Merge remote-tracking branch 'ups/develop' into refine/mklml/dyload

.pre-commit-config.yaml

@@ -23,7 +23,7 @@ repos:
     -   id: clang-format-with-version-check
         name: clang-format
         description: Format files with ClangFormat.
-        entry: bash ./.clang_format.hook -i
+        entry: bash ./tools/codestyle/clang_format.hook -i
         language: system
         files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$
 -   repo: local
@@ -52,7 +52,7 @@ repos:
     hooks:
     -   id: copyright_checker
         name: copyright_checker
-        entry: python ./.copyright.hook
+        entry: python ./tools/codestyle/copyright.hook
         language: system
         files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto|py)$
         exclude: (?!.*third_party)^.*$ | (?!.*book)^.*$

benchmark/fluid/Dockerfile

 FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04
+
+# Use UBUNTU_MIRROR can speed up apt-get speed.
+# ARG UBUNTU_MIRROR
+# RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
+
 RUN apt-get update && apt-get install -y python python-pip iputils-ping libgtk2.0-dev wget vim net-tools iftop python-opencv
 RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/lib/libcudnn.so && ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/lib/libnccl.so
-RUN pip install -U pip
-RUN pip install -U kubernetes paddlepaddle
 
 # IMPORTANT:
 # Add "ENV http_proxy=http://ip:port" if your download is slow, and don't forget to unset it at runtime.
+# exmaple: unset http_proxy && unset https_proxy && python fluid_benchmark.py ...
+
+RUN pip install -U pip
+RUN pip install -U kubernetes paddlepaddle
 
 RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()\npaddle.dataset.flowers.fetch()" | python'
 RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.mnist.train()\npaddle.dataset.mnist.test()\npaddle.dataset.imdb.fetch()" | python'
@@ -14,9 +21,11 @@ RUN pip uninstall -y paddlepaddle && mkdir /workspace
 
 ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin
 ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/k8s_tools.py /root
+RUN chmod +x /usr/bin/paddle_k8s
 
 ADD *.whl /
-RUN pip install /*.whl && rm -f /*.whl && chmod +x /usr/bin/paddle_k8s
+RUN pip install /*.whl && rm -f /*.whl 
 
 ENV LD_LIBRARY_PATH=/usr/local/lib
-ADD fluid_benchmark.py recordio_converter.py models/ /workspace/
+ADD fluid_benchmark.py recordio_converter.py args.py recordio_converter.py run.sh run_fluid_benchmark.sh /workspace/
+ADD models/ /workspace/models/

benchmark/fluid/fluid_benchmark.py

@@ -264,8 +264,6 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
                     break
             else:
                 loss, = exe.run([avg_loss.name], feed=feeder.feed(data))
-            if args.update_method == "pserver":
-                exe.bcast_params()
             if args.use_reader_op:
                 num_samples += args.batch_size * args.gpus
             else:
@@ -301,9 +299,18 @@ def print_train_time(start_time, end_time, num_samples):
           (num_samples, train_elapsed, examples_per_sec))
 
 
+def print_paddle_envs():
+    print('----------- Configuration envs -----------')
+    for k in os.environ:
+        if "PADDLE_" in k:
+            print "ENV %s:%s" % (k, os.environ[k])
+    print('------------------------------------------------')
+
+
 def main():
     args = parse_args()
     print_arguments(args)
+    print_paddle_envs()
 
     # the unique trainer id, starting from 0, needed by trainer
     # only

benchmark/fluid/kube_gen_job.py

@@ -17,6 +17,7 @@ import copy
 import argparse
 import random
 import os
+import copy
 from kube_templates import pserver, trainer, envs
 
 
@@ -109,10 +110,9 @@ def gen_job():
 
     envs.append({"name": "PADDLE_JOB_NAME", "value": args.jobname})
     envs.append({"name": "PADDLE_TRAINERS", "value": str(args.trainers)})
-    envs.append({"name": "PSERVERS", "value": str(args.pservers)})
+    envs.append({"name": "PADDLE_PSERVERS", "value": str(args.pservers)})
     envs.append({"name": "ENTRY", "value": args.entry})
     envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
-    envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
     # NOTE: these directories below are cluster specific, please modify
     # this settings before you run on your own cluster.
     envs.append({
@@ -166,7 +166,7 @@ def gen_job():
     tn["spec"]["template"]["spec"]["volumes"] = volumes
     tn_container["volumeMounts"] = volumeMounts
 
-    ps_container["env"] = envs
+    ps_container["env"] = copy.deepcopy(envs)
     ps_container["env"].append({
         "name": "PADDLE_TRAINING_ROLE",
         "value": "PSERVER"

doc/fluid/api/gen_doc.sh

 #!/bin/bash
 python gen_doc.py layers --submodules control_flow device io nn ops tensor detection learning_rate_scheduler metric > layers.rst
 
-for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer
+for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer transpiler
 do
   python gen_doc.py ${module} > ${module}.rst
 done

doc/fluid/api/transpiler.rst

+..  THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
+    !DO NOT EDIT THIS FILE MANUALLY!
+
+==========
+transpiler
+==========
+
+DistributeTranspiler
+--------------------
+
+..  autoclass:: paddle.fluid.transpiler.DistributeTranspiler
+    :members:
+    :noindex:
+
+InferenceTranspiler
+-------------------
+
+..  autoclass:: paddle.fluid.transpiler.InferenceTranspiler
+    :members:
+    :noindex:
+
+memory_optimize
+---------------
+
+..  autofunction:: paddle.fluid.transpiler.memory_optimize
+    :noindex:
+
+release_memory
+--------------
+
+..  autofunction:: paddle.fluid.transpiler.release_memory
+    :noindex:
+
+HashName
+--------
+
+..  autoclass:: paddle.fluid.transpiler.HashName
+    :members:
+    :noindex:
+
+RoundRobin
+----------
+
+..  autoclass:: paddle.fluid.transpiler.RoundRobin
+    :members:
+    :noindex:

doc/fluid/howto/inference/build_and_install_lib_cn.rst

@@ -13,6 +13,7 @@ cpu_noavx_openblas       `fluid.tgz <https://guest:@paddleci.ngrok.io/repository
 cuda7.5_cudnn5_avx_mkl   `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/fluid.tgz>`_
 cuda8.0_cudnn5_avx_mkl   `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/fluid.tgz>`_
 cuda8.0_cudnn7_avx_mkl   `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/fluid.tgz>`_
+cuda9.0_cudnn7_avx_mkl   `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/fluid.tgz>`_
 ======================   ========================================
 
 从源码编译

paddle/contrib/inference/demo/simple_on_word2vec.cc

@@ -40,10 +40,9 @@ void Main(bool use_gpu) {
     //# 2. Prepare input.
     int64_t data[4] = {1, 2, 3, 4};
 
-    PaddleBuf buf{.data = data, .length = sizeof(data)};
     PaddleTensor tensor{.name = "",
                         .shape = std::vector<int>({4, 1}),
-                        .data = buf,
+                        .data = PaddleBuf(data, sizeof(data)),
                         .dtype = PaddleDType::INT64};
 
     // For simplicity, we set all the slots with the same data.
@@ -55,14 +54,12 @@ void Main(bool use_gpu) {
 
     //# 4. Get output.
     ASSERT_EQ(outputs.size(), 1UL);
-    LOG(INFO) << "output buffer size: " << outputs.front().data.length;
-    const size_t num_elements = outputs.front().data.length / sizeof(float);
+    LOG(INFO) << "output buffer size: " << outputs.front().data.length();
+    const size_t num_elements = outputs.front().data.length() / sizeof(float);
     // The outputs' buffers are in CPU memory.
     for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
-      LOG(INFO) << static_cast<float*>(outputs.front().data.data)[i];
+      LOG(INFO) << static_cast<float*>(outputs.front().data.data())[i];
     }
-    // TODO(Superjomn): this is should be free automatically
-    free(outputs[0].data.data);
   }
 }
 
@@ -86,10 +83,9 @@ void MainThreads(int num_threads, bool use_gpu) {
       for (int batch_id = 0; batch_id < num_batches; ++batch_id) {
         // 2. Dummy Input Data
         int64_t data[4] = {1, 2, 3, 4};
-        PaddleBuf buf{.data = data, .length = sizeof(data)};
         PaddleTensor tensor{.name = "",
                             .shape = std::vector<int>({4, 1}),
-                            .data = buf,
+                            .data = PaddleBuf(data, sizeof(data)),
                             .dtype = PaddleDType::INT64};
         std::vector<PaddleTensor> inputs(4, tensor);
         std::vector<PaddleTensor> outputs;
@@ -99,13 +95,13 @@ void MainThreads(int num_threads, bool use_gpu) {
         // 4. Get output.
         ASSERT_EQ(outputs.size(), 1UL);
         LOG(INFO) << "TID: " << tid << ", "
-                  << "output buffer size: " << outputs.front().data.length;
-        const size_t num_elements = outputs.front().data.length / sizeof(float);
+                  << "output buffer size: " << outputs.front().data.length();
+        const size_t num_elements =
+            outputs.front().data.length() / sizeof(float);
         // The outputs' buffers are in CPU memory.
         for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
-          LOG(INFO) << static_cast<float*>(outputs.front().data.data)[i];
+          LOG(INFO) << static_cast<float*>(outputs.front().data.data())[i];
         }
-        free(outputs[0].data.data);
       }
     });
   }

paddle/contrib/inference/paddle_inference_api.cc

@@ -13,3 +13,53 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/contrib/inference/paddle_inference_api.h"
+
+namespace paddle {
+
+PaddleBuf::PaddleBuf(PaddleBuf&& other)
+    : data_(other.data_),
+      length_(other.length_),
+      memory_owned_(other.memory_owned_) {
+  other.memory_owned_ = false;
+  other.data_ = nullptr;
+  other.length_ = 0;
+}
+
+PaddleBuf::PaddleBuf(const PaddleBuf& other) { *this = other; }
+
+PaddleBuf& PaddleBuf::operator=(const PaddleBuf& other) {
+  // only the buffer with external memory can be copied
+  assert(!other.memory_owned_);
+  data_ = other.data_;
+  length_ = other.length_;
+  memory_owned_ = other.memory_owned_;
+  return *this;
+}
+
+void PaddleBuf::Resize(size_t length) {
+  // Only the owned memory can be reset, the external memory can't be changed.
+  if (length_ == length) return;
+  assert(memory_owned_);
+  Free();
+  data_ = new char[length];
+  length_ = length;
+  memory_owned_ = true;
+}
+
+void PaddleBuf::Reset(void* data, size_t length) {
+  Free();
+  memory_owned_ = false;
+  data_ = data;
+  length_ = length;
+}
+
+void PaddleBuf::Free() {
+  if (memory_owned_ && data_) {
+    assert(length_ > 0);
+    delete static_cast<char*>(data_);
+    data_ = nullptr;
+    length_ = 0;
+  }
+}
+
+}  // namespace paddle
\ No newline at end of file

paddle/contrib/inference/paddle_inference_api.h

@@ -21,6 +21,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <cassert>
 #include <memory>
 #include <string>
 #include <vector>
@@ -32,12 +33,38 @@ enum PaddleDType {
   INT64,
 };
 
-struct PaddleBuf {
-  void* data;     // pointer to the data memory.
-  size_t length;  // number of memory bytes.
+class PaddleBuf {
+ public:
+  PaddleBuf() = default;
+  PaddleBuf(PaddleBuf&& other);
+  // Copy only available when memory is managed externally.
+  explicit PaddleBuf(const PaddleBuf&);
+  PaddleBuf& operator=(const PaddleBuf&);
+  // Do not own the memory.
+  PaddleBuf(void* data, size_t length)
+      : data_(data), length_(length), memory_owned_{false} {}
+  // Own memory.
+  PaddleBuf(size_t length)
+      : data_(new char[length]), length_(length), memory_owned_(true) {}
+  // Resize to `length` bytes.
+  void Resize(size_t length);
+  // Reset to external memory.
+  void Reset(void* data, size_t length);
+  bool empty() const { return length_ == 0; }
+  void* data() const { return data_; }
+  size_t length() const { return length_; }
+
+  ~PaddleBuf() { Free(); }
+
+ private:
+  void Free();
+  void* data_{nullptr};  // pointer to the data memory.
+  size_t length_{0};     // number of memory bytes.
+  bool memory_owned_{true};
 };
 
 struct PaddleTensor {
+  PaddleTensor() = default;
   std::string name;  // variable name.
   std::vector<int> shape;
   // TODO(Superjomn) for LoD support, add a vector<vector<int>> field if needed.
@@ -67,8 +94,9 @@ class PaddlePredictor {
 
   // Predict an record.
   // The caller should be responsible for allocating and releasing the memory of
-  // `inputs`. `inputs` should be alive until Run returns. caller should be
-  // responsible for releasing the memory of `output_data`.
+  // `inputs`. `inputs` should be available until Run returns. Caller should be
+  // responsible for the output tensor's buffer, either allocated or passed from
+  // outside.
   virtual bool Run(const std::vector<PaddleTensor>& inputs,
                    std::vector<PaddleTensor>* output_data) = 0;
 

paddle/contrib/inference/paddle_inference_api_anakin_engine.cc

@@ -48,7 +48,7 @@ bool PaddleInferenceAnakinPredictor::Run(
     auto d_tensor_in_p = executor_.get_in(input.name);
     float *d_data_p = d_tensor_in_p->mutable_data();
     if (cudaMemcpy(d_data_p,
-                   static_cast<float *>(input.data.data),
+                   static_cast<float *>(input.data.data()),
                    d_tensor_in_p->valid_size() * sizeof(float),
                    cudaMemcpyHostToDevice) != 0) {
       LOG(ERROR) << "copy data from CPU to GPU error";
@@ -65,8 +65,11 @@ bool PaddleInferenceAnakinPredictor::Run(
   for (auto &output : *output_data) {
     auto *tensor = executor_.get_out(output.name);
     output.shape = tensor->shape();
+    if (output.data.length() < tensor->valid_size() * sizeof(float)) {
+      output.data.Resize(tensor->valid_size() * sizeof(float));
+    }
     // Copy data from GPU -> CPU
-    if (cudaMemcpy(output.data.data,
+    if (cudaMemcpy(output.data.data(),
                    tensor->mutable_data(),
                    tensor->valid_size() * sizeof(float),
                    cudaMemcpyDeviceToHost) != 0) {

paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc

@@ -37,28 +37,26 @@ TEST(inference, anakin) {
 
   float data[1 * 3 * 224 * 224] = {1.0f};
 
-  PaddleBuf buf{.data = data, .length = sizeof(data)};
   PaddleTensor tensor{.name = "input_0",
                       .shape = std::vector<int>({1, 3, 224, 224}),
-                      .data = buf,
+                      .data = PaddleBuf(data, sizeof(data)),
                       .dtype = PaddleDType::FLOAT32};
 
   // For simplicity, we set all the slots with the same data.
-  std::vector<PaddleTensor> paddle_tensor_feeds(1, tensor);
+  std::vector<PaddleTensor> paddle_tensor_feeds;
+  paddle_tensor_feeds.emplace_back(std::move(tensor));
 
-  float data_out[1000];
-
-  PaddleBuf buf_out{.data = data_out, .length = sizeof(data)};
   PaddleTensor tensor_out{.name = "prob_out",
                           .shape = std::vector<int>({1000, 1}),
-                          .data = buf_out,
+                          .data = PaddleBuf(),
                           .dtype = PaddleDType::FLOAT32};
 
-  std::vector<PaddleTensor> outputs(1, tensor_out);
+  std::vector<PaddleTensor> outputs;
+  outputs.emplace_back(std::move(tensor_out));
 
   ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs));
 
-  float* data_o = static_cast<float*>(outputs[0].data.data);
+  float* data_o = static_cast<float*>(outputs[0].data.data());
   for (size_t j = 0; j < 1000; ++j) {
     LOG(INFO) << "output[" << j << "]: " << data_o[j];
   }

paddle/contrib/inference/paddle_inference_api_impl.cc

@@ -178,8 +178,8 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
 
     // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
     std::memcpy(static_cast<void *>(input_ptr),
-                inputs[i].data.data,
-                inputs[i].data.length);
+                inputs[i].data.data(),
+                inputs[i].data.length());
     feeds->push_back(input);
   }
   return true;
@@ -241,10 +241,11 @@ bool NativePaddlePredictor::GetFetch(
     }
 
     outputs->at(i).shape = shape;
-    outputs->at(i).data.length = sizeof(float) * data.size();
-    outputs->at(i).data.data = malloc(outputs->at(i).data.length);
-    std::memcpy(
-        outputs->at(i).data.data, data.data(), outputs->at(i).data.length);
+    auto &buffer = outputs->at(i).data;
+    if (buffer.empty() || buffer.length() < sizeof(float) * data.size()) {
+      buffer.Resize(sizeof(float) * data.size());
+    }
+    std::memcpy(buffer.data(), data.data(), buffer.length());
     outputs->at(i).dtype = PaddleDType::FLOAT32;
     // TODO(panyx0718): support other types? fill tensor name? avoid a copy.
   }

paddle/contrib/inference/test_paddle_inference_api_impl.cc

@@ -27,13 +27,12 @@ namespace paddle {
 
 PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
   PaddleTensor pt;
-  pt.data.data = t->data<void>();
 
   if (t->type() == typeid(int64_t)) {
-    pt.data.length = t->numel() * sizeof(int64_t);
+    pt.data.Reset(t->data<void>(), t->numel() * sizeof(int64_t));
     pt.dtype = PaddleDType::INT64;
   } else if (t->type() == typeid(float)) {
-    pt.data.length = t->numel() * sizeof(float);
+    pt.data.Reset(t->data<void>(), t->numel() * sizeof(float));
     pt.dtype = PaddleDType::FLOAT32;
   } else {
     LOG(FATAL) << "unsupported type.";
@@ -79,8 +78,8 @@ void MainWord2Vec(bool use_gpu) {
   std::vector<PaddleTensor> outputs;
   ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs));
   ASSERT_EQ(outputs.size(), 1UL);
-  size_t len = outputs[0].data.length;
-  float* data = static_cast<float*>(outputs[0].data.data);
+  size_t len = outputs[0].data.length();
+  float* data = static_cast<float*>(outputs[0].data.data());
   for (size_t j = 0; j < len / sizeof(float); ++j) {
     ASSERT_LT(data[j], 1.0);
     ASSERT_GT(data[j], -1.0);
@@ -103,8 +102,6 @@ void MainWord2Vec(bool use_gpu) {
     EXPECT_LT(lod_data[i] - data[i], 1e-3);
     EXPECT_GT(lod_data[i] - data[i], -1e-3);
   }
-
-  free(outputs[0].data.data);
 }
 
 void MainImageClassification(bool use_gpu) {
@@ -143,13 +140,12 @@ void MainImageClassification(bool use_gpu) {
   std::vector<PaddleTensor> outputs;
   ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs));
   ASSERT_EQ(outputs.size(), 1UL);
-  size_t len = outputs[0].data.length;
-  float* data = static_cast<float*>(outputs[0].data.data);
+  size_t len = outputs[0].data.length();
+  float* data = static_cast<float*>(outputs[0].data.data());
   float* lod_data = output1.data<float>();
   for (size_t j = 0; j < len / sizeof(float); ++j) {
     EXPECT_NEAR(lod_data[j], data[j], 1e-3);
   }
-  free(data);
 }
 
 void MainThreadsWord2Vec(bool use_gpu) {
@@ -192,8 +188,8 @@ void MainThreadsWord2Vec(bool use_gpu) {
 
       // check outputs range
       ASSERT_EQ(local_outputs.size(), 1UL);
-      const size_t len = local_outputs[0].data.length;
-      float* data = static_cast<float*>(local_outputs[0].data.data);
+      const size_t len = local_outputs[0].data.length();
+      float* data = static_cast<float*>(local_outputs[0].data.data());
       for (size_t j = 0; j < len / sizeof(float); ++j) {
         ASSERT_LT(data[j], 1.0);
         ASSERT_GT(data[j], -1.0);
@@ -205,7 +201,6 @@ void MainThreadsWord2Vec(bool use_gpu) {
       for (int i = 0; i < refs[tid].numel(); ++i) {
         EXPECT_NEAR(ref_data[i], data[i], 1e-3);
       }
-      free(data);
     });
   }
   for (int i = 0; i < num_jobs; ++i) {
@@ -251,14 +246,13 @@ void MainThreadsImageClassification(bool use_gpu) {
 
       // check outputs correctness
       ASSERT_EQ(local_outputs.size(), 1UL);
-      const size_t len = local_outputs[0].data.length;
-      float* data = static_cast<float*>(local_outputs[0].data.data);
+      const size_t len = local_outputs[0].data.length();
+      float* data = static_cast<float*>(local_outputs[0].data.data());
       float* ref_data = refs[tid].data<float>();
       EXPECT_EQ(refs[tid].numel(), len / sizeof(float));
       for (int i = 0; i < refs[tid].numel(); ++i) {
         EXPECT_NEAR(ref_data[i], data[i], 1e-3);
       }
-      free(data);
     });
   }
   for (int i = 0; i < num_jobs; ++i) {

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -57,6 +57,7 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
   for (auto &p : params) {
     grad_names_.insert(GradVarName(p));
   }
+  balance_vars_.resize(places_.size(), 0);
 }
 
 void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result,
@@ -140,11 +141,30 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp(
          checker(op.InputArgumentNames(), recv_vars);
 }
 
+size_t MultiDevSSAGraphBuilder::GetAppropriateDeviceID(
+    const std::vector<std::string> &var_names) const {
+  int64_t numel_sum = 0;
+  for (auto var_name : var_names) {
+    auto var_desc = all_vars_.at(var_name);
+    PADDLE_ENFORCE_NOT_NULL(var_desc);
+    auto dim = framework::make_ddim(var_desc->GetShape());
+    int64_t numel = framework::product(dim);
+    PADDLE_ENFORCE_GT(numel, 0);
+    numel_sum += numel;
+  }
+
+  auto smallest =
+      std::min_element(std::begin(balance_vars_), std::end(balance_vars_));
+  size_t dev_id =
+      static_cast<size_t>(std::distance(std::begin(balance_vars_), smallest));
+  balance_vars_[dev_id] += numel_sum;
+  return dev_id;
+}
+
 std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
     const ProgramDesc &program) const {
-  std::unordered_map<std::string, VarDesc *> all_vars;
   for (auto *var : program.Block(0).AllVars()) {
-    all_vars[var->Name()] = var;
+    all_vars_.emplace(var->Name(), var);
   }
 
   auto graph = new SSAGraph();
@@ -161,35 +181,16 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   auto send_vars = FindDistTrainSendVars(program);
   auto recv_vars = FindDistTrainRecvVars(program);
 
-  std::vector<std::unordered_set<std::string>> var_name_on_devices;
   std::vector<std::unordered_set<std::string>> bcast_var_name_set;
-  var_name_on_devices.resize(places_.size());
   bcast_var_name_set.resize(places_.size());
 
   size_t cur_device_id = 0;
-  std::vector<int64_t> balance_grads(places_.size(), 0);
-
-  auto get_appropriate_dev = [&](std::string &g_name) -> size_t {
-    auto var_desc = all_vars.at(g_name);
-    PADDLE_ENFORCE_NOT_NULL(var_desc);
-    auto dim = framework::make_ddim(var_desc->GetShape());
-    int64_t numel = framework::product(dim);
-    PADDLE_ENFORCE_GE(numel, 0);
-    auto smallest =
-        std::min_element(std::begin(balance_grads), std::end(balance_grads));
-    size_t dev_id =
-        static_cast<size_t>(std::distance(std::begin(balance_grads), smallest));
-    balance_grads[dev_id] += numel;
-    return dev_id;
-  };
-
   bool is_forwarding = true;
+
   for (auto *op : program.Block(0).AllOps()) {
     if (boost::get<int>(
             op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) ==
         static_cast<int>(OpRole::kRPC)) {
-      // append rpc op if program is distributed trainer main program.
-      // always use the first device
       CreateRPCOp(&result, *op);
     } else if (IsDistTrainOp(*op, send_vars, recv_vars)) {
       CreateDistTrainOp(&result, *op);
@@ -199,15 +200,19 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
           BuildStrategy::GradientScaleStrategy::kCustomized) {
         CreateScaleLossGradOp(&result);
       }
+      // This assumes the backward generating code will ensure IsScaleLossOp
+      // is true only for the op that scale the final scalar loss.
+      // It also assumes backward op will always follow the forward op in
+      // the block.
       is_forwarding = false;
     } else {
-      int op_dev_id = GetOpDeviceID(var_name_on_devices, *op);
+      int op_dev_id = GetOpDeviceID(*op);
       if (op_dev_id == -1) {  // var on all device
         CreateComputationalOps(&result, *op, places_.size());
       } else {
         CreateComputationalOp(&result, *op, op_dev_id);
         for (auto &var_name : op->OutputArgumentNames()) {
-          var_name_on_devices[op_dev_id].emplace(var_name);
+          var_name_on_devices_.emplace(var_name, op_dev_id);
         }
       }
       if (!is_forwarding && places_.size() > 1) {
@@ -230,19 +235,22 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
 
               switch (strategy_.reduce_) {
                 case BuildStrategy::ReduceStrategy::kReduce:
-                  cur_device_id = get_appropriate_dev(g_name);
+                  cur_device_id = GetAppropriateDeviceID({g_name});
                   CreateReduceOp(&result, g_name, cur_device_id);
-                  var_name_on_devices[cur_device_id].emplace(g_name);
+                  var_name_on_devices_.emplace(g_name, cur_device_id);
                   bcast_var_name_set[cur_device_id].emplace(p_name);
                   break;
                 case BuildStrategy::ReduceStrategy::kAllReduce:
-                  if (IsSparseGradient(all_vars, g_name)) {
+                  if (IsSparseGradient(g_name)) {
                     CreateReduceOp(&result, g_name, 0);
                     CreateBroadcastOp(&result, g_name, 0);
                   } else {
                     InsertAllReduceOp(&result, g_name);
                   }
                   break;
+                default:
+                  LOG(FATAL) << "Unknown reduce strategy ";
+                  break;
               }
             }
           } catch (boost::bad_get e) {
@@ -261,7 +269,7 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   }
   /*
     Dependency graph has been constructed. However, there are still data
-    harzaeds need to be handled.
+    hazards need to be handled.
    */
   PolishGraphToSupportDataHazards(&result);
 
@@ -273,11 +281,9 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   return std::unique_ptr<SSAGraph>(graph);
 }
 
-bool MultiDevSSAGraphBuilder::IsSparseGradient(
-    const std::unordered_map<std::string, VarDesc *> &all_vars,
-    const std::string &og) const {
-  PADDLE_ENFORCE(all_vars.count(og) != 0);
-  if (all_vars.at(og)->GetType() == proto::VarType::SELECTED_ROWS) {
+bool MultiDevSSAGraphBuilder::IsSparseGradient(const std::string &og) const {
+  PADDLE_ENFORCE(all_vars_.count(og) != 0);
+  if (all_vars_.at(og)->GetType() == proto::VarType::SELECTED_ROWS) {
     return true;
   }
   return false;
@@ -363,24 +369,23 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
   return is_pg_once;
 }
 
-int MultiDevSSAGraphBuilder::GetOpDeviceID(
-    const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
-    const OpDesc &op) const {
+int MultiDevSSAGraphBuilder::GetOpDeviceID(const OpDesc &op) const {
   if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) {
     return -1;
   }
 
-  int var_dev_id = -1;
-  for (auto &var_name : op.InputArgumentNames()) {
-    if (var_dev_id != -1) break;
-    for (size_t i = 0; i < var_name_on_devices.size(); ++i) {
-      if (var_name_on_devices[i].count(var_name)) {
-        var_dev_id = static_cast<int>(i);
-        break;
-      }
+  for (auto &varname : op.InputArgumentNames()) {
+    int dev_id = GetVarDeviceID(varname);
+    if (dev_id != -1) {
+      return dev_id;
     }
   }
-  return var_dev_id;
+  return -1;
+}
+
+int MultiDevSSAGraphBuilder::GetVarDeviceID(const std::string &varname) const {
+  auto got = var_name_on_devices_.find(varname);
+  return got == var_name_on_devices_.end() ? -1 : got->second;
 }
 
 void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const {
@@ -449,6 +454,8 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result,
   return var;
 }
 
+// Find the first occurence of `prev_op_name` and make current `op` depend
+// on it.
 void MultiDevSSAGraphBuilder::ConnectOp(SSAGraph *result, OpHandleBase *op,
                                         const std::string &prev_op_name) const {
   for (auto &prev_op : result->ops_) {
@@ -463,16 +470,66 @@ void MultiDevSSAGraphBuilder::ConnectOp(SSAGraph *result, OpHandleBase *op,
 
 void MultiDevSSAGraphBuilder::CreateDistTrainOp(SSAGraph *result,
                                                 const OpDesc &op) const {
-  CreateComputationalOp(result, op, 0);
+  int op_dev_id = -1;
+  if (op.Type() == "split_byref") {
+    op_dev_id = GetVarDeviceID(op.InputArgumentNames()[0]);
+    if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
+      op_dev_id = GetAppropriateDeviceID(op.InputArgumentNames());
+      for (auto &varname : op.InputArgumentNames()) {
+        var_name_on_devices_.emplace(varname, op_dev_id);
+      }
+    }
+    for (auto &varname : op.OutputArgumentNames()) {
+      var_name_on_devices_.emplace(varname, op_dev_id);
+    }
+  } else if (op.Type() == "concat") {
+    op_dev_id = GetVarDeviceID(op.InputArgumentNames()[0]);
+  } else {
+    PADDLE_ENFORCE(
+        "the distribute training related op should be in [split_byref, "
+        "concat].");
+  }
+
+  PADDLE_ENFORCE(op_dev_id != -1,
+                 "can not find right place for distributed op: %s", op.Type());
+
+  CreateComputationalOp(result, op, op_dev_id);
   if (op.Type() == "concat") {
     ConnectOp(result, result->ops_.back().get(), "fetch_barrier");
   }
 }
 
+// Create RPC related op handles that connects its in ops and out ops.
 void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result,
                                           const OpDesc &op) const {
-  result->ops_.emplace_back(
-      new RPCOpHandle(op, local_scopes_[0], op.Type(), places_[0]));
+  int op_dev_id = -1;
+  if (op.Type() == "send") {
+    op_dev_id = GetVarDeviceID(op.InputArgumentNames()[0]);
+    // the variable name which contains .block means it was splited by
+    // split_byref op
+    // so that we can balance the variable blocks to all the pserver instances.
+    if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce &&
+        op.InputArgumentNames()[0].find(".block") == std::string::npos) {
+      op_dev_id = GetAppropriateDeviceID(op.InputArgumentNames());
+      for (auto &varname : op.InputArgumentNames()) {
+        var_name_on_devices_.emplace(varname, op_dev_id);
+      }
+    }
+  } else if (op.Type() == "recv") {
+    op_dev_id = GetAppropriateDeviceID(op.OutputArgumentNames());
+    for (auto &varname : op.OutputArgumentNames()) {
+      var_name_on_devices_.emplace(varname, op_dev_id);
+    }
+  } else {
+    // send_barrier and fetch_barrier op can be scheduled on device 0
+    op_dev_id = 0;
+  }
+
+  PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s",
+                 op.Type());
+
+  result->ops_.emplace_back(new RPCOpHandle(op, local_scopes_[op_dev_id],
+                                            op.Type(), places_[op_dev_id]));
 
   if (op.Type() == "send_barrier") {
     ConnectOp(result, result->ops_.back().get(), "send");
@@ -488,9 +545,7 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result,
         "send, send_barrier. recv, fetch_barrier]");
   }
 
-  // TODO(Yancey1989): schedule rpc op on different place may
-  // increate throughput
-  CreateOpHandleIOs(result, op, 0);
+  CreateOpHandleIOs(result, op, op_dev_id);
 }
 
 bool MultiDevSSAGraphBuilder::IsScaleLossOp(const OpDesc &op) const {

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -47,10 +47,11 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
+  int GetVarDeviceID(const std::string &varname) const;
 
  private:
   void CreateOpHandleIOs(SSAGraph *result, const OpDesc &op,
-                         size_t place_id) const;
+                         size_t device_id) const;
 
  private:
   std::string loss_var_name_;
@@ -96,21 +97,23 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
       const std::string &og,
       std::unordered_set<std::string> *og_has_been_broadcast) const;
 
-  int GetOpDeviceID(
-      const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
-      const OpDesc &op) const;
+  int GetOpDeviceID(const OpDesc &op) const;
 
   void InsertAllReduceOp(SSAGraph *result, const std::string &og) const;
 
   void CreateBroadcastOp(SSAGraph *result, const std::string &p_name,
                          size_t src_dev_id) const;
 
-  bool IsSparseGradient(
-      const std::unordered_map<std::string, VarDesc *> &all_vars,
-      const std::string &og) const;
+  bool IsSparseGradient(const std::string &og) const;
+
+  size_t GetAppropriateDeviceID(
+      const std::vector<std::string> &var_names) const;
 
  private:
   BuildStrategy strategy_;
+  mutable std::unordered_map<std::string, VarDesc *> all_vars_;
+  mutable std::unordered_map<std::string, int> var_name_on_devices_;
+  mutable std::vector<int64_t> balance_vars_;
 
   void SetCommunicationContext(OpHandleBase *op_handle,
                                const platform::Place &p) const;

paddle/fluid/framework/details/ssa_graph_builder.h

@@ -30,6 +30,7 @@ class SSAGraphBuilder {
   SSAGraphBuilder() {}
   virtual ~SSAGraphBuilder() {}
   virtual std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const = 0;
+  virtual int GetVarDeviceID(const std::string &var_name) const { return -1; }
 
   DISABLE_COPY_AND_ASSIGN(SSAGraphBuilder);
 

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -96,6 +96,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     auto cur_ready_vars = ready_vars.PopAll(1, &timeout);
 
     if (timeout) {
+      std::lock_guard<std::mutex> l(exception_mu_);
       if (exception_) {
         auto exp = *exception_;
         exception_.reset();
@@ -199,6 +200,7 @@ void ThreadedSSAGraphExecutor::RunOp(
       ready_var_q->Extend(op->Outputs());
       VLOG(10) << op << " " << op->Name() << "Signal posted";
     } catch (platform::EnforceNotMet ex) {
+      std::lock_guard<std::mutex> l(exception_mu_);
       exception_.reset(new platform::EnforceNotMet(ex));
     } catch (...) {
       LOG(FATAL) << "Unknown exception catched";

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -56,6 +56,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
   platform::DeviceContextPool fetch_ctxs_;
+  std::mutex exception_mu_;
   std::unique_ptr<platform::EnforceNotMet> exception_;
   std::atomic<int> running_ops_;
 

paddle/fluid/framework/executor.cc

@@ -21,7 +21,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/reader.h"
 #ifdef PADDLE_WITH_DISTRIBUTE
-#include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/operators/distributed/grpc_client.h"
 #endif
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -49,8 +49,8 @@ Executor::Executor(const platform::Place& place) : place_(place) {}
 
 #ifdef PADDLE_WITH_DISTRIBUTE
 void Executor::Complete() {
-  ::paddle::operators::detail::RPCClient::GetInstance<
-      ::paddle::operators::detail::GRPCClient>()
+  ::paddle::operators::distributed::RPCClient::GetInstance<
+      ::paddle::operators::distributed::GRPCClient>()
       ->SendComplete();
 }
 #endif
@@ -321,7 +321,8 @@ std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
 }
 
 void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
-                                  bool create_local_scope, bool create_vars) {
+                                  bool create_local_scope, bool create_vars,
+                                  bool keep_kids) {
   Scope* local_scope = scope;
   if (create_vars) {
     if (create_local_scope) {
@@ -344,12 +345,20 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
     }
   }
   platform::DeviceContextPool::Instance().Get(place_)->Wait();
-  if (create_vars && create_local_scope) {
+  if (local_scope != scope) {
     scope->DeleteScope(local_scope);
   } else {
-    // Delete the local scopes created in operators.
-    scope->DropKids();
+    if (!keep_kids) {
+      // By default, we should delete all kid scopes after run executor because
+      // some operators may create local scope when running, such as while_op.
+      // But when while_op also create a local executor to run it's sub block,
+      // the sub scopes it created should not be dropped immediately, because
+      // while_grad_op will use some variables created during while_op run, so
+      // we need to keep the kids and wait for the outer executor to drop them.
+      scope->DropKids();
+    }
   }
+
   if (FLAGS_benchmark) {
     VLOG(2) << "-------------------------------------------------------";
     VLOG(2) << "Memory used after deleting local scope: "

paddle/fluid/framework/executor.h

@@ -78,7 +78,7 @@ class Executor {
 
   void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                           bool create_local_scope = true,
-                          bool create_vars = true);
+                          bool create_vars = true, bool keep_kids = false);
 
   void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                           std::map<std::string, const LoDTensor*>* feed_targets,

paddle/fluid/framework/parallel_executor.cc

@@ -110,7 +110,6 @@ ParallelExecutor::ParallelExecutor(
 
   // Step 3. Convert main_program to SSA form and dependency graph. Also, insert
   // ncclOp
-
   details::SSAGraphBuilderFactory builder_factory(
       member_->places_, loss_var_name, params, member_->local_scopes_,
       build_strategy);
@@ -122,9 +121,10 @@ ParallelExecutor::ParallelExecutor(
 #endif
   }
 
+  builder_ = std::move(builder_factory.Create());
   member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
       exec_strategy, member_->local_scopes_, places,
-      builder_factory.Create()->Build(main_program)));
+      builder_->Build(main_program)));
 
   member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
       exec_strategy, member_->local_scopes_, std::move(var_infos),
@@ -133,10 +133,22 @@ ParallelExecutor::ParallelExecutor(
 
 void ParallelExecutor::BCastParamsToGPUs(
     const std::unordered_set<std::string> &vars) const {
-  auto *main_scope = member_->local_scopes_[0];
+  // the the initialize bcast, all vars would be bcast from device(0), otherwise
+  // bcast from the specified device.
+  bool initialize = builder_.get() == nullptr ? true : false;
 
   for (auto &var : vars) {
-    auto *main_var = main_scope->FindVar(var);
+    int var_dev_id =
+        builder_.get() == nullptr ? -1 : builder_->GetVarDeviceID(var);
+    if (!initialize && var_dev_id == -1) continue;
+
+    framework::Variable *main_var = nullptr;
+    if (initialize) {
+      main_var = member_->local_scopes_[0]->FindVar(var);
+    } else {
+      main_var = member_->local_scopes_[var_dev_id]->FindVar(var);
+    }
+
     if (main_var == nullptr || !main_var->IsType<LoDTensor>()) {
       continue;
     }
@@ -151,7 +163,8 @@ void ParallelExecutor::BCastParamsToGPUs(
       for (size_t i = 0; i < member_->places_.size(); ++i) {
         auto place = member_->places_[i];
         void *buffer;
-        if (i == 0) {
+
+        if ((initialize && i == 0) || (!initialize && i == var_dev_id)) {
           buffer = const_cast<void *>(main_tensor.data<void>());
         } else {
           auto local_scope = member_->local_scopes_[i];

paddle/fluid/framework/parallel_executor.h

@@ -19,12 +19,14 @@ limitations under the License. */
 #include <unordered_set>
 #include <vector>
 #include "paddle/fluid/framework/details/execution_strategy.h"
+#include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/platform/device_context.h"
+
 namespace paddle {
 namespace framework {
 
@@ -68,6 +70,7 @@ class ParallelExecutor {
 
  private:
   ParallelExecutorPrivate *member_;
+  std::unique_ptr<details::SSAGraphBuilder> builder_;
 };
 
 }  // namespace framework

paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc

@@ -27,7 +27,7 @@ void TensorRTSubGraphPass::Run(DataFlowGraph *graph) {
   SubGraphFuse(graph, node_inside_subgraph_teller_);
 }
 
-}  // analysis
-}  // inference
+}  // namespace analysis
+}  // namespace inference
 
-}  // paddle
+}  // namespace paddle

paddle/fluid/operators/CMakeLists.txt

@@ -184,9 +184,9 @@ else()
     set(DEPS_OPS ${DEPS_OPS} nccl_op)
 endif()
 
-add_subdirectory(detail)
 if(WITH_DISTRIBUTE)
-
+    add_subdirectory(distributed)
+    
     set(DISTRIBUTE_DEPS "")
     if(WITH_GRPC)
         set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf)
@@ -195,18 +195,11 @@ if(WITH_DISTRIBUTE)
     endif()
 
     set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
-    op_library(prefetch_op DEPS ${DISTRIBUTE_DEPS})
-    set_source_files_properties(prefetch_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-    op_library(recv_op DEPS ${DISTRIBUTE_DEPS})
-    set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-    op_library(listen_and_serv_op DEPS ${DISTRIBUTE_DEPS})
-    set_source_files_properties(listen_and_serv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-    op_library(send_op DEPS ${DISTRIBUTE_DEPS})
-    set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-    op_library(send_barrier_op DEPS ${DISTRIBUTE_DEPS})
-    op_library(fetch_barrier_op DEPS ${DISTRIBUTE_DEPS})
-    set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-    set_source_files_properties(fetch_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    foreach(dist_op "prefetch_op" "listen_and_serv_op" "send_op" "recv_op" "send_barrier_op" "fetch_barrier_op")
+        op_library(${dist_op} DEPS ${DISTRIBUTE_DEPS})
+        set_source_files_properties(${dist_op}.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    endforeach()
+    
     #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op
     #        listen_and_serv_op sum_op executor SERIAL)

paddle/fluid/operators/batch_norm_mkldnn_op.cc

@@ -21,8 +21,6 @@ namespace operators {
 
 using batch_norm_bwd = mkldnn::batch_normalization_backward;
 using batch_norm_fwd = mkldnn::batch_normalization_forward;
-using framework::DataLayout;
-using framework::Tensor;
 using mkldnn::memory;
 using mkldnn::primitive;
 using mkldnn::reorder;
@@ -31,18 +29,6 @@ using paddle::platform::MKLDNNDeviceContext;
 using paddle::platform::MKLDNNMemDesc;
 using platform::to_void_cast;
 
-template <typename T>
-using EigenArrayMap =
-    Eigen::Map<Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
-template <typename T>
-using ConstEigenArrayMap =
-    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
-template <typename T>
-using EigenVectorArrayMap = Eigen::Map<Eigen::Array<T, Eigen::Dynamic, 1>>;
-template <typename T>
-using ConstEigenVectorArrayMap =
-    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, 1>>;
-
 namespace {
 template <typename T>
 struct bn_type_traits {

paddle/fluid/operators/batch_norm_op.cc

@@ -22,22 +22,6 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-using Tensor = framework::Tensor;
-using LoDTensor = framework::LoDTensor;
-using DataLayout = framework::DataLayout;
-
-template <typename T>
-using EigenArrayMap =
-    Eigen::Map<Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
-template <typename T>
-using ConstEigenArrayMap =
-    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
-template <typename T>
-using EigenVectorArrayMap = Eigen::Map<Eigen::Array<T, Eigen::Dynamic, 1>>;
-template <typename T>
-using ConstEigenVectorArrayMap =
-    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, 1>>;
-
 class BatchNormOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;

paddle/fluid/operators/batch_norm_op.h

@@ -19,6 +19,22 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+using DataLayout = framework::DataLayout;
+
+template <typename T>
+using EigenArrayMap =
+    Eigen::Map<Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using ConstEigenArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using EigenVectorArrayMap = Eigen::Map<Eigen::Array<T, Eigen::Dynamic, 1>>;
+template <typename T>
+using ConstEigenVectorArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, 1>>;
+
 template <typename DeviceContext, typename T>
 class BatchNormKernel : public framework::OpKernel<T> {
  public:

paddle/fluid/operators/bilinear_interp_op.cc

@@ -110,6 +110,7 @@ REGISTER_OPERATOR(bilinear_interp, ops::BilinearInterpOp,
                   ops::BilinearInterpOpMaker,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(bilinear_interp_grad, ops::BilinearInterpOpGrad);
-REGISTER_OP_CPU_KERNEL(bilinear_interp, ops::BilinearInterpKernel<float>);
+REGISTER_OP_CPU_KERNEL(bilinear_interp, ops::BilinearInterpKernel<float>,
+                       ops::BilinearInterpKernel<uint8_t>);
 REGISTER_OP_CPU_KERNEL(bilinear_interp_grad,
                        ops::BilinearInterpGradKernel<float>);

paddle/fluid/operators/bilinear_interp_op.h

@@ -46,8 +46,10 @@ class BilinearInterpKernel : public framework::OpKernel<T> {
     int in_chw = channels * in_hw;
     int out_chw = channels * out_hw;
 
-    T ratio_h = (out_h > 1) ? static_cast<T>(in_h - 1) / (out_h - 1) : 0.f;
-    T ratio_w = (out_w > 1) ? static_cast<T>(in_w - 1) / (out_w - 1) : 0.f;
+    float ratio_h =
+        (out_h > 1) ? static_cast<float>(in_h - 1) / (out_h - 1) : 0.f;
+    float ratio_w =
+        (out_w > 1) ? static_cast<float>(in_w - 1) / (out_w - 1) : 0.f;
 
     if (in_h == out_h && in_w == out_w) {
       memcpy(output, input, input_t->numel() * sizeof(T));
@@ -56,24 +58,24 @@ class BilinearInterpKernel : public framework::OpKernel<T> {
         for (int i = 0; i < out_h; ++i) {     // loop for images
           int h = ratio_h * i;
           int hid = (h < in_h - 1) ? 1 : 0;
-          T h1lambda = ratio_h * i - h;
-          T h2lambda = 1 - h1lambda;
+          float h1lambda = ratio_h * i - h;
+          float h2lambda = 1.f - h1lambda;
 
           for (int j = 0; j < out_w; ++j) {
             int w = ratio_w * j;
             int wid = (w < in_w - 1) ? 1 : 0;
-            T w1lambda = ratio_w * j - w;
-            T w2lambda = 1 - w1lambda;
+            float w1lambda = ratio_w * j - w;
+            float w2lambda = 1.f - w1lambda;
             // calculate four position for bilinear interpolation
             const T* in_pos = &input[k * in_chw + h * in_w + w];
             T* out_pos = &output[k * out_chw + i * out_w + j];
 
             for (int c = 0; c < channels; ++c) {  // loop for channels
               // bilinear interpolation
-              out_pos[0] =
+              out_pos[0] = static_cast<T>(
                   h2lambda * (w2lambda * in_pos[0] + w1lambda * in_pos[wid]) +
                   h1lambda * (w2lambda * in_pos[hid * in_w] +
-                              w1lambda * in_pos[hid * in_w + wid]);
+                              w1lambda * in_pos[hid * in_w + wid]));
               in_pos += in_hw;
               out_pos += out_hw;
             }
@@ -117,8 +119,10 @@ class BilinearInterpGradKernel : public framework::OpKernel<T> {
     int in_chw = channels * in_hw;
     int out_chw = channels * out_hw;
 
-    T ratio_h = (out_h > 1) ? static_cast<T>(in_h - 1) / (out_h - 1) : 0.f;
-    T ratio_w = (out_w > 1) ? static_cast<T>(in_w - 1) / (out_w - 1) : 0.f;
+    float ratio_h =
+        (out_h > 1) ? static_cast<float>(in_h - 1) / (out_h - 1) : 0.f;
+    float ratio_w =
+        (out_w > 1) ? static_cast<float>(in_w - 1) / (out_w - 1) : 0.f;
 
     if (in_h == out_h && in_w == out_w) {
       memcpy(d_input, d_output, d_input_t->numel() * sizeof(T));
@@ -127,22 +131,24 @@ class BilinearInterpGradKernel : public framework::OpKernel<T> {
         for (int i = 0; i < out_h; ++i) {     // loop for images
           int h = ratio_h * i;
           int hid = (h < in_h - 1) ? 1 : 0;
-          T h1lambda = ratio_h * i - h;
-          T h2lambda = 1 - h1lambda;
+          float h1lambda = ratio_h * i - h;
+          float h2lambda = 1 - h1lambda;
 
           for (int j = 0; j < out_w; ++j) {
             int w = ratio_w * j;
             int wid = (w < in_w - 1) ? 1 : 0;
-            T w1lambda = ratio_w * j - w;
-            T w2lambda = 1 - w1lambda;
+            float w1lambda = ratio_w * j - w;
+            float w2lambda = 1 - w1lambda;
             T* in_pos = &d_input[k * in_chw + h * in_w + w];
             const T* out_pos = &d_output[k * out_chw + i * out_w + j];
 
             for (int c = 0; c < channels; ++c) {  // loop for channels
-              in_pos[0] += h2lambda * w2lambda * out_pos[0];
-              in_pos[wid] += h2lambda * w1lambda * out_pos[0];
-              in_pos[hid * in_w] += h1lambda * w2lambda * out_pos[0];
-              in_pos[hid * in_w + wid] += h1lambda * w1lambda * out_pos[0];
+              in_pos[0] += static_cast<T>(h2lambda * w2lambda * out_pos[0]);
+              in_pos[wid] += static_cast<T>(h2lambda * w1lambda * out_pos[0]);
+              in_pos[hid * in_w] +=
+                  static_cast<T>(h1lambda * w2lambda * out_pos[0]);
+              in_pos[hid * in_w + wid] +=
+                  static_cast<T>(h1lambda * w1lambda * out_pos[0]);
               in_pos += in_hw;
               out_pos += out_hw;
             }

paddle/fluid/operators/detail/macros.h

@@ -15,13 +15,13 @@
 #pragma once
 
 #ifdef PADDLE_WITH_GRPC
-#include "paddle/fluid/operators/detail/grpc_client.h"
-#include "paddle/fluid/operators/detail/grpc_server.h"
-#define RPCSERVER_T detail::AsyncGRPCServer
-#define RPCCLIENT_T detail::GRPCClient
+#include "paddle/fluid/operators/distributed/grpc_client.h"
+#include "paddle/fluid/operators/distributed/grpc_server.h"
+#define RPCSERVER_T distributed::AsyncGRPCServer
+#define RPCCLIENT_T distributed::GRPCClient
 #else
-#include "paddle/fluid/operators/detail/brpc_client.h"
-#include "paddle/fluid/operators/detail/brpc_server.h"
-#define RPCSERVER_T detail::AsyncBRPCServer
-#define RPCCLIENT_T detail::BRPCClient
+#include "paddle/fluid/operators/distributed/brpc_client.h"
+#include "paddle/fluid/operators/distributed/brpc_server.h"
+#define RPCSERVER_T distributed::AsyncBRPCServer
+#define RPCCLIENT_T distributed::BRPCClient
 #endif

paddle/fluid/operators/detail/CMakeLists.txt → paddle/fluid/operators/distributed/CMakeLists.txt

-if(NOT WITH_DISTRIBUTE)
-    return()
-endif()
-
-
 if(WITH_GRPC)
   grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc
       request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor

paddle/fluid/operators/detail/brpc_client.cc → paddle/fluid/operators/distributed/brpc_client.cc

@@ -12,12 +12,12 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/operators/detail/brpc_client.h"
+#include "paddle/fluid/operators/distributed/brpc_client.h"
 #include "paddle/fluid/framework/threadpool.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 DEFINE_int32(brpc_channel_num, 24,
              "Number of channels to send requests connected to one server");
@@ -175,6 +175,6 @@ ChannelQueuePtr BRPCClient::GetChannel(const std::string& ep) {
   return q;
 }
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/brpc_client.h → paddle/fluid/operators/distributed/brpc_client.h

@@ -31,13 +31,13 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
-#include "paddle/fluid/operators/detail/rpc_client.h"
-#include "paddle/fluid/operators/detail/send_recv.pb.h"
+#include "paddle/fluid/operators/distributed/rpc_client.h"
+#include "paddle/fluid/operators/distributed/send_recv.pb.h"
 #include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 struct ChannelContext {
   brpc::Channel channel;
@@ -95,6 +95,6 @@ class BRPCClient : public RPCClient {
   DISABLE_COPY_AND_ASSIGN(BRPCClient);
 };
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/brpc_server.cc → paddle/fluid/operators/distributed/brpc_server.cc

@@ -12,13 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/operators/detail/brpc_server.h"
-#include "paddle/fluid/operators/detail/request_handler.h"
+#include "paddle/fluid/operators/distributed/brpc_server.h"
+#include "paddle/fluid/operators/distributed/request_handler.h"
 
 namespace sendrecv {
 
 typedef std::unordered_map<std::string,
-                           paddle::operators::detail::RequestHandler*>
+                           paddle::operators::distributed::RequestHandler*>
     HandlerMap;
 
 class BRPCServiceImpl : public SendRecvService {
@@ -27,17 +27,17 @@ class BRPCServiceImpl : public SendRecvService {
       : request_send_h_(nullptr),
         request_get_h_(nullptr),
         request_prefetch_h_(nullptr) {
-    auto it = rpc_call_map.find(paddle::operators::detail::kRequestSend);
+    auto it = rpc_call_map.find(paddle::operators::distributed::kRequestSend);
     if (it != rpc_call_map.end()) {
       request_send_h_ = it->second;
     }
 
-    it = rpc_call_map.find(paddle::operators::detail::kRequestSend);
+    it = rpc_call_map.find(paddle::operators::distributed::kRequestSend);
     if (it != rpc_call_map.end()) {
       request_get_h_ = it->second;
     }
 
-    it = rpc_call_map.find(paddle::operators::detail::kRequestPrefetch);
+    it = rpc_call_map.find(paddle::operators::distributed::kRequestPrefetch);
     if (it != rpc_call_map.end()) {
       request_prefetch_h_ = it->second;
     }
@@ -88,15 +88,15 @@ class BRPCServiceImpl : public SendRecvService {
   }
 
  private:
-  paddle::operators::detail::RequestHandler* request_send_h_;
-  paddle::operators::detail::RequestHandler* request_get_h_;
-  paddle::operators::detail::RequestHandler* request_prefetch_h_;
+  paddle::operators::distributed::RequestHandler* request_send_h_;
+  paddle::operators::distributed::RequestHandler* request_get_h_;
+  paddle::operators::distributed::RequestHandler* request_prefetch_h_;
 };
 }  // namespace sendrecv
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 void AsyncBRPCServer::StartServer() {
   // Instance of your service.
@@ -139,6 +139,6 @@ void AsyncBRPCServer::WaitServerReady() {
   VLOG(3) << "AsyncGRPCServer WaitSeverReady";
 }
 
-};  // namespace detail
+};  // namespace distributed
 };  // namespace operators
 };  // namespace paddle

paddle/fluid/operators/detail/brpc_server.h → paddle/fluid/operators/distributed/brpc_server.h

@@ -19,12 +19,12 @@ limitations under the License. */
 #include <string>
 
 #include "brpc/server.h"
-#include "paddle/fluid/operators/detail/rpc_server.h"
-#include "paddle/fluid/operators/detail/send_recv.pb.h"
+#include "paddle/fluid/operators/distributed/rpc_server.h"
+#include "paddle/fluid/operators/distributed/send_recv.pb.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 class AsyncBRPCServer final : public RPCServer {
  public:
@@ -48,6 +48,6 @@ class AsyncBRPCServer final : public RPCServer {
   int ready_;
 };
 
-};  // namespace detail
+};  // namespace distributed
 };  // namespace operators
 };  // namespace paddle

paddle/fluid/operators/detail/bytebuffer_stream.cc → paddle/fluid/operators/distributed/bytebuffer_stream.cc

@@ -17,11 +17,11 @@ limitations under the License. */
 //       file and did some modifications so that we can send gRPC
 //       requests without too much copying of the tensor data.
 
-#include "paddle/fluid/operators/detail/bytebuffer_stream.h"
+#include "paddle/fluid/operators/distributed/bytebuffer_stream.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 GrpcByteBufferSource::GrpcByteBufferSource() {}
 
@@ -83,6 +83,6 @@ google::protobuf::int64 GrpcByteBufferSource::ByteCount() const {
   return byte_count_;
 }
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/bytebuffer_stream.h → paddle/fluid/operators/distributed/bytebuffer_stream.h

@@ -106,7 +106,7 @@ class GrpcBufferReader final
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 // Source provides a way for a particular RPC implementation to provide
 // received data to ParseFrom.
 class Source {
@@ -183,6 +183,6 @@ class GrpcByteSource : public Source {
   char space_[sizeof(Reader)];
 };
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/grpc_client.cc → paddle/fluid/operators/distributed/grpc_client.cc

@@ -12,19 +12,19 @@ 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 "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/operators/distributed/grpc_client.h"
 
 #include <sys/time.h>
 
 #include <limits>
 
 #include "paddle/fluid/framework/threadpool.h"
-#include "paddle/fluid/operators/detail/request_handler.h"
+#include "paddle/fluid/operators/distributed/request_handler.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 void GRPCClient::InitImpl() { InitEventLoop(); }
 
@@ -276,6 +276,6 @@ std::shared_ptr<grpc::Channel> GRPCClient::GetChannel(const std::string& ep) {
   return ch;
 }
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/grpc_client.h → paddle/fluid/operators/distributed/grpc_client.h

@@ -38,13 +38,13 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
-#include "paddle/fluid/operators/detail/rpc_client.h"
-#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+#include "paddle/fluid/operators/distributed/rpc_client.h"
+#include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
 #include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 struct VarHandle {
   std::string ep;
@@ -226,6 +226,6 @@ class GRPCClient : public RPCClient {
   DISABLE_COPY_AND_ASSIGN(GRPCClient);
 };
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/grpc_serde_test.cc → paddle/fluid/operators/distributed/grpc_serde_test.cc

@@ -21,8 +21,8 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/framework/variable.h"
-#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
-#include "paddle/fluid/operators/detail/variable_response.h"
+#include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
+#include "paddle/fluid/operators/distributed/variable_response.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/string/printf.h"
@@ -50,7 +50,7 @@ void RunSerdeTestSelectedRows(platform::Place place) {
   for (int i = 0; i < 564; ++i) rows->push_back(i);
 
   ::grpc::ByteBuffer msg;
-  operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg);
+  operators::distributed::SerializeToByteBuffer("myvar", &var, ctx, &msg);
   EXPECT_GT(msg.Length(), static_cast<size_t>(0));
 
   // deserialize
@@ -81,10 +81,10 @@ void RunSerdeTestSelectedRows(platform::Place place) {
 
   // deserialize zero-copy
   // framework::Variable var2;
-  // operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2);
+  // operators::distributed::DeserializeFromByteBuffer(msg, ctx, &var2);
   framework::Scope scope;
   scope.Var("myvar");
-  operators::detail::VariableResponse resp(&scope, &ctx);
+  operators::distributed::VariableResponse resp(&scope, &ctx);
   EXPECT_EQ(resp.Parse(msg), 0);
 
   framework::Variable* var2 = resp.GetVar();
@@ -128,7 +128,7 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) {
   math::set_constant(ctx, tensor, 31.9);
 
   ::grpc::ByteBuffer msg;
-  operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg);
+  operators::distributed::SerializeToByteBuffer("myvar", &var, ctx, &msg);
   EXPECT_GT(msg.Length(), static_cast<size_t>(0));
 
   // deserialize
@@ -171,7 +171,7 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) {
   // deserialize zero-copy
   framework::Scope scope;
   scope.Var("myvar");
-  operators::detail::VariableResponse resp(&scope, &ctx);
+  operators::distributed::VariableResponse resp(&scope, &ctx);
   if (from_type == 0) {
     EXPECT_EQ(resp.Parse(msg), 0);
   } else {

paddle/fluid/operators/detail/grpc_server.cc → paddle/fluid/operators/distributed/grpc_server.cc

@@ -15,13 +15,13 @@ limitations under the License. */
 #include <limits>
 #include <string>
 
-#include "paddle/fluid/operators/detail/grpc_server.h"
+#include "paddle/fluid/operators/distributed/grpc_server.h"
 
 using ::grpc::ServerAsyncResponseWriter;
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 enum CallStatus { PROCESS = 0, FINISH };
 
 // reference:
@@ -74,7 +74,7 @@ class RequestSend final : public RequestBase {
     request_.reset(new VariableResponse(request_handler->scope(),
                                         request_handler->dev_ctx(),
                                         !request_handler->sync_mode()));
-    int method_id = static_cast<int>(detail::GrpcMethod::kSendVariable);
+    int method_id = static_cast<int>(distributed::GrpcMethod::kSendVariable);
     service_->RequestAsyncUnary(
         method_id, &ctx_, request_.get(), &responder_, cq_, cq_,
         reinterpret_cast<void*>(static_cast<intptr_t>(req_id)));
@@ -106,7 +106,7 @@ class RequestGet final : public RequestBase {
                       ::grpc::ServerCompletionQueue* cq,
                       RequestHandler* request_handler, int req_id)
       : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) {
-    auto method_id = static_cast<int>(detail::GrpcMethod::kGetVariable);
+    auto method_id = static_cast<int>(distributed::GrpcMethod::kGetVariable);
     service_->RequestAsyncUnary(
         method_id, &ctx_, &request_, &responder_, cq_, cq_,
         reinterpret_cast<void*>(static_cast<intptr_t>(req_id)));
@@ -150,7 +150,8 @@ class RequestPrefetch final : public RequestBase {
         local_scope_(nullptr) {
     request_.reset(new VariableResponse(request_handler->scope(),
                                         request_handler->dev_ctx(), true));
-    int method_id = static_cast<int>(detail::GrpcMethod::kPrefetchVariable);
+    int method_id =
+        static_cast<int>(distributed::GrpcMethod::kPrefetchVariable);
     service_->RequestAsyncUnary(
         method_id, &ctx_, request_.get(), &responder_, cq_, cq_,
         reinterpret_cast<void*>(static_cast<intptr_t>(req_id)));
@@ -354,6 +355,6 @@ void AsyncGRPCServer::HandleRequest(
   }
 }
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/grpc_server.h → paddle/fluid/operators/distributed/grpc_server.h

@@ -29,17 +29,17 @@ limitations under the License. */
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/var_type.h"
-#include "paddle/fluid/operators/detail/grpc_service.h"
-#include "paddle/fluid/operators/detail/request_handler.h"
-#include "paddle/fluid/operators/detail/rpc_server.h"
-#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h"
-#include "paddle/fluid/operators/detail/send_recv.pb.h"
-#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+#include "paddle/fluid/operators/distributed/grpc_service.h"
+#include "paddle/fluid/operators/distributed/request_handler.h"
+#include "paddle/fluid/operators/distributed/rpc_server.h"
+#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h"
+#include "paddle/fluid/operators/distributed/send_recv.pb.h"
+#include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 class RequestBase;
 
@@ -84,6 +84,6 @@ class AsyncGRPCServer final : public RPCServer {
   std::map<std::string, std::vector<RequestBase*>> rpc_reqs_;
 };
 
-};  // namespace detail
+};  // namespace distributed
 };  // namespace operators
 };  // namespace paddle

paddle/fluid/operators/detail/grpc_service.h → paddle/fluid/operators/distributed/grpc_service.h

@@ -23,7 +23,7 @@
 #include <grpc++/impl/codegen/stub_options.h>
 #include <grpc++/impl/codegen/sync_stream.h>
 #include <grpc++/support/byte_buffer.h>
-#include "paddle/fluid/operators/detail/variable_response.h"
+#include "paddle/fluid/operators/distributed/variable_response.h"
 
 #include "paddle/fluid/platform/profiler.h"
 
@@ -42,24 +42,25 @@ class ServerContext;
 // Support parsing/unparsing of tensorflow::VariableResponse.
 // Wire-format is identical to RecvVariableResponse.
 template <>
-class SerializationTraits<paddle::operators::detail::VariableResponse> {
+class SerializationTraits<paddle::operators::distributed::VariableResponse> {
  public:
   static Status Serialize(
-      const paddle::operators::detail::VariableResponse& msg,
+      const paddle::operators::distributed::VariableResponse& msg,
       grpc_byte_buffer** bp, bool* own_buffer) {
     PADDLE_ENFORCE(false, "SerializationTraits::Serialize not implemented!");
     return Status();
   }
-  static Status Deserialize(grpc_byte_buffer* buffer,
-                            paddle::operators::detail::VariableResponse* msg,
-                            int max_message_size = INT_MAX) {
+  static Status Deserialize(
+      grpc_byte_buffer* buffer,
+      paddle::operators::distributed::VariableResponse* msg,
+      int max_message_size = INT_MAX) {
     if (buffer == nullptr) {
       return Status(StatusCode::INTERNAL, "No payload");
     }
 
     Status result = g_core_codegen_interface->ok();
     if (result.ok()) {
-      paddle::operators::detail::GrpcByteSource source(buffer);
+      paddle::operators::distributed::GrpcByteSource source(buffer);
       int ret = msg->Parse(&source);
       if (ret != 0) {
         result = Status(StatusCode::INTERNAL, "VariableResponse parse error");
@@ -73,7 +74,7 @@ class SerializationTraits<paddle::operators::detail::VariableResponse> {
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 enum class GrpcMethod {
   kSendVariable,
@@ -118,6 +119,6 @@ class GrpcService final {
   };
 };
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/proto_encoder_helper.h → paddle/fluid/operators/distributed/proto_encoder_helper.h

@@ -26,7 +26,7 @@ limitations under the License. */
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 char* EncodeVarint32(char* dst, uint32_t v) {
   // Operate on characters as unsigneds
@@ -144,6 +144,6 @@ class ProtoEncodeHelper {
   char* limit_;  // Just for CHECKs
 };
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/request_handler.h → paddle/fluid/operators/distributed/request_handler.h

@@ -31,7 +31,7 @@
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 constexpr char kRequestSend[] = "RequestSend";
 constexpr char kRequestGet[] = "RequestGet";
@@ -124,6 +124,6 @@ class RequestHandler {
   RPCServer* rpc_server_;
 };
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/request_handler_impl.cc → paddle/fluid/operators/distributed/request_handler_impl.cc

@@ -20,12 +20,12 @@
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
-#include "paddle/fluid/operators/detail/request_handler_impl.h"
-#include "paddle/fluid/operators/detail/rpc_server.h"
+#include "paddle/fluid/operators/distributed/request_handler_impl.h"
+#include "paddle/fluid/operators/distributed/rpc_server.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 bool RequestSendHandler::Handle(const std::string& varname,
                                 framework::Scope* scope,
@@ -119,6 +119,6 @@ bool RequestPrefetchHandler::Handle(const std::string& varname,
   return true;
 }
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/request_handler_impl.h → paddle/fluid/operators/distributed/request_handler_impl.h

@@ -28,11 +28,11 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/var_type.h"
-#include "paddle/fluid/operators/detail/request_handler.h"
+#include "paddle/fluid/operators/distributed/request_handler.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 class RequestSendHandler final : public RequestHandler {
  public:
@@ -66,6 +66,6 @@ class RequestPrefetchHandler final : public RequestHandler {
               const std::string& out_var_name = "") override;
 };
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/rpc_client.cc → paddle/fluid/operators/distributed/rpc_client.cc

@@ -12,15 +12,15 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/operators/detail/rpc_client.h"
+#include "paddle/fluid/operators/distributed/rpc_client.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 std::once_flag RPCClient::init_flag_;
 std::unique_ptr<RPCClient> RPCClient::rpc_client_(nullptr);
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/rpc_client.h → paddle/fluid/operators/distributed/rpc_client.h

@@ -22,7 +22,7 @@
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 class RPCClient {
  public:
@@ -84,6 +84,6 @@ class RPCClient {
   static std::once_flag init_flag_;
   static std::unique_ptr<RPCClient> rpc_client_;
 };
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/rpc_server.cc → paddle/fluid/operators/distributed/rpc_server.cc

@@ -17,11 +17,11 @@
 #include <limits>
 #include <string>
 
-#include "paddle/fluid/operators/detail/rpc_server.h"
+#include "paddle/fluid/operators/distributed/rpc_server.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 void RPCServer::ShutDown() {
   LOG(INFO) << "RPCServer ShutDown ";
@@ -112,6 +112,6 @@ void RPCServer::WaitCond(const std::string& rpc_name) {
       lock, [=] { return (cur_cond_.load() == cond || exit_flag_.load()); });
 }
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/rpc_server.h → paddle/fluid/operators/distributed/rpc_server.h

@@ -19,11 +19,11 @@
 #include <thread>  // NOLINT
 #include <utility>
 #include <vector>
-#include "paddle/fluid/operators/detail/request_handler.h"
+#include "paddle/fluid/operators/distributed/request_handler.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 class RPCServer {
  public:
@@ -86,6 +86,6 @@ class RPCServer {
   friend class RequestHandler;
 };
 
-};  // namespace detail
+};  // namespace distributed
 };  // namespace operators
 };  // namespace paddle

paddle/fluid/operators/detail/rpc_server_test.cc → paddle/fluid/operators/distributed/rpc_server_test.cc

@@ -22,18 +22,18 @@ limitations under the License. */
 #include "paddle/fluid/framework/operator.h"
 
 #include "paddle/fluid/operators/detail/macros.h"
-#include "paddle/fluid/operators/detail/request_handler_impl.h"
-#include "paddle/fluid/operators/detail/rpc_client.h"
-#include "paddle/fluid/operators/detail/rpc_server.h"
+#include "paddle/fluid/operators/distributed/request_handler_impl.h"
+#include "paddle/fluid/operators/distributed/rpc_client.h"
+#include "paddle/fluid/operators/distributed/rpc_server.h"
 
 namespace framework = paddle::framework;
 namespace platform = paddle::platform;
-namespace detail = paddle::operators::detail;
+namespace distributed = paddle::operators::distributed;
 
 USE_OP(lookup_table);
 
-std::unique_ptr<detail::RPCServer> g_rpc_service;
-std::unique_ptr<detail::RequestHandler> g_req_handler;
+std::unique_ptr<distributed::RPCServer> g_rpc_service;
+std::unique_ptr<distributed::RequestHandler> g_req_handler;
 
 framework::BlockDesc* AppendPrefetchBlcok(framework::ProgramDesc* program) {
   auto root_block = program->MutableBlock(0);
@@ -113,19 +113,21 @@ void StartServer() {
   g_req_handler->SetScope(&scope);
   g_req_handler->SetExecutor(&exe);
 
-  g_rpc_service->RegisterRPC(detail::kRequestPrefetch, g_req_handler.get());
+  g_rpc_service->RegisterRPC(distributed::kRequestPrefetch,
+                             g_req_handler.get());
   g_req_handler->SetRPCServer(g_rpc_service.get());
 
   std::thread server_thread(
-      std::bind(&detail::RPCServer::StartServer, g_rpc_service.get()));
+      std::bind(&distributed::RPCServer::StartServer, g_rpc_service.get()));
 
   server_thread.join();
 }
 
 TEST(PREFETCH, CPU) {
-  g_req_handler.reset(new detail::RequestPrefetchHandler(true));
+  g_req_handler.reset(new distributed::RequestPrefetchHandler(true));
   g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1));
-  detail::RPCClient* client = detail::RPCClient::GetInstance<RPCCLIENT_T>();
+  distributed::RPCClient* client =
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
   std::thread server_thread(StartServer);
   g_rpc_service->WaitServerReady();

paddle/fluid/operators/detail/sendrecvop_utils.cc → paddle/fluid/operators/distributed/sendrecvop_utils.cc

@@ -12,7 +12,7 @@ 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. */
 
-#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+#include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
 
 #ifdef PADDLE_WITH_CUDA
 #include <nccl.h>
@@ -23,14 +23,14 @@ limitations under the License. */
 #include "google/protobuf/io/coded_stream.h"
 #include "google/protobuf/io/zero_copy_stream.h"
 #include "paddle/fluid/framework/data_type.h"
-#include "paddle/fluid/operators/detail/bytebuffer_stream.h"
-#include "paddle/fluid/operators/detail/proto_encoder_helper.h"
-#include "paddle/fluid/operators/detail/variable_response.h"
+#include "paddle/fluid/operators/distributed/bytebuffer_stream.h"
+#include "paddle/fluid/operators/distributed/proto_encoder_helper.h"
+#include "paddle/fluid/operators/distributed/variable_response.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 using VarMsg = sendrecv::VariableMessage;
 
@@ -222,11 +222,11 @@ void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg,
                                const platform::DeviceContext& ctx,
                                const framework::Scope* scope,
                                framework::Variable** var) {
-  operators::detail::VariableResponse resp(scope, &ctx);
+  operators::distributed::VariableResponse resp(scope, &ctx);
   PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!");
   *var = resp.GetVar();
 }
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/sendrecvop_utils.h → paddle/fluid/operators/distributed/sendrecvop_utils.h

@@ -25,12 +25,12 @@ limitations under the License. */
 #include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/framework/var_type.h"
 
-#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h"
-#include "paddle/fluid/operators/detail/send_recv.pb.h"
+#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h"
+#include "paddle/fluid/operators/distributed/send_recv.pb.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 typedef void (*DestroyCallback)(void*);
 
@@ -61,6 +61,6 @@ inline std::type_index ToTypeIndex(sendrecv::VariableMessage::Type type) {
   }
 }
 
-}  // namespace detail
+}  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/variable_response.cc → paddle/fluid/operators/distributed/variable_response.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/operators/detail/variable_response.h"
+#include "paddle/fluid/operators/distributed/variable_response.h"
 
 #include <string>
 #include <utility>
@@ -22,12 +22,12 @@
 #endif
 #include "paddle/fluid/platform/profiler.h"
 
-#include "paddle/fluid/operators/detail/send_recv.pb.h"
-#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+#include "paddle/fluid/operators/distributed/send_recv.pb.h"
+#include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 enum WireType {
   WIRETYPE_VARINT = 0,
@@ -158,13 +158,13 @@ bool VariableResponse::CopySelectRowsTensorData(
   slr->set_height(meta_.slr_height());
   auto* tensor = slr->mutable_value();
   tensor->Resize(dims);
-  PADDLE_ENFORCE_EQ(
-      static_cast<size_t>(tensor->numel()),
-      length / framework::SizeOfType(
-                   paddle::operators::detail::ToTypeIndex(meta_.data_type())));
+  PADDLE_ENFORCE_EQ(static_cast<size_t>(tensor->numel()),
+                    length / framework::SizeOfType(
+                                 paddle::operators::distributed::ToTypeIndex(
+                                     meta_.data_type())));
   void* tensor_data = tensor->mutable_data(
       ctx.GetPlace(),
-      paddle::operators::detail::ToTypeIndex(meta_.data_type()));
+      paddle::operators::distributed::ToTypeIndex(meta_.data_type()));
 
   if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) {
     return false;
@@ -480,6 +480,6 @@ int VariableResponse::Parse(Source* source) {
   return 0;
 }
 
-};  // namespace detail
+};  // namespace distributed
 };  // namespace operators
 };  // namespace paddle

paddle/fluid/operators/detail/variable_response.h → paddle/fluid/operators/distributed/variable_response.h

@@ -22,17 +22,17 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/var_type.h"
 
-#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h"
-#include "paddle/fluid/operators/detail/send_recv.pb.h"
+#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h"
+#include "paddle/fluid/operators/distributed/send_recv.pb.h"
 
 #include "google/protobuf/io/coded_stream.h"
 #include "google/protobuf/io/zero_copy_stream.h"
 #include "paddle/fluid/framework/tensor.h"
-#include "paddle/fluid/operators/detail/bytebuffer_stream.h"
+#include "paddle/fluid/operators/distributed/bytebuffer_stream.h"
 
 namespace paddle {
 namespace operators {
-namespace detail {
+namespace distributed {
 
 class VariableResponse {
  public:
@@ -99,6 +99,6 @@ class VariableResponse {
   sendrecv::VariableMessage meta_;
 };
 
-};  // namespace detail
+};  // namespace distributed
 };  // namespace operators
 };  // namespace paddle

paddle/fluid/operators/fetch_barrier_op.cc

@@ -42,8 +42,8 @@ class FetchBarrierOp : public framework::OperatorBase {
     // For profiling
     platform::RecordEvent record_event(Type(), &ctx);
 
-    detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<RPCCLIENT_T>();
+    distributed::RPCClient* rpc_client =
+        distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
     rpc_client->Wait();
 

paddle/fluid/operators/gen_nccl_id_op.cc

@@ -22,7 +22,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/operators/detail/macros.h"
-#include "paddle/fluid/operators/detail/request_handler_impl.h"
+#include "paddle/fluid/operators/distributed/request_handler_impl.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 
 namespace paddle {
@@ -60,7 +60,8 @@ class GenNCCLIdOp : public framework::OperatorBase {
 
     std::vector<std::string> endpoint_list =
         Attr<std::vector<std::string>>("endpoint_list");
-    detail::RPCClient* client = detail::RPCClient::GetInstance<RPCCLIENT_T>();
+    distributed::RPCClient* client =
+        distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
     for (auto& ep : endpoint_list) {
       VLOG(3) << "sending nccl id to " << ep;
@@ -80,11 +81,11 @@ class GenNCCLIdOp : public framework::OperatorBase {
     // NOTE: Can not use unique_ptr here because the default
     // deleter will call GRPC Server's base class's dtor and
     // that will cause a wired crash.
-    detail::RequestSendHandler rpc_h(true);
-    std::unique_ptr<detail::RPCServer> rpc_service(
+    distributed::RequestSendHandler rpc_h(true);
+    std::unique_ptr<distributed::RPCServer> rpc_service(
         new RPCSERVER_T(endpoint, 1));
 
-    rpc_service->RegisterRPC(detail::kRequestSend, &rpc_h);
+    rpc_service->RegisterRPC(distributed::kRequestSend, &rpc_h);
     rpc_h.SetRPCServer(rpc_service.get());
 
     framework::ProgramDesc empty_program;
@@ -95,11 +96,11 @@ class GenNCCLIdOp : public framework::OperatorBase {
     rpc_h.SetExecutor(&executor);
 
     std::thread server_thread(
-        std::bind(&detail::RPCServer::StartServer, rpc_service.get()));
+        std::bind(&distributed::RPCServer::StartServer, rpc_service.get()));
 
-    rpc_service->SetCond(detail::kRequestSend);
+    rpc_service->SetCond(distributed::kRequestSend);
     VLOG(3) << "start getting nccl id from trainer 0...";
-    rpc_service->WaitBarrier(detail::kRequestSend);
+    rpc_service->WaitBarrier(distributed::kRequestSend);
     VLOG(3) << "got nccl id and stop server...";
     rpc_service->ShutDown();
     VLOG(3) << "rpc server stopped";

paddle/fluid/operators/listen_and_serv_op.cc

@@ -21,14 +21,14 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/detail/macros.h"
 
-#include "paddle/fluid/operators/detail/request_handler_impl.h"
+#include "paddle/fluid/operators/distributed/request_handler_impl.h"
 #include "paddle/fluid/operators/listen_and_serv_op.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
 
-void RunServer(std::shared_ptr<detail::RPCServer> service) {
+void RunServer(std::shared_ptr<distributed::RPCServer> service) {
   service->StartServer();
   VLOG(4) << "RunServer thread end";
 }
@@ -121,12 +121,12 @@ void ListenAndServOp::RunSyncLoop(
   while (true) {
     // Get from multiple trainers, we don't care about the order in which
     // the gradients arrives, just add suffix 0~n and merge the gradient.
-    rpc_service_->SetCond(detail::kRequestSend);
-    rpc_service_->WaitBarrier(detail::kRequestSend);
+    rpc_service_->SetCond(distributed::kRequestSend);
+    rpc_service_->WaitBarrier(distributed::kRequestSend);
 
     if (rpc_service_->IsExit()) {
       LOG(WARNING) << "get exit!rpc_processor break!";
-      rpc_service_->SetCond(detail::kRequestGet);
+      rpc_service_->SetCond(distributed::kRequestGet);
       break;
     }
 
@@ -154,11 +154,11 @@ void ListenAndServOp::RunSyncLoop(
                           recv_scope);
     VLOG(2) << "run all blocks spent " << GetTimestamp() - ts << "(ms)";
 
-    rpc_service_->SetCond(detail::kRequestGet);
-    rpc_service_->WaitBarrier(detail::kRequestGet);
+    rpc_service_->SetCond(distributed::kRequestGet);
+    rpc_service_->WaitBarrier(distributed::kRequestGet);
     rpc_service_->ResetBarrierCounter();
     // reset received sparse vars to avoid reuse it in the next mini-batch
-    dynamic_cast<detail::RequestSendHandler *>(request_send_handler_.get())
+    dynamic_cast<distributed::RequestSendHandler *>(request_send_handler_.get())
         ->ResetSparseVarRecorder();
   }  // while(true)
 }
@@ -215,13 +215,13 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor,
 }
 
 static void FillRequestCtx(
-    detail::RequestHandler *h, framework::Scope *scope,
+    distributed::RequestHandler *h, framework::Scope *scope,
     platform::DeviceContext *dev_ctx, framework::Executor *executor,
     framework::ProgramDesc *program,
     std::unordered_map<std::string,
                        std::shared_ptr<framework::ExecutorPrepareContext>>
         *prefetch_ctx,
-    detail::RPCServer *rpc_server) {
+    distributed::RPCServer *rpc_server) {
   h->SetScope(scope);
   h->SetDevCtx(dev_ctx);
   h->SetExecutor(executor);
@@ -249,14 +249,16 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
 
   rpc_service_.reset(new RPCSERVER_T(endpoint, fan_in));
 
-  request_send_handler_.reset(new detail::RequestSendHandler(sync_mode));
-  request_get_handler_.reset(new detail::RequestGetHandler(sync_mode));
+  request_send_handler_.reset(new distributed::RequestSendHandler(sync_mode));
+  request_get_handler_.reset(new distributed::RequestGetHandler(sync_mode));
   request_prefetch_handler_.reset(
-      new detail::RequestPrefetchHandler(sync_mode));
+      new distributed::RequestPrefetchHandler(sync_mode));
 
-  rpc_service_->RegisterRPC(detail::kRequestSend, request_send_handler_.get());
-  rpc_service_->RegisterRPC(detail::kRequestGet, request_get_handler_.get());
-  rpc_service_->RegisterRPC(detail::kRequestPrefetch,
+  rpc_service_->RegisterRPC(distributed::kRequestSend,
+                            request_send_handler_.get());
+  rpc_service_->RegisterRPC(distributed::kRequestGet,
+                            request_get_handler_.get());
+  rpc_service_->RegisterRPC(distributed::kRequestPrefetch,
                             request_prefetch_handler_.get());
 
   auto *optimize_block = Attr<framework::BlockDesc *>(kOptimizeBlock);

paddle/fluid/operators/listen_and_serv_op.h

@@ -24,8 +24,8 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/threadpool.h"
-#include "paddle/fluid/operators/detail/request_handler.h"
-#include "paddle/fluid/operators/detail/rpc_server.h"
+#include "paddle/fluid/operators/distributed/request_handler.h"
+#include "paddle/fluid/operators/distributed/rpc_server.h"
 
 namespace paddle {
 namespace operators {
@@ -33,7 +33,7 @@ namespace operators {
 constexpr char kOptimizeBlock[] = "OptimizeBlock";
 constexpr char kPrefetchVarNameToBlockId[] = "prefetch_var_name_to_block_id";
 
-void RunServer(std::shared_ptr<detail::RPCServer> service);
+void RunServer(std::shared_ptr<distributed::RPCServer> service);
 
 class ListenAndServOp : public framework::OperatorBase {
  public:
@@ -62,10 +62,11 @@ class ListenAndServOp : public framework::OperatorBase {
                const platform::Place& dev_place) const override;
 
  protected:
-  mutable std::shared_ptr<detail::RPCServer> rpc_service_;
-  mutable std::shared_ptr<detail::RequestHandler> request_send_handler_;
-  mutable std::shared_ptr<detail::RequestHandler> request_get_handler_;
-  mutable std::shared_ptr<detail::RequestHandler> request_prefetch_handler_;
+  mutable std::shared_ptr<distributed::RPCServer> rpc_service_;
+  mutable std::shared_ptr<distributed::RequestHandler> request_send_handler_;
+  mutable std::shared_ptr<distributed::RequestHandler> request_get_handler_;
+  mutable std::shared_ptr<distributed::RequestHandler>
+      request_prefetch_handler_;
 
   mutable std::shared_ptr<std::thread> server_thread_;
 };

paddle/fluid/operators/logical_op.cc

@@ -146,6 +146,6 @@ REGISTER_UNARY_LOGICAL_OP(logical_not, "$$Out = !X$$");
 REGISTER_UNARY_LOGICAL_KERNEL(logical_not, CPU,
                               paddle::operators::LogicalNotFunctor);
 REGISTER_BINARY_LOGICAL_OP(logical_xor,
-                           "$$Out = (X || Y) \\, \\&\\& \\, !(X \\&\\& Y)$$");
+                           "$$Out = (X || Y) \\&\\& !(X \\&\\& Y)$$");
 REGISTER_BINARY_LOGICAL_KERNEL(logical_xor, CPU,
                                paddle::operators::LogicalXorFunctor);

paddle/fluid/operators/math/concat.cc

@@ -93,10 +93,10 @@ class ConcatGradFunctor<platform::CPUDeviceContext, T> {
     auto cpu_place = boost::get<platform::CPUPlace>(context.GetPlace());
 
     // computation
-    for (size_t k = 0; k < input_rows; ++k) {
+    for (int k = 0; k < input_rows; ++k) {
       const T* src_ptr = input.data<T>() + k * input_cols;
       int col_idx = 0;
-      for (int j = 0; j < num; ++j) {
+      for (size_t j = 0; j < num; ++j) {
         int col_len = output_cols[j];
         auto* out_tensor = outputs->at(j);
         if (out_tensor != nullptr) {

paddle/fluid/operators/math/math_function.cc

@@ -30,6 +30,7 @@ template struct SetConstant<platform::CPUDeviceContext, double>;
 template struct SetConstant<platform::CPUDeviceContext, int>;
 template struct SetConstant<platform::CPUDeviceContext, int64_t>;
 template struct SetConstant<platform::CPUDeviceContext, bool>;
+template struct SetConstant<platform::CPUDeviceContext, uint8_t>;
 
 #define DEFINE_CPU_TRANS(RANK)                                             \
   template struct Transpose<platform::CPUDeviceContext, platform::float16, \

paddle/fluid/operators/prefetch_op.cc

@@ -41,8 +41,8 @@ class PrefetchOp : public framework::OperatorBase {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
 
-    detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<RPCCLIENT_T>();
+    distributed::RPCClient* rpc_client =
+        distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
     for (size_t i = 0; i < ins.size(); i++) {
       if (NeedSend(scope, ins[i])) {

paddle/fluid/operators/recv_op.cc

@@ -43,8 +43,8 @@ class RecvOp : public framework::OperatorBase {
     // For profiling
     platform::RecordEvent record_event(Type(), &ctx);
 
-    detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<RPCCLIENT_T>();
+    distributed::RPCClient* rpc_client =
+        distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
     for (size_t i = 0; i < outs.size(); i++) {
       VLOG(3) << "getting " << outs[i] << " from " << epmap[i];

paddle/fluid/operators/send_barrier_op.cc

@@ -44,8 +44,8 @@ class SendBarrierOp : public framework::OperatorBase {
     // For profiling
     platform::RecordEvent record_event(Type(), &ctx);
 
-    detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<RPCCLIENT_T>();
+    distributed::RPCClient* rpc_client =
+        distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
     VLOG(3) << "SendBarrierOp sync_mode:" << sync_mode;
 

paddle/fluid/operators/send_op.cc

@@ -45,8 +45,8 @@ class SendOp : public framework::OperatorBase {
     // For profiling
     platform::RecordEvent record_event(Type(), &ctx);
 
-    detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<RPCCLIENT_T>();
+    distributed::RPCClient* rpc_client =
+        distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
     for (size_t i = 0; i < ins.size(); i++) {
       if (NeedSend(scope, ins[i])) {

paddle/fluid/operators/tensorrt_engine_op.cc

@@ -14,11 +14,14 @@
 
 #ifdef PADDLE_WITH_CUDA
 
-#include "paddle/fluid/operators/tensorrt_engine_op.h"
+#include <string>
+#include <vector>
+
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
 #include "paddle/fluid/inference/utils/singleton.h"
+#include "paddle/fluid/operators/tensorrt_engine_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/tensorrt_engine_op.h

@@ -16,10 +16,12 @@
 
 #ifdef PADDLE_WITH_CUDA
 
+#include <string>
+#include <vector>
+
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
-#include "paddle/fluid/inference/tensorrt/engine.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/tensorrt_engine_op_test.cc

@@ -179,7 +179,6 @@ void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) {
                         const std::string& z_name, bool x_created,
                         const shape_t& x_shape, const shape_t& y_shape,
                         const shape_t& z_shape) {
-
     LOG(INFO) << "create fc op";
     auto* fc = block_desc.AppendOp();
     fc->SetType("mul");

paddle/fluid/operators/test_send_nccl_id.cc

@@ -21,7 +21,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/operators/detail/macros.h"
-#include "paddle/fluid/operators/detail/request_handler_impl.h"
+#include "paddle/fluid/operators/distributed/request_handler_impl.h"
 #include "paddle/fluid/operators/listen_and_serv_op.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
@@ -37,11 +37,11 @@ USE_NO_KERNEL_OP(listen_and_serv);
 namespace f = paddle::framework;
 namespace p = paddle::platform;
 namespace m = paddle::operators::math;
-namespace detail = paddle::operators::detail;
+namespace distributed = paddle::operators::distributed;
 namespace string = paddle::string;
 
-std::unique_ptr<detail::RPCServer> g_rpc_service;
-std::unique_ptr<detail::RequestHandler> g_req_handler;
+std::unique_ptr<distributed::RPCServer> g_rpc_service;
+std::unique_ptr<distributed::RequestHandler> g_req_handler;
 
 void StartServer() {
   f::Scope scope;
@@ -57,14 +57,14 @@ void StartServer() {
   g_req_handler->SetProgram(&empty_program);
   g_req_handler->SetExecutor(&executor);
 
-  g_rpc_service->RegisterRPC(detail::kRequestSend, g_req_handler.get());
+  g_rpc_service->RegisterRPC(distributed::kRequestSend, g_req_handler.get());
   g_req_handler->SetRPCServer(g_rpc_service.get());
 
   std::thread server_thread(
-      std::bind(&detail::RPCServer::StartServer, g_rpc_service.get()));
+      std::bind(&distributed::RPCServer::StartServer, g_rpc_service.get()));
 
-  g_rpc_service->SetCond(detail::kRequestSend);
-  g_rpc_service->WaitBarrier(detail::kRequestSend);
+  g_rpc_service->SetCond(distributed::kRequestSend);
+  g_rpc_service->WaitBarrier(distributed::kRequestSend);
 
   LOG(INFO) << "got nccl id and stop server...";
   g_rpc_service->ShutDown();
@@ -72,7 +72,7 @@ void StartServer() {
 }
 
 TEST(SendNcclId, RPCServer) {
-  g_req_handler.reset(new detail::RequestSendHandler(true));
+  g_req_handler.reset(new distributed::RequestSendHandler(true));
   g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1));
 
   std::thread server_thread(StartServer);
@@ -91,7 +91,8 @@ TEST(SendNcclId, RPCServer) {
 
   std::string ep = string::Sprintf("127.0.0.1:%d", port);
 
-  detail::RPCClient* client = detail::RPCClient::GetInstance<RPCCLIENT_T>();
+  distributed::RPCClient* client =
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
   LOG(INFO) << "connect to server" << ep;
   client->AsyncSendVar(ep, dev_ctx, scope, NCCL_ID_VARNAME);

paddle/fluid/pybind/pybind.cc

@@ -159,6 +159,11 @@ PYBIND11_PLUGIN(core) {
              new (&instance) LoDTensor(new_offset_lod);
            })
       .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); })
+      // We implement offset based LOD in C++ while we use length based with
+      // Python API. So we changed set_lod to set_recursive_sequence_lengths to
+      // avoid misuse.
+      // The discussion is here:
+      // https://github.com/PaddlePaddle/Paddle/issues/10855
       .def("set_lod",
            [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
              // the input lod is offset-based level-of-detail info
@@ -199,6 +204,7 @@ PYBIND11_PLUGIN(core) {
              std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
              return new_lod;
            })
+      // Set above comments of set_lod.
       .def("recursive_sequence_lengths",
            [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
              // output the length-based lod info

paddle/fluid/pybind/tensor_py.h

@@ -97,7 +97,7 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
 inline pybind11::buffer_info CastToPyBuffer(const framework::Tensor &tensor) {
   auto buffer_info =
       details::CastToPyBufferImpl<true, 0, float, int, double, int64_t, bool,
-                                  platform::float16>()(tensor);
+                                  uint8_t, platform::float16>()(tensor);
   return buffer_info;
 }
 

python/paddle/fluid/__init__.py

@@ -44,7 +44,7 @@ import metrics
 import transpiler
 from param_attr import ParamAttr, WeightNormParamAttr
 from data_feeder import DataFeeder
-from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace
+from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace, Scope
 from transpiler import DistributeTranspiler, InferenceTranspiler, \
     memory_optimize, release_memory
 from concurrency import (Go, make_channel, channel_send, channel_recv,
@@ -83,6 +83,7 @@ __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + \
               'profiler',
               'unique_name',
               'recordio_writer',
+              'Scope',
           ]
 
 

python/paddle/fluid/average.py

@@ -36,6 +36,25 @@ def _is_number_or_matrix_(var):
 
 
 class WeightedAverage(object):
+    """
+    Calculate weighted average.
+
+    The average calculating is accomplished via Python totally. 
+    They do not change Paddle's Program, nor do anything to
+    modify NN model's configuration. They are completely 
+    wrappers of Python functions.
+
+    Examples:
+        .. code-block:: python
+            avg = fluid.average.WeightedAverage()
+            avg.add(value=2.0, weight=1)
+            avg.add(value=4.0, weight=2)
+            avg.eval()
+
+            # The result is 3.333333333.
+            # For (2.0 * 1 + 4.0 * 2) / (1 + 2) = 3.333333333
+    """
+
     def __init__(self):
         warnings.warn(
             "The %s is deprecated, please use fluid.metrics.Accuracy instead." %

python/paddle/fluid/backward.py

@@ -147,7 +147,7 @@ def _addup_repetitive_outputs_(op_descs):
             else:
                 if len(renamed_vars[var_name]) == 1:
                     new_name = var_name + "@RENAME@" + \
-                               str(var_rename_count[var_name])
+                        str(var_rename_count[var_name])
                     var_rename_count[var_name] += 1
                     # rename original var_name
                     renamed_vars[var_name][0] = new_name
@@ -155,7 +155,7 @@ def _addup_repetitive_outputs_(op_descs):
                     _rename_arg_(pending_sum_ops, var_name, new_name)
 
                 new_name = var_name + "@RENAME@" + \
-                           str(var_rename_count[var_name])
+                    str(var_rename_count[var_name])
                 var_rename_count[var_name] += 1
                 op_desc.rename_output(var_name, new_name)
                 renamed_vars[var_name].append(new_name)
@@ -435,18 +435,65 @@ def _get_stop_gradients_(program):
 def append_backward(loss, parameter_list=None, no_grad_set=None,
                     callbacks=None):
     """
-    Append backward part to main_program
+    Append backward part to main_program.
 
-    Args:
-        loss(Variable): The variable generated by cost function.
-        parameter_list(list[string]): Parameters that need to be updated by
-            optimizer. If None, it means all parameters need to be updated.
-        no_grad_set(set): Variables that have no gradients in Block 0.
-            All variables with `step_gradient=True` from all blocks will be
-            automatically added.
+    A complete neural network training is made up of forward and backward 
+    propagation. However, when we configure a network, we only need to 
+    specify its forwrd part. The backward part is generated automatically 
+    according to the forward part by this function.
 
-    Return:
-        (list[(Variable,Variable)]): list of (parameter, gradient) pair.
+    In most cases, users do not need to invoke this function manually. It 
+    will be automatically invoked by the optimizer's `minimize` function.
+
+    Args:
+        loss(Variable): The loss variable of the network.
+        parameter_list(list[string]|None): Names of parameters that need 
+                                           to be updated by optimizers. 
+                                           If it is None, all parameters 
+                                           will be updated.
+                                           Default: None
+        no_grad_set(set|None): Variables in the Block 0 whose gradients 
+                               should be ignored. All variables with 
+                               `step_gradient=True` from all blocks will 
+                               be automatically added into this set.
+                               Default: None
+        callbacks(list[callable object]|None): The callbacks are used for 
+                                               doing some custom jobs during 
+                                               backward part building. All 
+                                               callable objects in it will 
+                                               be invoked once each time a 
+                                               new gradient operator is added 
+                                               into the program. The callable 
+                                               object must has two input 
+                                               parameters: 'block' and 'context'. 
+                                               The 'block' is the block which 
+                                               the new gradient operator will 
+                                               be added to. The 'context' is a 
+                                               map, whose keys are gradient 
+                                               variable names and values are 
+                                               corresponding original variables.
+                                               In addition to this, the 'context' 
+                                               has another special key-value pair: 
+                                               the key is string '__current_op_desc__' 
+                                               and the value is the op_desc of the 
+                                               gradient operator who has just 
+                                               triggered the callable object. 
+
+    Returns:
+        list[(Variable,Variable)]: Pairs of parameter and its 
+        corresponding gradients. The key is the parameter and the 
+        value is gradient variable.
+
+    Raises:
+        AssertionError: If `loss` is not an instance of Variable.
+
+    Examples:
+        .. code-block:: python
+
+            # network configuration code
+            # ...
+            avg_loss = fluid.layers.mean(loss)
+            param_grad_list = fluid.backward.append_backward(loss=avg_loss)
     """
     assert isinstance(loss, framework.Variable)
 

python/paddle/fluid/data_feeder.py

@@ -29,6 +29,13 @@ class DataToLoDTensorConverter(object):
         self.place = place
         self.lod_level = lod_level
         self.shape = shape
+        negtive_count = 0
+        for s in self.shape:
+            if s < 0:
+                negtive_count += 1
+            if negtive_count > 1:
+                self.shape = None
+                break
         if dtype == core.VarDesc.VarType.FP32:
             self.dtype = 'float32'
         elif dtype == core.VarDesc.VarType.INT64:
@@ -61,7 +68,9 @@ class DataToLoDTensorConverter(object):
                 self._feed_impl_(each_data, lod[1:], lod_level - 1)
 
     def done(self):
-        arr = numpy.array(self.data, dtype=self.dtype).reshape(self.shape)
+        arr = numpy.array(self.data, dtype=self.dtype)
+        if self.shape:
+            arr = arr.reshape(self.shape)
         t = core.LoDTensor()
         t.set(arr, self.place)
         if self.lod_level > 0:
@@ -70,6 +79,61 @@ class DataToLoDTensorConverter(object):
 
 
 class DataFeeder(object):
+    """
+    DataFeeder converts the data that returned by a reader into a data
+    structure that can feed into Executor and ParallelExecutor. The reader
+    usually returns a list of mini-batch data entries. Each data entry in
+    the list is one sample. Each sample is a list or a tuple with one
+    feature or multiple features.
+
+    The simple usage shows below:
+
+    ..  code-block:: python
+
+        place = fluid.CPUPlace()
+        img = fluid.layers.data(name='image', shape=[1, 28, 28])
+        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        feeder = fluid.DataFeeder([img, label], fluid.CPUPlace())
+        result = feeder.feed([([0] * 784, [9]), ([1] * 784, [1])])
+
+
+    If you want to feed data into GPU side separately in advance when you
+    use multi-GPU to train a model, you can use `decorate_reader` function.
+
+    ..  code-block:: python
+
+        place=fluid.CUDAPlace(0)
+        feeder = fluid.DataFeeder(place=place, feed_list=[data, label])
+        reader = feeder.decorate_reader(
+            paddle.batch(flowers.train(), batch_size=16))
+
+    Args:
+        feed_list(list): The Variables or Variables'name that will
+            feed into model.
+        place(Place): place indicates feed data into CPU or GPU, if you want to
+            feed data into GPU, please using `fluid.CUDAPlace(i)` (`i` represents
+            the GPU id), or if you want to feed data into CPU, please using
+            `fluid.CPUPlace()`.
+        program(Program): The Program that will feed data into, if program
+            is None, it will use default_main_program(). Default None.
+
+    Raises:
+        ValueError: If some Variable is not in this Program.
+
+    Examples:
+        .. code-block:: python
+
+            # ...
+            place = fluid.CPUPlace()
+            feed_list = [
+                main_program.global_block().var(var_name) for var_name in feed_vars_name
+            ] # feed_vars_name is a list of variables' name.
+            feeder = fluid.DataFeeder(feed_list, place)
+            for data in reader():
+                outs = exe.run(program=main_program,
+                               feed=feeder.feed(data))
+    """
+
     def __init__(self, feed_list, place, program=None):
         self.feed_dtypes = []
         self.feed_names = []
@@ -99,6 +163,16 @@ class DataFeeder(object):
         self.place = place
 
     def feed(self, iterable):
+        """
+        According to feed_list and iterable, converters the input into
+        a data structure that can feed into Executor and ParallelExecutor.
+
+        Args:
+            iterable(list|tuple): the input data.
+
+        Returns:
+            dict: the result of conversion.
+        """
         converter = []
         for lod_level, shape, dtype in six.zip(
                 self.feed_lod_level, self.feed_shapes, self.feed_dtypes):
@@ -121,6 +195,20 @@ class DataFeeder(object):
         return ret_dict
 
     def feed_parallel(self, iterable, num_places=None):
+        """
+        Takes multiple mini-batches. Each mini-batch will be feed on each
+        device in advance.
+
+        Args:
+            iterable(list|tuple): the input data.
+            num_places(int): the number of devices. Default None.
+
+        Returns:
+            dict: the result of conversion.
+
+        Notes:
+            The number of devices and number of mini-batches must be same.
+        """
         if isinstance(self.place, core.CUDAPlace):
             places = [
                 core.CUDAPlace(i)
@@ -159,6 +247,24 @@ class DataFeeder(object):
                         multi_devices,
                         num_places=None,
                         drop_last=True):
+        """
+        Converter the input data into a data that returned by reader into
+        multiple mini-batches. Each mini-batch will be feed on each device.
+
+        Args:
+            reader(fun): the input data.
+            multi_devices(bool): the number of places. Default None.
+            num_places(int): the number of places. Default None.
+            drop_last(bool): the number of places. Default None.
+
+        Returns:
+            dict: the result of conversion.
+
+        Raises:
+            ValueError: If drop_last is False and the data batch which cannot
+            fit for devices.
+        """
+
         def __reader_creator__():
             if not multi_devices:
                 for item in reader():

python/paddle/fluid/executor.py

@@ -25,6 +25,13 @@ g_scope = core.Scope()
 
 
 def global_scope():
+    """
+    Get the global/default scope instance. There are a lot of APIs use
+    :code:`global_scope` as its default value, e.g., :code:`Executor.run`
+
+    Returns:
+        Scope: The global/default scope instance.
+    """
     return g_scope
 
 
@@ -37,6 +44,19 @@ def switch_scope(scope):
 
 @contextlib.contextmanager
 def scope_guard(scope):
+    """
+    Change the global/default scope instance by Python `with` statement. All
+    variable in runtime will assigned to the new scope.
+
+    Examples:
+        >>> import paddle.fluid as fluid
+        >>> new_scope = fluid.Scope()
+        >>> with fluid.scope_guard(new_scope):
+        >>>     ...
+
+    Args:
+        scope: The new global/default scope.
+    """
     ex = switch_scope(scope)
     yield
     switch_scope(ex)
@@ -135,14 +155,18 @@ def has_fetch_operators(block, fetch_targets, fetch_holder_name):
 
 def fetch_var(name, scope=None, return_numpy=True):
     """
-    Fetch the value of the variable with the given name from the given scope
+    Fetch the value of the variable with the given name from the
+    given scope.
+
     Args:
         name(str): name of the variable. Typically, only persistable variables
             can be found in the scope used for running the program.
         scope(core.Scope|None): scope object. It should be the scope where
             you pass to Executor.run() when running your program.
-            If None, global_scope() will be used.
-        return_numpy(bool): whether convert the tensor to numpy.ndarray
+            If None, global_scope() will be used. Default None.
+        return_numpy(bool): whether convert the tensor to numpy.ndarray.
+            Default True.
+
     Returns:
        LodTensor|numpy.ndarray
     """

python/paddle/fluid/lod_tensor.py

@@ -19,33 +19,41 @@ __all__ = ['create_lod_tensor', 'create_random_int_lodtensor']
 
 
 def create_lod_tensor(data, lod, place):
-    """Create a lod tensor from a numpy array, a list, or an existing lod tensor.
+    """
+    Create a lod tensor from a numpy array, a list, or an existing lod tensor.
 
     Create a lod tensor by doing the following:
+
     1. Check that the length-based input lod is valid.
+
     2. Convert the length-based lod to a offset-based LoD.
-    3. Copy the data from a numpy array, a list or a existing lod tensor to 
+
+    3. Copy the data from a numpy array, a list or a existing lod tensor to
        CPU or GPU device (based on input place).
+
     4. Set the level of detail (LoD) using the offset-based LoD.
     
-    Use example:
-    Suppose we want LoDTensor to hold data for sequences of word, where each word is
-    represented by an integer. If we want to create a LoDTensor to represent two 
-    sentences, one of 2 words, and one of 3 words. 
+    Examples:
 
-    Then 'data' can be a numpy array of integers with shape (5, 1).
-    'lod' will be [[2, 3]], indicating the length(# of words) in each sentence.
-    This length-based input lod [[2, 3]] will be converted to offset-based lod [[0, 2, 5]]
-    inside the function call.
+        Suppose we want LoDTensor to hold data for sequences of word, where each
+        word is represented by an integer. If we want to create a LoDTensor to
+        represent two  sentences, one of 2 words, and one of 3 words.
 
-    Please refer to 
-    github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/lod_tensor.md
-    for more details regarding LoD.
+        Then :code:`data` can be a numpy array of integers with shape (5, 1).
+        :code:`lod` will be [[2, 3]], indicating the length(# of words) in each
+        sentence. This length-based input lod [[2, 3]] will be converted to
+        offset-based lod [[0, 2, 5]] inside the function call.
+
+    Please reference :ref:`api_guide_low_level_lod_tensor` for more details
+    regarding LoD.
 
     Args:
-        data: a numpy array or a LoDTensor or a list holding the data to be copied.
-        lod: a list of lists indicating the length-based LoD info specified by the user. 
-        place: CPU or GPU place indicating where the data in the new LoDTensor will be stored.
+        data(numpy.ndarray|list|LoDTensor): a numpy array or a LoDTensor or a
+            list holding the data to be  copied.
+        lod(list): a list of lists indicating the length-based LoD info
+            specified by the user.
+        place(Place): CPU or GPU place indicating where the data in the new
+            LoDTensor will be stored.
 
     Returns:
         A fluid LoDTensor object with tensor data and lod info.
@@ -77,31 +85,38 @@ def create_lod_tensor(data, lod, place):
 
 
 def create_random_int_lodtensor(lod, base_shape, place, low, high):
-    """Create a LoDTensor containing random integers.
+    """
+    Create a LoDTensor containing random integers.
 
-    This function is frequently used in the book examples. So we revised it based on 
-    the new create_lod_tensor API and put it here in the lod_tensor module to simplify 
-    the code. 
+    This function is frequently used in the book examples. So we revised it
+    based on the new create_lod_tensor API and put it here in the lod_tensor
+    module to simplify the code.
 
     The function does the following:
-    1. Calculate the overall shape of the LoDTensor based on the length-based 'lod' input 
-    and the shape of the basic element in 'base_shape'.
+
+    1. Calculate the overall shape of the LoDTensor based on the length-based
+       :code:`lod` input and the shape of the basic element in
+       :code:`base_shape`.
+
     2. Create a numpy array of this shape.
+
     3. Create the LoDTensor using create_lod_tensor API.
 
-    Suppose we want LoDTensor to hold data for sequences of word, where each word is
-    represented by an integer. If we want to create a LoDTensor to represent two 
-    sentences, one of 2 words, and one of 3 words. Then 'base_shape' is [1], input 
-    length-based 'lod' is [[2, 3]]. Then the overall shape of the LoDTensor would be 
-    [5, 1], holding 5 words for two sentences. 
+    Suppose we want LoDTensor to hold data for sequences of word, where each
+    word is represented by an integer. If we want to create a LoDTensor to
+    represent two sentences, one of 2 words, and one of 3 words. Then
+    'base_shape' is [1], input length-based 'lod' is [[2, 3]]. Then the overall
+    shape of the LoDTensor would be [5, 1], holding 5 words for two sentences.
 
     Args:
-        data: a numpy array or a LoDTensor holding the data to be copied.
-        lod: a list of lists indicating the length-based LoD info specified by the user.
-        base_shape: the shape of the basic element to be held by the LoDTensor. 
-        place: CPU or GPU place indicating where the data in the new LoDTensor will be stored.
-        low: the lower bound of the random integers.
-        high: the upper bound of the random integers.
+        lod(list): a list of lists indicating the length-based LoD info
+            specified by the user.
+        base_shape(list): the shape of the basic element to be held by the
+            LoDTensor.
+        place(Place): CPU or GPU place indicating where the data in the new
+            LoDTensor will be stored.
+        low(int): the lower bound of the random integers.
+        high(int): the upper bound of the random integers.
 
     Returns:
         A fluid LoDTensor object with tensor data and lod info. 

python/paddle/fluid/metrics.py

@@ -325,14 +325,14 @@ class Auc(MetricBase):
     """
 
     def __init__(self, name, curve='ROC', num_thresholds=200):
-        super(MetricBase, self).__init__(name, curve, num_thresholds)
+        super(Auc, self).__init__(name=name)
         self._curve = curve
         self._num_thresholds = num_thresholds
         self._epsilon = 1e-6
-        self.tp_list = np.ndarray((num_thresholds, ))
-        self.fn_list = np.ndarray((num_thresholds, ))
-        self.tn_list = np.ndarray((num_thresholds, ))
-        self.fp_list = np.ndarray((num_thresholds, ))
+        self.tp_list = np.zeros((num_thresholds, ))
+        self.fn_list = np.zeros((num_thresholds, ))
+        self.tn_list = np.zeros((num_thresholds, ))
+        self.fp_list = np.zeros((num_thresholds, ))
 
     def update(self, labels, predictions, axis=1):
         if not _is_numpy_(labels):
@@ -350,12 +350,12 @@ class Auc(MetricBase):
             tp, fn, tn, fp = 0, 0, 0, 0
             for i, lbl in enumerate(labels):
                 if lbl:
-                    if predictions[i, 0] >= thresh:
+                    if predictions[i, 1] >= thresh:
                         tp += 1
                     else:
                         fn += 1
                 else:
-                    if predictions[i, 0] >= thresh:
+                    if predictions[i, 1] >= thresh:
                         fp += 1
                     else:
                         tn += 1

python/paddle/fluid/parallel_executor.py

@@ -27,6 +27,40 @@ BuildStrategy = core.ParallelExecutor.BuildStrategy
 
 
 class ParallelExecutor(object):
+    """
+    ParallelExecutor can run program in parallel.
+
+    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 provied, it will share variables
+            from the specified ParallelExecutor. 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.
+
+    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,
@@ -37,42 +71,6 @@ class ParallelExecutor(object):
                  num_trainers=1,
                  trainer_id=0,
                  **kwargs):
-        """
-        ParallelExecutor can run program in parallel.
-
-        Args:
-            use_cuda(bool): Whether to use CUDA or not.
-            loss_name(str, default None): The loss name must set in training.
-            main_program(Program, default None): The program that need to run,
-                if not provided, then default_main_program will be used.
-            share_vars_from(ParallelExecutor, default None): If provied,
-                it will share variables from the specified ParallelExecutor.
-            num_trainers(int, default 1): If greater than 1, NCCL will be
-                initialized with multpile rank of nodes, each node should have
-                same number of GPUs. Distributed training will be enabled then.
-            trainer_id(int, default 0): Must use together with num_trainers.
-                trainer_id is the "rank" of current node starts from 0.
-
-        Returns:
-            A 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)
-        """
         if len(kwargs) != 0:
             err_msg = ""
             for key in kwargs:
@@ -131,10 +129,16 @@ class ParallelExecutor(object):
         main = main_program
         main = main if main else framework.default_main_program()
         scope = executor.global_scope()
+        # FIXME(Yancey1989): it's a temporary approach to determinate the distribute
+        # train program, call self.bcast_param() at the end of each mini-batch.
+        self.is_dist = True if "recv" in [
+            op.type for op in main.global_block().ops
+        ] else False
 
         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 []
 
@@ -166,12 +170,14 @@ class ParallelExecutor(object):
         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)
@@ -182,18 +188,40 @@ class ParallelExecutor(object):
         >>>               {"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.
+                to each device. Default None.
             feed_dict: Alias for feed parameter, for backward compatibility.
-                This parameter is deprecated.
+                This parameter has been deprecated. Default None.
+
+        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.
 
-        Returns: fetched result list.
+        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
@@ -238,9 +266,17 @@ class ParallelExecutor(object):
         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 self.is_dist:
+            self.bcast_params()
+
         return [arr[i] for i in range(len(arr))]
 
     def bcast_params(self):
+        """
+        Broadcast the parameters to other devices. It is used during
+        distributed training.
+        """
         self.executor.bcast_params(set(self.persistable_vars))
 
     @property

python/paddle/fluid/recordio_writer.py

@@ -36,6 +36,45 @@ def convert_reader_to_recordio_file(
         compressor=core.RecordIOWriter.Compressor.Snappy,
         max_num_records=1000,
         feed_order=None):
+    """
+    Convert a Python Reader to a recordio file.
+
+    Please see :ref:`api_guide_python_reader` and :ref:`api_guide_reader_op` for
+    details.
+
+    Examples:
+
+        >>> import paddle.fluid as fluid
+        >>> import paddle.dataset.mnist as mnist
+        >>> import paddle
+        >>>
+        >>> tmp_program = fluid.Program()
+        >>> with fluid.program_guard(tmp_program):
+        >>>     img = fluid.layers.data(name='img', shape=[784])
+        >>>     label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        >>> feeder = fluid.DataFeeder(feed_list=[img, label], place=fluid.CPUPlace())
+        >>> # mnist.recordio will be generated in current directory
+        >>> fluid.recordio_writer.convert_reader_to_recordio_file(
+        >>>                     filename="mnist.recordio",
+        >>>                     reader_creator=paddle.batch(mnist.train(), batch_size=32),
+        >>>                     feeder=feeder)
+
+    Args:
+        filename(str): The recordio filename.
+        reader_creator(callable): The Python Reader Creator. See
+            :ref:`api_guide_python_reader`.
+        feeder(DataFeeder): The DataFeeder instance. Used to convert
+            :code:`reader_creator` to :code: `lod_tensor`
+        compressor: Must in fluid.core.RecordIOWriter.Compressor.Snappy or
+            fluid.core.RecordIOWriter.Compressor.NoCompress. Use :code:`Snappy`
+            by default.
+        max_num_records(int): Maximum number of records in one chuck. Each record
+            is each return value from reader function
+        feed_order(list): The order of variable names that the reader returns
+
+    Returns:
+        int: the number of record that saved.
+    """
     if feed_order is None:
         feed_order = feeder.feed_names
     counter = 0
@@ -58,6 +97,17 @@ def convert_reader_to_recordio_files(
         compressor=core.RecordIOWriter.Compressor.Snappy,
         max_num_records=1000,
         feed_order=None):
+    """
+    convert a python reader to many recordio files.
+
+    This API is basically same as :code:`convert_reader_to_recordio_file`,
+    instead of it will create many recordio files. Each file contains at
+    most :code:`batch_per_file` records.
+
+    Please reference
+    :ref:`api_fluid_recordio_writer_convert_reader_to_recordio_file` for more
+    details.
+    """
     if feed_order is None:
         feed_order = feeder.feed_names
     f_name, f_ext = os.path.splitext(filename)

python/paddle/fluid/transpiler/memory_optimization_transpiler.py

@@ -383,6 +383,16 @@ def memory_optimize(input_program, skip_opt_set=None, print_log=False, level=0):
 
 
 def release_memory(input_program, skip_opt_set=None):
+    """
+    Modify the input program and insert :code:`delete_op` to early drop not used
+    variables. The modification will be performed inplace.
+
+    Notes: This is an experimental API and could be removed in next few
+    releases. Users should not use this API.
+
+    Args:
+        input_program(Program): The program will be inserted :code:`delete_op`.
+    """
     cfgs = _get_cfgs(input_program)
     for cfg in cfgs:
         cfg.release_memory(skip_opt_set=skip_opt_set)

python/paddle/fluid/unique_name.py

@@ -16,7 +16,7 @@ import collections
 import contextlib
 import sys
 
-__all__ = ['generate', 'switch', 'guard', 'UniqueNameGenerator']
+__all__ = ['generate', 'switch', 'guard']
 
 
 class UniqueNameGenerator(object):