed_tensor_dict)
-            elif isinstance(feed, list) or isinstance(feed, tuple):
-                if len(feed) != len(self._act_places):
-                    raise ValueError(
-                        "Feed a list of tensor, the list should be the same size as places"
-                    )
-
-                res = list()
-
-                for i, each in enumerate(feed):
-                    if not isinstance(each, dict):
-                        raise TypeError(
-                            "Each element of feed list should be a dict")
-                    res_dict = dict()
-                    for feed_name in each:
-                        tensor = each[feed_name]
-                        if not isinstance(tensor, core.LoDTensor):
-                            tmp = core.LoDTensor()
-                            tmp.set(tensor, self._act_places[i])
-                            tensor = tmp
-                        res_dict[feed_name] = tensor
-                    res.append(res_dict)
-                self.executor.feed_tensors_into_local_scopes(res)
-
-            fetch_var_name = '@FETCHED_VAR_NAME@'
-            self.executor.run(fetch_list, fetch_var_name)
-            arr = self.scope.find_var(fetch_var_name).get_lod_tensor_array()
-
-            if return_numpy:
-                return executor.as_numpy(arr)
-
-            return [arr[i] for i in range(len(arr))]
-
-        @property
-        def device_count(self):
-            return len(self._act_places)
+        if feed is None and feed_dict is not None:
+            feed = feed_dict
+            print(
+                "`feed_dict` is deprecated. Please use `feed=`",
+                file=sys.stderr)
+
+        if isinstance(feed, dict):
+            feed_tensor_dict = dict()
+            for feed_name in feed:
+                feed_tensor = feed[feed_name]
+                if not isinstance(feed_tensor, core.LoDTensor):
+                    feed_tensor = core.LoDTensor()
+                    # always set to CPU place, since the tensor need to be splitted
+                    # it is fast in CPU
+                    feed_tensor.set(feed[feed_name], core.CPUPlace())
+                feed_tensor_dict[feed_name] = feed_tensor
+
+            self.executor.feed_and_split_tensor_into_local_scopes(
+                feed_tensor_dict)
+        elif isinstance(feed, list) or isinstance(feed, tuple):
+            if len(feed) != len(self._act_places):
+                raise ValueError(
+                    "Feed a list of tensor, the list should be the same size as places"
+                )
+
+            res = list()
+
+            for i, each in enumerate(feed):
+                if not isinstance(each, dict):
+                    raise TypeError(
+                        "Each element of feed list should be a dict")
+                res_dict = dict()
+                for feed_name in each:
+                    tensor = each[feed_name]
+                    if not isinstance(tensor, core.LoDTensor):
+                        tmp = core.LoDTensor()
+                        tmp.set(tensor, self._act_places[i])
+                        tensor = tmp
+                    res_dict[feed_name] = tensor
+                res.append(res_dict)
+            self.executor.feed_tensors_into_local_scopes(res)
+
+        fetch_var_name = '@FETCHED_VAR_NAME@'
+        self.executor.run(fetch_list, fetch_var_name)
+        arr = self.scope.find_var(fetch_var_name).get_lod_tensor_array()
+
+        if return_numpy:
+            return executor.as_numpy(arr)
+
+        return [arr[i] for i in range(len(arr))]
+
+    @property
+    def device_count(self):
+        return len(self._act_places)