merge develop branch

CMakeLists.txt

@@ -207,6 +207,10 @@ endif()
 
 include(external/threadpool)
 
+include(flags)              # set paddle compile flags
+include(cudnn)              # set cudnn libraries, must before configure
+include(configure)          # add paddle env configuration
+
 if(WITH_GPU)
     include(cuda)
     include(tensorrt)

benchmark/fluid/fluid_benchmark.py

@@ -151,6 +151,7 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
                 if data == None:
                     break
             if iters == args.iterations:
+                reader_generator.close()
                 break
             if iters == args.skip_batch_num:
                 start_time = time.time()
@@ -252,6 +253,7 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
                 if data == None:
                     break
             if iters == args.iterations:
+                reader_generator.close()
                 break
             if args.profile and pass_id == 0 and batch_id == 5:
                 profiler.start_profiler("All")

cmake/external/anakin.cmake

@@ -52,9 +52,8 @@ ExternalProject_Add(
     extern_anakin
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLML_PROJECT}
-    # Anakin codes error on Intel(R) Xeon(R) Gold 5117 CPU, temporary do not compile avx512 related code.
-    GIT_REPOSITORY      "https://github.com/luotao1/Anakin"
-    GIT_TAG             "211d1fc5d813d70c0c14072f9083cf25f40940ea"
+    GIT_REPOSITORY      "https://github.com/PaddlePaddle/Anakin"
+    GIT_TAG             "9424277cf9ae180a14aff09560d3cd60a49c76d2"
     PREFIX              ${ANAKIN_SOURCE_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DUSE_GPU_PLACE=YES

cmake/external/boost.cmake

@@ -46,8 +46,13 @@ ExternalProject_Add(
     ${BOOST_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DOWNLOAD_DIR          ${BOOST_DOWNLOAD_DIR}
+<<<<<<< HEAD
     DOWNLOAD_COMMAND      "wget --no-check-certificate ${BOOST_URL} -c -q -O ${BOOST_TAR}.tar.gz
                           && tar zxf ${BOOST_TAR}.tar.gz"
+=======
+    DOWNLOAD_COMMAND      wget --no-check-certificate ${BOOST_URL} -c -q -O ${BOOST_TAR}.tar.gz
+    && tar zxf ${BOOST_TAR}.tar.gz
+>>>>>>> origin/develop
     DOWNLOAD_NO_PROGRESS  1
     PREFIX                ${BOOST_SOURCES_DIR}
     CONFIGURE_COMMAND     ""
@@ -57,7 +62,7 @@ ExternalProject_Add(
 )
 endif(NOT WIN32)
 
-if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
+if (${CMAKE_VERSION} VERSION_LESS "3.3.0" OR NOT WIN32)
     set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c)
     file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";")
     add_library(boost STATIC ${dummyfile})

cmake/external/openblas.cmake

@@ -17,6 +17,9 @@ IF(USE_EIGEN_FOR_BLAS)
 ENDIF(USE_EIGEN_FOR_BLAS)
 
 INCLUDE(cblas)
+# IF(WIN32 AND NOT ${CBLAS_FOUND})
+
+
 
 IF(NOT ${CBLAS_FOUND})
 

cmake/generic.cmake

@@ -218,14 +218,18 @@ function(merge_static_libs TARGET_NAME)
 
     foreach(lib ${libs})
       # Get the file names of the libraries to be merged
+      #if(NOT $<TARGET_FILE:${lib}> MATCHES "lib.*\\.lib")
+      #  message("library" ${lib})
+      #  set(libfiles ${libfiles} lib$<TARGET_FILE:${lib}>)
+      #else()
       set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
+      #endif()
     endforeach()
-    
-    # msvc will put libarary in directory of "/Release/xxxlib" by default 
-    #       COMMAND cmake -E remove "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/${TARGET_NAME}.lib"
+   
+    # windows cmd return error in clean env.
+    # COMMAND del "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/${TARGET_NAME}.lib"
     add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
-      COMMAND cmake -E make_directory "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}"
-      COMMAND lib /OUT:${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/lib${TARGET_NAME}.lib ${libfiles}
+      COMMAND lib /OUT:${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.lib ${libfiles}
       )
   endif(WIN32)
 endfunction(merge_static_libs)

doc/fluid/dev/releasing_process_en.md

@@ -50,6 +50,33 @@ pop-up box, choose the current release branch and click "Run Build" button. You
 * pypi does not allow overwrite the already uploaded version of wheel package, even if you delete the
   old version. you must change the version number before upload a new one.
 
+### Publish wheel Packages for MacOS
+
+You need to build the binary wheel package for MacOS before publishing, to
+make sure that the package can be used by many versions of MacOS
+(10.11, 10.12, 10.13) and different python installs (python.org, homebrew, etc.),
+you must build the package ***exactly*** following below steps:
+
+Build steps:
+
+1. install python from python.org downloads, and make sure it's currently in use
+   in your system.
+1. `export MACOSX_DEPLOYMENT_TARGET=10.11`, use `10.11` is enough for recent versions.
+1. `git clone https://github.com/PaddlePaddle/Paddle.git && cd Paddle && mkdir build && cd build`
+1. `cmake -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_SYSTEM_BLAS=OFF  ..`, make sure the output of `cmake` command is using the correct python interpreter installed from python.org
+1. `make -j`
+1. `pip install delocate`
+1. `mkdir fixed_wheel && delocate-wheel -w fixed_wheel python/dist/*.whl`
+
+Then the whl under `fixed_wheel` is ready to upload.
+
+Install steps:
+
+1. run `pip install paddlepaddle...whl`
+1. find the `libpython.dylib` that are currently in use:
+    - for python.org package installs, do nothing.
+    - for other python installs, find the path of `libpython*.dylib` and `export LD_LIBRARY_PATH=you path && DYLD_LIBRARY_PATH=your path`
+
 ## Publish Docker Images
 
 Our CI tool will push latest images to DockerHub, so we only need to push a version tag like:

doc/fluid/new_docs/advanced_usage/deploy/native_infer.rst

@@ -9,8 +9,6 @@ Paddle 预测 API
 
 -  头文件 ``paddle_inference_api.h`` 定义了所有的接口
 -  库文件\ ``libpaddle_fluid.so`` 或 ``libpaddle_fluid.a``
--  库文件 ``libpaddle_inference_api.so`` 或
-   ``libpaddle_inference_api.a``
 
 编译和依赖可以参考 :ref:`install_or_build_cpp_inference_lib` 。
 
@@ -97,8 +95,7 @@ engine
     CHECK(predictor->Run(slots, &outputs));
     // 获取 outputs ...
 
-编译时，联编 ``libpaddle_fluid.a/.so`` 和
-``libpaddle_inference_api.a/.so`` 便可。
+编译时，联编 ``libpaddle_fluid.a/.so`` 即可。
 
 详细代码参考
 ------------

doc/fluid/new_docs/beginners_guide/basics/index.rst

@@ -10,9 +10,9 @@
 ..  toctree::
     :maxdepth: 2
 
-    image_classification/index.md
-    word2vec/index.md
-    recommender_system/index.md
-    understand_sentiment/index.md
-    label_semantic_roles/index.md
-    machine_translation/index.md
+    image_classification/README.cn.md
+    word2vec/README.cn.md
+    recommender_system/README.cn.md
+    understand_sentiment/README.cn.md
+    label_semantic_roles/README.cn.md
+    machine_translation/README.cn.md

doc/fluid/new_docs/beginners_guide/install/install_doc.rst

@@ -57,7 +57,28 @@ paddlepaddle-gpu==0.11.0            使用CUDA 7.5和cuDNN 5编译的0.11.0版
 您可以在 `Release History <https://pypi.org/project/paddlepaddle-gpu/#history>`_
 中找到paddlepaddle-gpu的各个发行版本。
 
-如果需要获取并安装最新的PaddlePaddle开发分支，可以从我们的 `CI系统 <https://paddleci.ngrok.io/project.html?projectId=Manylinux1&tab=projectOverview>`_ 中下载最新的whl安装包和c-api开发包并安装。如需登录，请点击“Log in as guest”。
+如果需要获取并安装最新的（开发分支）PaddlePaddle，可以从我们的CI系统中下载最新的whl
+安装包和c-api开发包并安装，您可以从下面的表格中找到需要的版本：
+
+如果在点击下面链接时出现如下登陆界面，点击“Log in as guest”即可开始下载：
+
+.. image:: paddleci.png
+   :scale: 50 %
+   :align: center
+
+..  csv-table:: 各个版本最新的whl包
+    :header: "版本说明", "cp27-cp27mu", "cp27-cp27m"
+    :widths: 1, 3, 3
+
+    "stable_cuda9.0_cudnn7", "`paddlepaddle_gpu-0.14.0-cp27-cp27mu-manylinux1_x86_64.whl <https://files.pythonhosted.org/packages/ee/ee/5d96e99d4a6d57bd1a7a8c4c98124a5ba0f6f0e07f38f4cee1365e0d9734/paddlepaddle_gpu-0.14.0-cp27-cp27mu-manylinux1_x86_64.whl>`__", "`paddlepaddle_gpu-0.14.0-cp27-cp27m-manylinux1_x86_64.whl <https://files.pythonhosted.org/packages/2e/65/3c1e44417dfc4afc7004f4db06789876b1237a0b6b234e0bd4213f3258b7/paddlepaddle_gpu-0.14.0-cp27-cp27m-manylinux1_x86_64.whl>`__"
+    "stable_cuda8.0_cudnn7", "`paddlepaddle_gpu-0.14.0.post87-cp27-cp27mu-manylinux1_x86_64.whl <https://files.pythonhosted.org/packages/a1/eb/261d920ede38d4b2b8dfb5817d7f7d25c526b1a70260f23312ad6029c0d3/paddlepaddle_gpu-0.14.0.post87-cp27-cp27mu-manylinux1_x86_64.whl>`__", "`paddlepaddle_gpu-0.14.0.post87-cp27-cp27m-manylinux1_x86_64.whl <https://files.pythonhosted.org/packages/54/1d/2c2a5c8665634b47fa925839108752611202a7c08ba4d65c2ee79f825a0e/paddlepaddle_gpu-0.14.0.post87-cp27-cp27m-manylinux1_x86_64.whl>`__"
+    "stable_cuda8.0_cudnn5", "`paddlepaddle_gpu-0.14.0.post85-cp27-cp27mu-manylinux1_x86_64.whl <https://files.pythonhosted.org/packages/60/50/94d16d34976f06b3cd8818d9b7bf40a9ff16bc48120ac9254d976f8ffc35/paddlepaddle_gpu-0.14.0.post85-cp27-cp27mu-manylinux1_x86_64.whl>`__", "`paddlepaddle_gpu-0.14.0.post85-cp27-cp27m-manylinux1_x86_64.whl <https://files.pythonhosted.org/packages/24/dd/25c1db09524f654c80baa83e7aafdd67109449bd5b500964f4005047dcf8/paddlepaddle_gpu-0.14.0.post85-cp27-cp27m-manylinux1_x86_64.whl>`__"
+    "cpu_avx_mkl", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/845:id/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/845:id/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`__"
+    "cpu_avx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/846:id/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/846:id/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`__"
+    "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/847:id/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/847:id/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_"
+    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/841:id/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/841:id/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
+    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/843:id/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/843:id/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
+    "cuda9.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/842:id/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/842:id/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
 
 .. _FAQ:
 

paddle/fluid/API.spec

@@ -36,6 +36,7 @@ paddle.fluid.default_startup_program ArgSpec(args=[], varargs=None, keywords=Non
 paddle.fluid.default_main_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
 paddle.fluid.program_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
 paddle.fluid.get_var ArgSpec(args=['name', 'program'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.name_scope ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
 paddle.fluid.Executor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Executor.close ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False))
@@ -55,9 +56,10 @@ paddle.fluid.Inferencer.__init__ ArgSpec(args=['self', 'infer_func', 'param_path
 paddle.fluid.Inferencer.infer ArgSpec(args=['self', 'inputs', 'return_numpy'], varargs=None, keywords=None, defaults=(True,))
 paddle.fluid.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True))
+paddle.fluid.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode', 'startup_program'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True, None))
 paddle.fluid.InferenceTranspiler.__init__ 
 paddle.fluid.InferenceTranspiler.transpile ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0))
@@ -113,6 +115,7 @@ paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size
 paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None))
+paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None))
 paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
@@ -143,9 +146,12 @@ paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_
 paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1))
 paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None))
+paddle.fluid.layers.squeeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.unsqueeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lod_reset ArgSpec(args=['x', 'y', 'target_lod'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.lrn ArgSpec(args=['input', 'n', 'k', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(5, 1.0, 0.0001, 0.75, None))
 paddle.fluid.layers.pad ArgSpec(args=['x', 'paddings', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None))
+paddle.fluid.layers.pad_constant_like ArgSpec(args=['x', 'y', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None))
 paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 0.1, 'float32', None))
 paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0))
 paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,))
@@ -162,7 +168,10 @@ paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs
 paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.prelu ArgSpec(args=['x', 'mode', 'param_attr', 'name'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.flatten ArgSpec(args=['x', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None))
+paddle.fluid.layers.sequence_mask ArgSpec(args=['x', 'maxlen', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 'int64', None))
 paddle.fluid.layers.stack ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=(0,))
+paddle.fluid.layers.pad2d ArgSpec(args=['input', 'paddings', 'mode', 'pad_value', 'data_format', 'name'], varargs=None, keywords=None, defaults=([0, 0, 0, 0], 'constant', 0.0, 'NCHW', None))
+paddle.fluid.layers.unstack ArgSpec(args=['x', 'axis', 'num'], varargs=None, keywords=None, defaults=(0, None))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_recordio_file ArgSpec(args=['filename', 'shapes', 'lod_levels', 'dtypes', 'pass_num', 'for_parallel'], varargs=None, keywords=None, defaults=(1, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
@@ -251,6 +260,7 @@ paddle.fluid.layers.logical_xor ArgSpec(args=[], varargs='args', keywords='kwarg
 paddle.fluid.layers.logical_not ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.uniform_random_batch_size_like ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.gaussian_random ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
+paddle.fluid.layers.sampling_id ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.gaussian_random_batch_size_like ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.sum ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.slice ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
@@ -293,8 +303,10 @@ paddle.fluid.layers.target_assign ArgSpec(args=['input', 'matched_indices', 'neg
 paddle.fluid.layers.detection_output ArgSpec(args=['loc', 'scores', 'prior_box', 'prior_box_var', 'background_label', 'nms_threshold', 'nms_top_k', 'keep_top_k', 'score_threshold', 'nms_eta'], varargs=None, keywords=None, defaults=(0, 0.3, 400, 200, 0.01, 1.0))
 paddle.fluid.layers.ssd_loss ArgSpec(args=['location', 'confidence', 'gt_box', 'gt_label', 'prior_box', 'prior_box_var', 'background_label', 'overlap_threshold', 'neg_pos_ratio', 'neg_overlap', 'loc_loss_weight', 'conf_loss_weight', 'match_type', 'mining_type', 'normalize', 'sample_size'], varargs=None, keywords=None, defaults=(None, 0, 0.5, 3.0, 0.5, 1.0, 1.0, 'per_prediction', 'max_negative', True, None))
 paddle.fluid.layers.detection_map ArgSpec(args=['detect_res', 'label', 'class_num', 'background_label', 'overlap_threshold', 'evaluate_difficult', 'has_state', 'input_states', 'out_states', 'ap_version'], varargs=None, keywords=None, defaults=(0, 0.3, True, None, None, None, 'integral'))
-paddle.fluid.layers.rpn_target_assign ArgSpec(args=['loc', 'scores', 'anchor_box', 'gt_box', 'rpn_batch_size_per_im', 'fg_fraction', 'rpn_positive_overlap', 'rpn_negative_overlap'], varargs=None, keywords=None, defaults=(256, 0.25, 0.7, 0.3))
+paddle.fluid.layers.rpn_target_assign ArgSpec(args=['loc', 'scores', 'anchor_box', 'anchor_var', 'gt_box', 'rpn_batch_size_per_im', 'fg_fraction', 'rpn_positive_overlap', 'rpn_negative_overlap'], varargs=None, keywords=None, defaults=(256, 0.25, 0.7, 0.3))
 paddle.fluid.layers.anchor_generator ArgSpec(args=['input', 'anchor_sizes', 'aspect_ratios', 'variance', 'stride', 'offset', 'name'], varargs=None, keywords=None, defaults=(None, None, [0.1, 0.1, 0.2, 0.2], None, 0.5, None))
+paddle.fluid.layers.generate_proposal_labels ArgSpec(args=['rpn_rois', 'gt_classes', 'gt_boxes', 'im_scales', 'batch_size_per_im', 'fg_fraction', 'fg_thresh', 'bg_thresh_hi', 'bg_thresh_lo', 'bbox_reg_weights', 'class_nums'], varargs=None, keywords=None, defaults=(256, 0.25, 0.25, 0.5, 0.0, [0.1, 0.1, 0.2, 0.2], None))
+paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'im_info', 'anchors', 'variances', 'pre_nms_top_n', 'post_nms_top_n', 'nms_thresh', 'min_size', 'eta', 'name'], varargs=None, keywords=None, defaults=(6000, 1000, 0.5, 0.1, 1.0, None))
 paddle.fluid.layers.iou_similarity ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.box_coder ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.polygon_box_transform ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
@@ -330,9 +342,10 @@ paddle.fluid.contrib.BeamSearchDecoder.update_array ArgSpec(args=['self', 'array
 paddle.fluid.contrib.memory_usage ArgSpec(args=['program', 'batch_size'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.transpiler.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.transpiler.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.transpiler.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.transpiler.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.transpiler.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True))
+paddle.fluid.transpiler.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode', 'startup_program'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True, None))
 paddle.fluid.transpiler.InferenceTranspiler.__init__ 
 paddle.fluid.transpiler.InferenceTranspiler.transpile ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.transpiler.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0))
@@ -377,7 +390,7 @@ paddle.fluid.LoDTensor.__init__ 1. __init__(self: paddle.fluid.core.LoDTensor, a
 paddle.fluid.LoDTensor.has_valid_recursive_sequence_lengths has_valid_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> bool
 paddle.fluid.LoDTensor.lod lod(self: paddle.fluid.core.LoDTensor) -> List[List[int]]
 paddle.fluid.LoDTensor.recursive_sequence_lengths recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> List[List[int]]
-paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None  2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None  3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None  4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None  5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None  6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None  7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None  8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPlace) -> None  9. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPlace) -> None  10. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPlace) -> None  11. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPlace) -> None  12. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPlace) -> None  13. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPlace) -> None  14. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPlace) -> None  15. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPinnedPlace) -> None  16. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPinnedPlace) -> None  17. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPinnedPlace) -> None  18. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPinnedPlace) -> None  19. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPinnedPlace) -> None  20. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPinnedPlace) -> None  21. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPinnedPlace) -> None
+paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None  2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None  3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None  4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None  5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None  6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None  7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None  8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CPUPlace) -> None  9. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPlace) -> None  10. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPlace) -> None  11. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPlace) -> None  12. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPlace) -> None  13. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPlace) -> None  14. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPlace) -> None  15. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPlace) -> None  16. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPlace) -> None  17. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPinnedPlace) -> None  18. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPinnedPlace) -> None  19. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPinnedPlace) -> None  20. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPinnedPlace) -> None  21. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPinnedPlace) -> None  22. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPinnedPlace) -> None  23. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPinnedPlace) -> None  24. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPinnedPlace) -> None
 paddle.fluid.LoDTensor.set_lod set_lod(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None
 paddle.fluid.LoDTensor.set_recursive_sequence_lengths set_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None
 paddle.fluid.LoDTensor.shape shape(self: paddle.fluid.core.Tensor) -> List[int]

paddle/fluid/framework/CMakeLists.txt

+
 # windows treat symbolic file as a real file, which is different with unix
 # We create a hidden file and compile it instead of origin source file.
 function(windows_symbolic TARGET)
@@ -140,11 +141,11 @@ cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
 cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog)
 
 if(WITH_DISTRIBUTE)
-  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr)
+  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 else()
-  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method)
+  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
 endif()
 
 if (NOT WIN32)

paddle/fluid/framework/data_layout_transform.cc

@@ -46,7 +46,7 @@ struct CastDataLayout {
   const std::vector<int> axis_;
 
   template <typename T>
-  void operator()() {
+  void apply() {
     auto place = ctx_->GetPlace();
 
     if (platform::is_cpu_place(place)) {

paddle/fluid/framework/data_type.cc

@@ -64,6 +64,7 @@ static DataTypeMap* InitDataTypeMap() {
   RegType(size_t, proto::VarType::SIZE_T);
   RegType(int16_t, proto::VarType::INT16);
   RegType(uint8_t, proto::VarType::UINT8);
+  RegType(int8_t, proto::VarType::INT8);
 
 #undef RegType
   return retv;

paddle/fluid/framework/data_type.h

@@ -31,28 +31,31 @@ template <typename Visitor>
 inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
   switch (type) {
     case proto::VarType::FP16:
-      visitor.template operator()<platform::float16>();
+      visitor.template apply<platform::float16>();
       break;
     case proto::VarType::FP32:
-      visitor.template operator()<float>();
+      visitor.template apply<float>();
       break;
     case proto::VarType::FP64:
-      visitor.template operator()<double>();
+      visitor.template apply<double>();
       break;
     case proto::VarType::INT32:
-      visitor.template operator()<int>();
+      visitor.template apply<int>();
       break;
     case proto::VarType::INT64:
-      visitor.template operator()<int64_t>();
+      visitor.template apply<int64_t>();
       break;
     case proto::VarType::BOOL:
-      visitor.template operator()<bool>();
+      visitor.template apply<bool>();
       break;
     case proto::VarType::UINT8:
-      visitor.template operator()<uint8_t>();
+      visitor.template apply<uint8_t>();
       break;
     case proto::VarType::INT16:
-      visitor.template operator()<int16_t>();
+      visitor.template apply<int16_t>();
+      break;
+    case proto::VarType::INT8:
+      visitor.template apply<int8_t>();
       break;
     default:
       PADDLE_THROW("Not supported %d", type);

paddle/fluid/framework/data_type_transform.cc

@@ -37,7 +37,7 @@ struct CastDataType {
   const platform::DeviceContext* ctx_;
 
   template <typename OutType>
-  void apply()() {
+  void apply() {
     auto* in_begin = in_.data<InType>();
     auto* in_end = in_begin + in_.numel();
     auto* out_begin = out_->mutable_data<OutType>(in_.place());

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -625,19 +625,11 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
 void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
     ir::Graph *result, const std::string &loss_grad_name) const {
   for (size_t i = 0; i < places_.size(); ++i) {
-// Insert ScaleCost OpHandle
-#ifdef PADDLE_WITH_CUDA
-    auto *communication_dev_ctx =
-        nccl_ctxs_ ? nccl_ctxs_->DevCtx(places_[i])
-                   : platform::DeviceContextPool::Instance().Get(places_[i]);
-#else
-    auto *communication_dev_ctx =
-        platform::DeviceContextPool::Instance().Get(platform::CPUPlace());
-#endif
+    // Insert ScaleCost OpHandle
+    auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]);
     auto *op_handle = new ScaleLossGradOpHandle(
         result->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation),
-        local_scopes_.size(), local_scopes_[i], places_[i],
-        communication_dev_ctx);
+        local_scopes_.size(), local_scopes_[i], places_[i], dev_ctx);
     result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
 
     // FIXME: Currently ScaleLossGradOp only use device_count as scale
@@ -744,7 +736,7 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
           .emplace(varname, op_dev_id);
     }
   } else {
-    PADDLE_ENFORCE(
+    PADDLE_THROW(
         "the distribute training related op should be in [split_byref, "
         "concat].");
   }
@@ -754,17 +746,26 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
                  node->Op()->Type());
 
   CreateComputationalOp(result, node, op_dev_id);
-  if (node->Op()->Type() == "concat") {
-    ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(),
-              "fetch_barrier");
+}
+
+void SetOpInputsAllPlaces(ir::Graph *result, ir::Node *node, int num_places) {
+  auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
+  for (ir::Node *input : node->inputs) {
+    VarHandle *var = nullptr;
+    for (int place_offset = 0; place_offset < num_places; ++place_offset) {
+      auto &var_holders = result->Get<GraphVars>(kGraphVars)[place_offset];
+      auto &var_holder = var_holders[input->Name()];
+      if (!var_holder.empty()) {
+        var = var_holder.rbegin()->get();
+        op_handle->AddInput(var);
+      }
+    }
   }
 }
 
 // Create RPC related op handles that connects its in ops and out ops.
 void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
                                           ir::Node *node) const {
-  // FIXME(typhoonzero): Cleanup this deps for both sync mode and async mode
-  //                     put them into transpiler.
   int op_dev_id = -1;
   if (node->Op()->Type() == "send") {
     // TODO(paddle-dev): getting the first var is not safe.
@@ -799,8 +800,6 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
     }
     auto recv_param_grad = boost::get<std::vector<std::string>>(
         node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
-    // FIXME(typhoonzero): assume each recv op output one param
-    // Use the same place as send.
     if (recv_param_grad.size() == 2U) {
       op_dev_id = GetVarDeviceID(*result, recv_param_grad[1]);
       VLOG(10) << "recv param " << recv_param_grad[0]
@@ -814,34 +813,44 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
           .emplace(varname, op_dev_id);
     }
   } else {
-    // send_barrier and fetch_barrier op can be scheduled on device 0
+    // send_barrier, fetch_barrier will run on place 0;
     op_dev_id = 0;
   }
 
   PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s",
                  node->Op()->Type());
-
   result->Get<GraphOps>(kGraphOps).emplace_back(new RPCOpHandle(
       result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
       node->Op()->Type(), places_[op_dev_id]));
 
-  // TODO(panyx0718): This might not be needed anymore.
-  if (node->Op()->Type() == "send_barrier") {
-    ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(), "send");
-  } else if (node->Op()->Type() == "recv") {
-    ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(),
-              "send_barrier");
-  } else if (node->Op()->Type() == "fetch_barrier") {
-    ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(), "recv");
-  } else if (node->Op()->Type() == "send") {
-    // do nothing
+  if (node->Op()->Type() == "send") {
+    CreateOpHandleIOs(result, node, op_dev_id);
   } else {
-    PADDLE_THROW(
-        "rpc op should be in ["
-        "send, send_barrier. recv, fetch_barrier]");
-  }
+    // send_barrier, recv, fetch_barrier's inputs are deps var, get them from
+    // all places
+    auto p = places_[op_dev_id];
+    auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
+    op_handle->SetDeviceContext(p,
+                                platform::DeviceContextPool::Instance().Get(p));
 
-  CreateOpHandleIOs(result, node, op_dev_id);
+    SetOpInputsAllPlaces(result, node, places_.size());
+    for (ir::Node *output : node->outputs) {
+      int outvar_dev_id = op_dev_id;
+      if (node->Op()->Type() == "fetch_barrier") {
+        outvar_dev_id = GetVarDeviceID(*result, output->Name());
+        PADDLE_ENFORCE_NE(outvar_dev_id, -1);
+      }
+      p = places_[outvar_dev_id];
+      ir::Node *new_node = nullptr;
+      if (output->Var()) {
+        new_node = result->CreateVarNode(output->Var());
+      } else {
+        new_node =
+            result->CreateEmptyNode(output->Name(), ir::Node::Type::kVariable);
+      }
+      CreateOpOutput(result, op_handle, new_node, p, outvar_dev_id);
+    }
+  }
 }
 
 bool MultiDevSSAGraphBuilder::IsScaleLossOp(ir::Node *node) const {

paddle/fluid/framework/details/reduce_and_gather.h

@@ -31,7 +31,7 @@ struct ReduceLoDTensor {
       : src_tensors_(src), dst_tensor_(*dst) {}
 
   template <typename T>
-  void operator()() const {
+  void apply() const {
     PADDLE_ENFORCE(!src_tensors_.empty());
     auto &t0 = *src_tensors_[0];
     PADDLE_ENFORCE_NE(t0.numel(), 0);

paddle/fluid/framework/executor.h

@@ -60,6 +60,7 @@ class Executor {
   void Run(const ProgramDesc& prog, Scope* scope, int block_id,
            bool create_local_scope = true, bool create_vars = true);
 
+  // This API is very slow.
   void Run(const ProgramDesc& program, Scope* scope,
            std::map<std::string, const LoDTensor*>* feed_targets,
            std::map<std::string, LoDTensor*>* fetch_targets,
@@ -79,6 +80,7 @@ class Executor {
                           bool create_local_scope = true,
                           bool create_vars = true, bool keep_kids = false);
 
+  // This API is very slow.
   void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                           std::map<std::string, const LoDTensor*>* feed_targets,
                           std::map<std::string, LoDTensor*>* fetch_targets,