Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into feature/support_op_role

doc/v2/build_and_install/build_from_source_cn.rst

@@ -19,8 +19,9 @@
 ----------------
 
 PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安装编译依赖的步骤，可选的不同编译环境Docker镜像
-可以在 `这里 <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ 找到。或者
-参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。
+可以在 `这里 <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ 找到，您也可以
+在 `这里 <https://github.com/PaddlePaddle/Paddle/tree/develop/tools/manylinux1/>`_ 找到 paddle_manylinux_devel
+镜像的编译以及使用方法。或者参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。
 
 如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。
 

doc/v2/build_and_install/build_from_source_en.rst

@@ -22,6 +22,8 @@ How To Build
 You need to use Docker to build PaddlePaddle
 to avoid installing dependencies by yourself. We have several pre-built
 Docker images `here <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ ,
+you can also find how to build and use paddle_manylinux_devel Docker image from
+`here <https://github.com/PaddlePaddle/Paddle/tree/develop/tools/manylinux1/>`_
 Or you can build your own image from source as the optional step below:
 
 .. code-block:: bash

paddle/fluid/inference/analysis/dot.h

@@ -21,6 +21,7 @@
 
 #include <glog/logging.h>
 #include <sstream>
+#include <string>
 #include <unordered_map>
 #include <vector>
 

paddle/fluid/inference/engine.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
 #include "paddle/fluid/framework/framework.pb.h"
 
 namespace paddle {
@@ -58,8 +59,8 @@ class EngineBase {
 
 struct Buffer {
   void* buffer{nullptr};               // buffer should be allocated only once.
-  int max_size;                        // buffer allocated space.
-  int size;                            // data size.
+  size_t max_size;                     // buffer allocated space.
+  size_t size;                         // data size.
   DeviceType device{DeviceType::UNK};  // tells which device this buffer is on.
 };
 

paddle/fluid/inference/tensorrt/CMakeLists.txt

 nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto)
 nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
 nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine)
-
 add_subdirectory(convert)

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

-nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc op_converter.h DEPS ${FLUID_CORE_MODULES})
-nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
+nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES})
+nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc
   DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine)
 nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)

paddle/fluid/inference/tensorrt/convert/io_converter.cc

@@ -23,26 +23,42 @@ namespace tensorrt {
 using platform::is_gpu_place;
 using platform::is_cpu_place;
 
-class DefaultInputConverter : public EngineInputConverter {
+class DefaultIOConverter : public EngineIOConverter {
  public:
-  DefaultInputConverter() {}
+  DefaultIOConverter() {}
   // NOTE out is GPU memory.
   virtual void operator()(const LoDTensor& in, void* out,
                           size_t max_size) override {
     PADDLE_ENFORCE(out != nullptr);
-    PADDLE_ENFORCE_LE(in.memory_size(), max_size);
+    PADDLE_ENFORCE(stream_ != nullptr);
     const auto& place = in.place();
+    size_t size = in.memory_size();
+    PADDLE_ENFORCE_LE(size, max_size);
     if (is_cpu_place(place)) {
-      PADDLE_ENFORCE(stream_ != nullptr);
-      PADDLE_ENFORCE_EQ(0,
-                        cudaMemcpyAsync(out, in.data<float>(), in.memory_size(),
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data<float>(), size,
                                            cudaMemcpyHostToDevice, *stream_));
-
     } else if (is_gpu_place(place)) {
-      PADDLE_ENFORCE_EQ(0,
-                        cudaMemcpyAsync(out, in.data<float>(), in.memory_size(),
-                                        cudaMemcpyHostToHost, *stream_));
-
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data<float>(), size,
+                                           cudaMemcpyDeviceToDevice, *stream_));
+    } else {
+      PADDLE_THROW("Unknown device for converter");
+    }
+    cudaStreamSynchronize(*stream_);
+  }
+  // NOTE in is GPU memory.
+  virtual void operator()(const void* in, LoDTensor* out,
+                          size_t max_size) override {
+    PADDLE_ENFORCE(in != nullptr);
+    PADDLE_ENFORCE(stream_ != nullptr);
+    const auto& place = out->place();
+    size_t size = out->memory_size();
+    PADDLE_ENFORCE_LE(size, max_size);
+    if (is_cpu_place(place)) {
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data<float>(), in, size,
+                                           cudaMemcpyDeviceToHost, *stream_));
+    } else if (is_gpu_place(place)) {
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data<float>(), in, size,
+                                           cudaMemcpyDeviceToDevice, *stream_));
     } else {
       PADDLE_THROW("Unknown device for converter");
     }
@@ -50,7 +66,8 @@ class DefaultInputConverter : public EngineInputConverter {
   }
 };
 
-REGISTER_TENSORRT_INPUT_CONVERTER(default, DefaultInputConverter);
+// fluid LodTensor <-> tensorrt ITensor
+REGISTER_TENSORRT_IO_CONVERTER(default, DefaultIOConverter);
 
 }  // namespace tensorrt
 }  // namespace inference

paddle/fluid/inference/tensorrt/convert/io_converter.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
 #include <unordered_map>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/utils/singleton.h"
@@ -25,43 +26,57 @@ namespace tensorrt {
 using framework::LoDTensor;
 
 /*
- * Convert Input from Fluid to an Engine.
- * TensorRT's ITensor follows row major, NCHW. Fluid is also row major, so in
- * most cases just need to copy the data.
+ * Convert Input from Fluid to TensorRT Engine.
+ * Convert Output from TensorRT Engine to Fluid.
+ *
+ * Note that TensorRT's ITensor follows row major, NCHW. Fluid is also row
+ * major,
+ * so in the default case just need to copy the data.
  */
-class EngineInputConverter {
+class EngineIOConverter {
  public:
-  EngineInputConverter() {}
+  EngineIOConverter() {}
 
   virtual void operator()(const LoDTensor& in, void* out, size_t max_size) {}
+  virtual void operator()(const void* in, LoDTensor* out, size_t max_size) {}
 
   void SetStream(cudaStream_t* stream) { stream_ = stream; }
 
-  static void Run(const std::string& in_op_type, const LoDTensor& in, void* out,
-                  size_t max_size, cudaStream_t* stream) {
+  static void ConvertInput(const std::string& op_type, const LoDTensor& in,
+                           void* out, size_t max_size, cudaStream_t* stream) {
     PADDLE_ENFORCE(stream != nullptr);
-    auto* converter = Registry<EngineInputConverter>::Lookup(
-        in_op_type, "default" /* default_type */);
+    auto* converter = Registry<EngineIOConverter>::Lookup(
+        op_type, "default" /* default_type */);
     PADDLE_ENFORCE_NOT_NULL(converter);
     converter->SetStream(stream);
     (*converter)(in, out, max_size);
   }
 
-  virtual ~EngineInputConverter() {}
+  static void ConvertOutput(const std::string& op_type, const void* in,
+                            LoDTensor* out, size_t max_size,
+                            cudaStream_t* stream) {
+    PADDLE_ENFORCE(stream != nullptr);
+    auto* converter = Registry<EngineIOConverter>::Lookup(
+        op_type, "default" /* default_type */);
+    PADDLE_ENFORCE_NOT_NULL(converter);
+    converter->SetStream(stream);
+    (*converter)(in, out, max_size);
+  }
+
+  virtual ~EngineIOConverter() {}
 
  protected:
   cudaStream_t* stream_{nullptr};
 };
 
+#define REGISTER_TENSORRT_IO_CONVERTER(op_type__, Converter__)        \
+  struct trt_io_##op_type__##_converter {                             \
+    trt_io_##op_type__##_converter() {                                \
+      Registry<EngineIOConverter>::Register<Converter__>(#op_type__); \
+    }                                                                 \
+  };                                                                  \
+  trt_io_##op_type__##_converter trt_io_##op_type__##_converter__;
+
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle
-
-#define REGISTER_TENSORRT_INPUT_CONVERTER(in_op_type__, Converter__) \
-  struct trt_input_##in_op_type__##_converter {                      \
-    trt_input_##in_op_type__##_converter() {                         \
-      ::paddle::inference::Registry<EngineInputConverter>::Register< \
-          Converter__>(#in_op_type__);                               \
-    }                                                                \
-  };                                                                 \
-  trt_input_##in_op_type__##_converter trt_input_##in_op_type__##_converter__;

paddle/fluid/inference/tensorrt/convert/test_activation_op.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/place.h"
@@ -26,7 +27,7 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-void Compare(float input, float expect) {
+void Compare(const std::string op_type, float input, float expect) {
   framework::Scope scope;
   platform::CUDAPlace place;
   platform::CUDADeviceContext ctx(place);
@@ -35,6 +36,7 @@ void Compare(float input, float expect) {
   auto x_var = scope.Var("X");
   auto x_tensor = x_var->GetMutable<framework::LoDTensor>();
   x_tensor->Resize({1, 1});
+  x_tensor->mutable_data<float>(place);
   std::vector<float> init;
   init.push_back(input);
   framework::TensorFromVector(init, ctx, x_tensor);
@@ -45,14 +47,15 @@ void Compare(float input, float expect) {
   out_tensor->mutable_data<float>(place);
 
   framework::OpDesc op_desc;
-  op_desc.SetType("relu");
+  op_desc.SetType(op_type);
   op_desc.SetInput("X", {"X"});
   op_desc.SetOutput("Out", {"Out"});
 
-  auto relu_op = framework::OpRegistry::CreateOp(*op_desc.Proto());
+  auto op = framework::OpRegistry::CreateOp(*op_desc.Proto());
 
   // run fluid op
-  relu_op->Run(scope, place);
+  op->Run(scope, place);
+  // get fluid output
   std::vector<float> out1;
   framework::TensorToVector(*out_tensor, ctx, &out1);
 
@@ -63,21 +66,28 @@ void Compare(float input, float expect) {
   engine->InitNetwork();
   engine->DeclareInput("X", nvinfer1::DataType::kFLOAT,
                        nvinfer1::DimsCHW{1, 1, 1});
-
+  // convert op
   OpConverter op_converter;
   op_converter.ConvertOp(*op_desc.Proto(), engine);
 
   engine->DeclareOutput("Out");
   engine->FreezeNetwork();
-  engine->SetInputFromCPU("X", &input, 1 * sizeof(float));
 
-  // run tensorrt op
+  // convert LoDTensor to ITensor
+  size_t size = x_tensor->memory_size();
+  EngineIOConverter::ConvertInput(op_type, *x_tensor,
+                                  engine->buffer("X").buffer, size, &stream);
+  // run tensorrt Outp
   engine->Execute(1);
-
-  float out2;
-  engine->GetOutputInCPU("Out", &out2, 1 * sizeof(float));
-
-  ASSERT_EQ(out1[0], out2);
+  // convert ITensor to LoDTensor
+  EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer,
+                                   out_tensor, size, &stream);
+  // get tensorrt output
+  std::vector<float> out2;
+  framework::TensorToVector(*out_tensor, ctx, &out2);
+
+  // compare
+  ASSERT_EQ(out1[0], out2[0]);
   ASSERT_EQ(out1[0], expect);
 
   delete engine;
@@ -85,8 +95,8 @@ void Compare(float input, float expect) {
 }
 
 TEST(OpConverter, ConvertRelu) {
-  Compare(1, 1);   // relu(1) = 1
-  Compare(-5, 0);  // relu(-5) = 0
+  Compare("relu", 1, 1);   // relu(1) = 1
+  Compare("relu", -5, 0);  // relu(-5) = 0
 }
 
 }  // namespace tensorrt

paddle/fluid/inference/tensorrt/convert/test_io_converter.cc

@@ -12,40 +12,63 @@ 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. */
 
+#include <gtest/gtest.h>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
 
-#include <gtest/gtest.h>
-
 namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-class EngineInputConverterTester : public ::testing::Test {
- public:
-  void SetUp() override { tensor.Resize({10, 10}); }
+void IOConverterTester(const platform::DeviceContext& ctx) {
+  cudaStream_t stream;
+  ASSERT_EQ(0, cudaStreamCreate(&stream));
 
-  framework::LoDTensor tensor;
-};
+  // init fluid in_tensor
+  framework::LoDTensor in_tensor;
+  in_tensor.Resize({10, 10});
+  auto place = ctx.GetPlace();
+  in_tensor.mutable_data<float>(place);
+  std::vector<float> init;
+  for (int64_t i = 0; i < 10 * 10; ++i) {
+    init.push_back(i);
+  }
+  framework::TensorFromVector(init, ctx, &in_tensor);
 
-TEST_F(EngineInputConverterTester, DefaultCPU) {
+  // init tensorrt buffer
   void* buffer;
-  tensor.mutable_data<float>(platform::CPUPlace());
-  ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
+  size_t size = in_tensor.memory_size();
+  ASSERT_EQ(cudaMalloc(&buffer, size), 0);
 
-  cudaStream_t stream;
-  EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
-                            &stream);
+  // convert fluid in_tensor to tensorrt buffer
+  EngineIOConverter::ConvertInput("test", in_tensor, buffer, size, &stream);
+
+  // convert tensorrt buffer to fluid out_tensor
+  framework::LoDTensor out_tensor;
+  out_tensor.Resize({10, 10});
+  out_tensor.mutable_data<float>(place);
+  EngineIOConverter::ConvertOutput("test", buffer, &out_tensor, size, &stream);
+
+  // compare in_tensor and out_tensor
+  std::vector<float> result;
+  framework::TensorToVector(out_tensor, ctx, &result);
+  EXPECT_EQ(init.size(), result.size());
+  for (size_t i = 0; i < init.size(); i++) {
+    EXPECT_EQ(init[i], result[i]);
+  }
+  cudaStreamDestroy(stream);
 }
 
-TEST_F(EngineInputConverterTester, DefaultGPU) {
-  void* buffer;
-  tensor.mutable_data<float>(platform::CUDAPlace());
-  ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
+TEST(EngineIOConverterTester, DefaultCPU) {
+  platform::CPUPlace place;
+  platform::CPUDeviceContext ctx(place);
+  IOConverterTester(ctx);
+}
 
-  cudaStream_t stream;
-  EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
-                            &stream);
+TEST(EngineIOConverterTester, DefaultGPU) {
+  platform::CUDAPlace place;
+  platform::CUDADeviceContext ctx(place);
+  IOConverterTester(ctx);
 }
 
 }  // namespace tensorrt

python/paddle/fluid/tests/book/notest_understand_sentiment.py

@@ -170,7 +170,7 @@ def train(word_dict,
         assert save_dirname is None
 
     adagrad = fluid.optimizer.Adagrad(learning_rate=0.002)
-    optimize_ops, params_grads = adagrad.minimize(cost)
+    adagrad.minimize(cost)
 
     train_data = paddle.batch(
         paddle.reader.shuffle(

python/paddle/fluid/tests/book/test_fit_a_line.py

@@ -33,7 +33,7 @@ def train(use_cuda, save_dirname, is_local):
     avg_cost = fluid.layers.mean(cost)
 
     sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+    sgd_optimizer.minimize(avg_cost)
 
     BATCH_SIZE = 20
 

python/paddle/fluid/tests/book/test_image_classification.py

@@ -125,7 +125,7 @@ def train(net_type, use_cuda, save_dirname, is_local):
     test_program = fluid.default_main_program().clone(for_test=True)
 
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
-    optimize_ops, params_grads = optimizer.minimize(avg_cost)
+    optimizer.minimize(avg_cost)
 
     BATCH_SIZE = 128
     PASS_NUM = 1

python/paddle/fluid/tests/book/test_label_semantic_roles.py

@@ -175,7 +175,7 @@ def train(use_cuda, save_dirname=None, is_local=True):
             decay_steps=100000,
             decay_rate=0.5,
             staircase=True))
-    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+    sgd_optimizer.minimize(avg_cost)
 
     # TODO(qiao)
     # add dependency track and move this config before optimizer

python/paddle/fluid/tests/book/test_machine_translation.py

@@ -185,7 +185,7 @@ def train_main(use_cuda, is_sparse, is_local=True):
         learning_rate=1e-4,
         regularization=fluid.regularizer.L2DecayRegularizer(
             regularization_coeff=0.1))
-    optimize_ops, params_grads = optimizer.minimize(avg_cost)
+    optimizer.minimize(avg_cost)
 
     train_data = paddle.batch(
         paddle.reader.shuffle(

python/paddle/fluid/tests/book/test_recognize_digits.py

@@ -95,7 +95,7 @@ def train(nn_type,
     test_program = fluid.default_main_program().clone(for_test=True)
 
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
-    optimize_ops, params_grads = optimizer.minimize(avg_loss)
+    optimizer.minimize(avg_loss)
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 

python/paddle/fluid/tests/book/test_recommender_system.py

@@ -160,7 +160,7 @@ def train(use_cuda, save_dirname, is_local=True):
     test_program = fluid.default_main_program().clone(for_test=True)
 
     sgd_optimizer = SGDOptimizer(learning_rate=0.2)
-    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+    sgd_optimizer.minimize(avg_cost)
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 

python/paddle/fluid/tests/book/test_word2vec.py

@@ -101,7 +101,7 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True):
         avg_cost = fluid.layers.mean(pd())
 
     sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+    sgd_optimizer.minimize(avg_cost)
 
     train_reader = paddle.batch(
         paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE)

python/paddle/fluid/tests/unittests/test_dist_transpiler.py

+#   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 unittest
+
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+import paddle.fluid.layers as layers
+from paddle.fluid.transpiler.distribute_transpiler import delete_ops
+import numpy
+
+
+class TestDistTranspiler(unittest.TestCase):
+    def setUp(self):
+        self.trainer_id = 0
+        self.trainers = 2
+        self.pservers = 2
+        self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175"
+        self.current_pserver_ep = "127.0.0.1:6174"
+
+    def net_conf(self):
+        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
+
+        y_predict = fluid.layers.fc(input=x,
+                                    size=1000,
+                                    act=None,
+                                    param_attr=fluid.ParamAttr(name='fc_w'))
+
+        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+        avg_cost = fluid.layers.mean(cost)
+        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1)
+
+        optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+        return optimize_ops, params_grads
+
+    def test_transpiler(self):
+        trainer = self.get_trainer()
+        pserver, startup = self.get_pserver(self.current_pserver_ep)
+
+        self.assertEqual([op.type for op in trainer.global_block().ops],
+                         self.get_expect_trainer_ops())
+
+        self.assertEqual(len(pserver.blocks), 3)
+        # block0: listen_and_serv
+        self.assertEqual([op.type for op in pserver.blocks[0].ops],
+                         ["listen_and_serv"])
+        # block2: optimize pass
+        self.assertEqual([op.type for op in pserver.blocks[1].ops],
+                         ["sum", "scale", "sgd"])
+
+        # confirm startup program
+
+        self.assertEqual([op.type for op in startup.global_block().ops], [
+            "fill_constant", "fill_constant", "uniform_random", "uniform_random"
+        ])
+
+        # the variable #fc_w will be split into two blocks 
+        fc_w_var = startup.global_block().var("fc_w.block1")
+        self.assertEqual(fc_w_var.shape, (500, 1000))
+
+    def get_main_program(self):
+        main = fluid.Program()
+
+        with fluid.program_guard(main):
+            self.net_conf()
+
+        return main
+
+    def get_expect_trainer_ops(self):
+        trainer = fluid.Program()
+
+        with fluid.program_guard(trainer):
+            optimize_ops, params_grads = self.net_conf()
+
+        delete_ops(trainer.global_block(), optimize_ops)
+        return [op.type for op in trainer.global_block().ops
+                ] + ["split_byref", "send", "concat"]
+
+    def get_trainer(self):
+        return self._transpiler_instance().get_trainer_program()
+
+    def get_pserver(self, ep):
+        t = self._transpiler_instance()
+        pserver = t.get_pserver_program(ep)
+        startup = t.get_startup_program(ep, pserver)
+        return pserver, startup
+
+    def _transpiler_instance(self):
+        main = self.get_main_program()
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            self.trainer_id,
+            program=main,
+            pservers=self.pserver_eps,
+            trainers=self.trainers)
+        return t
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/transpiler/__init__.py

@@ -11,6 +11,7 @@
 # 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.
+
 from distribute_transpiler import DistributeTranspiler
 from inference_transpiler import InferenceTranspiler
 from memory_optimization_transpiler import memory_optimize, release_memory

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -17,7 +17,7 @@ from __future__ import print_function
 import math
 
 import distributed_splitter as splitter
-from .. import core
+from .. import core, framework
 from ..framework import Program, default_main_program, \
                         default_startup_program, \
                         Variable, Parameter, grad_var_name
@@ -417,7 +417,7 @@ class DistributeTranspiler:
         def __append_optimize_op__(op, block, grad_to_block_id):
             if self._is_opt_op(op):
                 self._append_pserver_ops(block, op, endpoint, grad_to_block_id,
-                                         default_main_program())
+                                         self.origin_program)
             else:
                 self._append_pserver_non_opt_ops(block, op)
 

tools/manylinux1/README.md

@@ -28,3 +28,38 @@ git clone https://github.com/paddlepaddle/paddle
 cd paddle/tools/manylinux1
 REPO=[yourrepo] ./build_all.sh
 ```
+
+## Build PaddlePaddle for the different Python ABIs
+
+Choose one of the following Python ABI and set the correct environment variables.
+
+- cp27-cp27m
+
+  ```bash
+  export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:}
+  export PATH=/opt/python/cp27-cp27m/bin/:${PATH}
+  export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python
+        -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7
+        -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so"
+  ```
+
+- cp27-cp27mu
+
+  ```bash
+  export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:}
+  export PATH=/opt/python/cp27-cp27mu/bin/:${PATH}
+  export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python
+        -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7
+        -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so"
+  ```
+
+And then add the `PYTHON_FLAGS` as your cmake flags:
+
+```bash
+cmake ..
+  ${PYTHON_FLAGS} \
+  -DWITH_GPU=OFF \
+  ...
+```
+
+You can find more details about cmake flags at [here](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html#appendix-build-options)