diff --git a/CMakeLists.txt b/CMakeLists.txt
index a51552d96a462f297f2711f9aaa70f4fc6c8ba00..9ec632e20690eafdc558e24f160270a89b29ee41 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -33,9 +33,7 @@ if(WIN32)
     set(CMAKE_CXX_FLAGS_RELEASE   "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
 endif(WIN32)
 
-if(NOT CMAKE_CROSSCOMPILING)
-    find_package(CUDA QUIET)
-endif(NOT CMAKE_CROSSCOMPILING)
+find_package(CUDA QUIET)
 find_package(Git REQUIRED)
 find_package(Threads REQUIRED)
 
@@ -49,7 +47,6 @@ option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FO
 option(WITH_NGRAPH      "Compile PaddlePaddle with nGraph support."     OFF)
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        OFF)
-option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
 option(WITH_PYTHON      "Compile PaddlePaddle with python interpreter"  ON)
 option(WITH_DOUBLE      "Compile PaddlePaddle with double precision"    OFF)
 option(WITH_RDMA        "Compile PaddlePaddle with RDMA support"        OFF)
@@ -60,11 +57,9 @@ option(WITH_DOC         "Compile PaddlePaddle with documentation"       OFF)
 option(WITH_COVERAGE    "Compile PaddlePaddle with code coverage"       OFF)
 option(COVERALLS_UPLOAD "Package code coverage data to coveralls"       OFF)
 option(ON_TRAVIS        "Exclude special unit test on Travis CI"        OFF)
-option(WITH_C_API       "Compile PaddlePaddle with C-API(Prediction)"   OFF)
 option(WITH_FLUID_ONLY  "Compile PaddlePaddle fluid only"               OFF)
 option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
 option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
-option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
 option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
 option(WITH_PSLIB       "Compile with pslib support"                    OFF)
 option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
@@ -96,37 +91,6 @@ if(NOT CMAKE_BUILD_TYPE)
       FORCE)
 endif()
 
-if(ANDROID OR IOS)
-    if(ANDROID)
-        if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
-            message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
-        endif()
-    endif()
-
-    set(WITH_GPU OFF CACHE STRING
-        "Disable GPU when cross-compiling for Android and iOS" FORCE)
-    set(WITH_AVX OFF CACHE STRING
-        "Disable AVX when cross-compiling for Android and iOS" FORCE)
-    set(WITH_PYTHON OFF CACHE STRING
-        "Disable PYTHON when cross-compiling for Android and iOS" FORCE)
-    set(WITH_RDMA OFF CACHE STRING
-        "Disable RDMA when cross-compiling for Android and iOS" FORCE)
-    set(WITH_MKL OFF CACHE STRING
-        "Disable MKL when cross-compiling for Android and iOS" FORCE)
-    set(WITH_NGRAPH OFF CACHE STRING
-        "Disable nGraph when cross-compiling for Android and iOS" FORCE)
-    set(WITH_GOLANG OFF CACHE STRING
-        "Disable golang when cross-compiling for Android and iOS" FORCE)
-
-    # Compile PaddlePaddle mobile inference library
-    if (NOT WITH_C_API)
-        set(WITH_C_API ON CACHE STRING
-            "Always compile the C_API when cross-compiling for Android and iOS" FORCE)
-    endif()
-    set(MOBILE_INFERENCE ON)
-    add_definitions(-DPADDLE_MOBILE_INFERENCE)
-endif()
-
 if (APPLE)
     set(WITH_MKL OFF CACHE STRING
         "Disable MKL for building on mac" FORCE)
@@ -135,8 +99,6 @@ endif()
 if (WIN32)
     set(WITH_DISTRIBUTE OFF CACHE STRING
             "Disable DISTRIBUTE when compiling for Windows" FORCE)
-    set(WITH_C_API OFF CACHE STRING
-            "Disable C_API when compiling for Windows" FORCE)
     set(WITH_FLUID_ONLY ON CACHE STRING
             "Enable FLUID_ONLY when compiling for Windows" FORCE)
 endif()
@@ -150,21 +112,7 @@ set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING
 set(FLUID_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_inference_install_dir" CACHE STRING
   "A path setting fluid inference shared and static libraries")
 
-if (WITH_C_API AND WITH_PYTHON)
-  message(WARNING "It is suggest not embedded a python interpreter in Paddle "
-    "when using C-API. It will give an unpredictable behavior when using a "
-    "different Python interpreter from compiling.")
-endif()
-
-if (WITH_C_API)
-  set(WITH_FLUID_ONLY OFF CACHE STRING "Disable install fluid when compile the C_API" FORCE)
-endif()
-
-if(MOBILE_INFERENCE)
-    set(THIRD_PARTY_BUILD_TYPE MinSizeRel)
-else()
-    set(THIRD_PARTY_BUILD_TYPE Release)
-endif()
+set(THIRD_PARTY_BUILD_TYPE Release)
 
 set(WITH_MKLML ${WITH_MKL})
 if (NOT DEFINED WITH_MKLDNN)
@@ -193,7 +141,6 @@ include(external/python)    # download, build, install python
 include(external/openblas)  # download, build, install openblas
 include(external/mkldnn)    # download, build, install mkldnn
 include(external/ngraph)    # download, build, install nGraph
-include(external/swig)      # download, build, install swig
 include(external/boost)     # download boost
 include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
@@ -312,11 +259,6 @@ if(WITH_MKLDNN)
     list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
 endif()
 
-if(USE_NNPACK)
-    include(external/nnpack)
-    list(APPEND EXTERNAL_LIBS ${NNPACK_LIBS})
-endif(USE_NNPACK)
-
 set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
 
 set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
diff --git a/Dockerfile.android b/Dockerfile.android
deleted file mode 100644
index 48db2efea21a648657e3f490c95429b9a29ede52..0000000000000000000000000000000000000000
--- a/Dockerfile.android
+++ /dev/null
@@ -1,42 +0,0 @@
-FROM ubuntu:16.04
-MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
-
-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'
-
-# ENV variables
-ARG ANDROID_ABI
-ARG ANDROID_API
-
-ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
-ENV ANDROID_API=${ANDROID_API:-21}
-
-ENV HOME=/root \
-    ANDROID_NDK_HOME=/opt/android-ndk-linux \
-    ANDROID_TOOLCHAINS_DIR=/opt/toolchains
-
-RUN apt-get update && \
-    apt-get install -y \
-    git python-dev python-pip python-numpy \
-    wget curl tar unzip gcc g++ locales clang-format-3.8 swig cmake && \
-    apt-get clean -y
-
-# git credential to skip password typing
-RUN git config --global credential.helper store
-
-# Fix locales to en_US.UTF-8
-RUN localedef -i en_US -f UTF-8 en_US.UTF-8
-
-RUN pip install --upgrade pip==9.0.3 && \
-    pip install -U 'protobuf==3.1.0' && \
-    pip install -U wheel sphinx && \
-    pip install pre-commit
-
-# Android NDK
-RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \
-    mkdir -p /opt/android-ndk-tmp && \
-    cd /opt/android-ndk-tmp && \
-    wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
-    unzip -q android-ndk-r14b-linux-x86_64.zip && \
-    mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
-    rm -rf /opt/android-ndk-tmp
diff --git a/cmake/cblas.cmake b/cmake/cblas.cmake
index 24de8d9d7ced5f8111cc5d65f761b7506bde048e..52ac31d1d125afb89fb0ae783fba94ab9a0c5a1a 100644
--- a/cmake/cblas.cmake
+++ b/cmake/cblas.cmake
@@ -64,24 +64,18 @@ endif()
 ## Then find the reference-cblas.  www.netlib.org/blas/
 set(REFERENCE_CBLAS_ROOT $ENV{REFERENCE_CBLAS_ROOT} CACHE PATH
   "Folder contains reference-cblas")
-if(NOT CMAKE_CROSSCOMPILING)
-  set(REFERENCE_CBLAS_INCLUDE_SEARCH_PATHS
-    ${REFERENCE_CBLAS_ROOT}/include
-    /usr/include
-    /usr/include/cblas
-  )
+set(REFERENCE_CBLAS_INCLUDE_SEARCH_PATHS
+  ${REFERENCE_CBLAS_ROOT}/include
+  /usr/include
+  /usr/include/cblas
+)
 
-  set(REFERENCE_CBLAS_LIB_SEARCH_PATHS
-    ${REFERENCE_CBLAS_ROOT}/lib
-    /usr/lib
-    /usr/lib/blas/reference/
-    /usr/lib/reference/
-  )
-else()
-  # Disable the finding of reference cblas under host's system path
-  set(REFERENCE_CBLAS_INCLUDE_SEARCH_PATHS ${REFERENCE_CBLAS_ROOT}/include)
-  set(REFERENCE_CBLAS_LIB_SEARCH_PATHS ${REFERENCE_CBLAS_ROOT}/lib)
-endif()
+set(REFERENCE_CBLAS_LIB_SEARCH_PATHS
+  ${REFERENCE_CBLAS_ROOT}/lib
+  /usr/lib
+  /usr/lib/blas/reference/
+  /usr/lib/reference/
+)
 
 if(WITH_SYSTEM_BLAS)
   find_path(REFERENCE_CBLAS_INCLUDE_DIR NAMES cblas.h PATHS
@@ -98,10 +92,3 @@ if(WITH_SYSTEM_BLAS)
     message(STATUS "Found reference-cblas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
   endif()
 endif()
-
-if(IOS_USE_VECLIB_FOR_BLAS AND VECLIB_FOUND)
-  set(CBLAS_FOUND ON)
-  set(CBLAS_PROVIDER vecLib)
-  set(CBLAS_INC_DIR ${VECLIB_INC_DIR})
-  add_definitions(-DPADDLE_USE_VECLIB)
-endif()
diff --git a/cmake/configure.cmake b/cmake/configure.cmake
index e3d856fb30d8103f50ebcb6dc16153c8ed2a97a6..076e839120d98d801de4374f2f8338ebd918b88f 100644
--- a/cmake/configure.cmake
+++ b/cmake/configure.cmake
@@ -49,12 +49,10 @@ if(NOT WITH_PROFILER)
     add_definitions(-DPADDLE_DISABLE_PROFILER)
 endif(NOT WITH_PROFILER)
 
-if(NOT CMAKE_CROSSCOMPILING)
-    if(WITH_AVX AND AVX_FOUND)
-        set(SIMD_FLAG ${AVX_FLAG})
-    elseif(SSE3_FOUND)
-        set(SIMD_FLAG ${SSE3_FLAG})
-    endif()
+if(WITH_AVX AND AVX_FOUND)
+    set(SIMD_FLAG ${AVX_FLAG})
+elseif(SSE3_FOUND)
+    set(SIMD_FLAG ${SSE3_FLAG})
 endif()
 
 if(WIN32)
diff --git a/cmake/cross_compiling/android.cmake b/cmake/cross_compiling/android.cmake
deleted file mode 100644
index 4cf2be3bdf07ed018c57cd6bc305a3eda9c9a23d..0000000000000000000000000000000000000000
--- a/cmake/cross_compiling/android.cmake
+++ /dev/null
@@ -1,236 +0,0 @@
-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
-# 
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-# 
-# http://www.apache.org/licenses/LICENSE-2.0
-# 
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# This is a toolchain file for cross-compiling for Android, and the
-# configuration refers to the open-source resposity:
-#     https://github.com/taka-no-me/android-cmake
-# Most of the variables are compatible with that used in
-#     https://developer.android.com/ndk/guides/cmake.html
-# The supported variables are listed belows:
-# 
-# ANDROID_STANDALONE_TOOLCHAIN
-# ANDROID_TOOLCHAIN
-# ANDROID_ABI
-# ANDROID_NATIVE_API_LEVEL
-# ANDROID_ARM_MODE
-# ANDROID_ARM_NEON
-#
-# For CMake >= 3.7.0, all the settings will be delivered to CMake system
-# variables to let CMake do the cross-compiling configurations itself.
-# More detail of cross-compiling settings
-#     https://cmake.org/cmake/help/v3.7/manual/cmake-toolchains.7.html
-
-IF(NOT ANDROID)
-    return()
-ENDIF()
-
-# check the exist of android standalone toolchain
-IF(NOT DEFINED ANDROID_STANDALONE_TOOLCHAIN)
-    SET(ANDROID_STANDALONE_TOOLCHAIN $ENV{ANDROID_STANDALONE_TOOLCHAIN}
-        CACHE PATH "Folder holds the standalone toolchain of Android NDK")
-ENDIF()
-IF(NOT ANDROID_STANDALONE_TOOLCHAIN)
-    MESSAGE(WARNING "It is recommended to set ANDROID_STANDALONE_TOOLCHAIN to "
-            "use a standalone toolchain.\n"
-            "To cross-compile for Android, you need to:\n"
-            "1. Download an Android NDK from"
-            " https://developer.android.com/ndk/downloads/index.html\n"
-            "2. Setup a standalone toolchain"
-            "https://developer.android.google.cn/ndk/guides/standalone_toolchain.html?hl=zh-cn\n")
-ENDIF()
-
-IF(NOT DEFINED CMAKE_SYSTEM_VERSION AND ANDROID_NATIVE_API_LEVEL)
-    IF(ANDROID_NATIVE_API_LEVEL MATCHES "^android-[0-9]+$")
-        STRING(REPLACE "android-" "" CMAKE_SYSTEM_VERSION "${CMAKE_MATCH_0}")
-    ELSEIF(ANDROID_NATIVE_API_LEVEL MATCHES "^[0-9]+$")
-        SET(CMAKE_SYSTEM_VERSION ${ANDROID_NATIVE_API_LEVEL})
-    ENDIF()
-ENDIF()
-
-IF(NOT DEFINED ANDROID_TOOLCHAIN)
-    SET(ANDROID_TOOLCHAIN clang)
-ENDIF()
-
-IF(NOT DEFINED ANDROID_ABI)
-    SET(ANDROID_ABI "armeabi-v7a")
-ENDIF()
-
-IF(NOT DEFINED ANDROID_ARM_MODE)
-    SET(ANDROID_ARM_MODE ON)
-ENDIF()
-IF(ANDROID_ARM_MODE)
-    SET(ANDROID_ARM_MODE_NAME "arm")
-ELSE(ANDROID_ARM_MODE)
-    SET(ANDROID_ARM_MODE_NAME "thumb")
-ENDIF(ANDROID_ARM_MODE)
-
-IF(NOT DEFINED ANDROID_ARM_NEON)
-    SET(ANDROID_ARM_NEON ON)
-ENDIF()
-
-IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
-    IF("${CMAKE_VERSION}" VERSION_LESS "3.1.0")
-        SET(CMAKE_SYSTEM_NAME "Linux")
-    ENDIF()
-    MESSAGE(WARNING "It is recommended to use CMake >= 3.7.0 (current version: "
-            "${CMAKE_VERSION}), when cross-compiling for Android.")
-
-    IF(ANDROID_STANDALONE_TOOLCHAIN)
-        # Use standalone toolchain
-        SET(CMAKE_SYSROOT "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot")
-
-        IF(NOT CMAKE_SYSTEM_VERSION)
-            SET(ANDROID_STANDALONE_TOOLCHAIN_API "")
-            SET(ANDROID_API_LEVEL_H_REGEX "^[\t ]*#[\t ]*define[\t ]+__ANDROID_API__[\t ]+([0-9]+)")
-            FILE(STRINGS "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot/usr/include/android/api-level.h"
-                ANDROID_API_LEVEL_H_CONTENT REGEX "${ANDROID_API_LEVEL_H_REGEX}")
-            IF(ANDROID_API_LEVEL_H_CONTENT MATCHES "${ANDROID_API_LEVEL_H_REGEX}")
-                SET(ANDROID_STANDALONE_TOOLCHAIN_API "${CMAKE_MATCH_1}")
-            ENDIF()
-            SET(CMAKE_SYSTEM_VERSION ${ANDROID_STANDALONE_TOOLCHAIN_API})
-        ENDIF()
-
-        # Toolchain
-        SET(ANDROID_TOOLCHAIN_ROOT ${ANDROID_STANDALONE_TOOLCHAIN})
-    ELSE(ANDROID_NDK)
-        # TODO: use android ndk
-    ENDIF()
-
-    IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
-        SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi)
-        IF(ANDROID_ABI STREQUAL "armeabi")
-            SET(CMAKE_SYSTEM_PROCESSOR armv5te)
-            SET(ANDROID_CLANG_TRIPLE armv5te-none-linux-androideabi)
-        ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
-            SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
-            SET(ANDROID_CLANG_TRIPLE armv7-none-linux-androideabi)
-        ENDIF()
-    ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
-        SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
-        SET(CMAKE_SYSTEM_PROCESSOR aarch64)
-        SET(ANDROID_CLANG_TRIPLE aarch64-none-linux-android)
-    ELSE()
-        MESSAGE(FATAL_ERROR "Invalid Android ABI: ${ANDROID_ABI}.")
-    ENDIF()
-    SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
-
-    IF(ANDROID_TOOLCHAIN STREQUAL clang)
-        SET(ANDROID_C_COMPILER_NAME clang)
-        SET(ANDROID_CXX_COMPILER_NAME clang++)
-        SET(CMAKE_C_COMPILER_TARGET   ${ANDROID_CLANG_TRIPLE})
-        SET(CMAKE_CXX_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE})
-    ELSEIF(ANDROID_TOOLCHAIN STREQUAL gcc)
-        SET(ANDROID_C_COMPILER_NAME gcc)
-        SET(ANDROID_CXX_COMPILER_NAME g++)
-    ELSE()
-        MESSAGE(FATAL_ERROR "Invalid Android toolchain: ${ANDROID_TOOLCHAIN}")
-    ENDIF()
-
-    # C compiler
-    IF(NOT CMAKE_C_COMPILER)
-        SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_C_COMPILER_NAME}")
-    ELSE()
-        GET_FILENAME_COMPONENT(ANDROID_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
-    ENDIF()
-    IF(NOT EXISTS ${ANDROID_C_COMPILER})
-        MESSAGE(FATAL_ERROR "Cannot find C compiler: ${ANDROID_C_COMPILER}")
-    ENDIF()
-
-    # CXX compiler
-    IF(NOT CMAKE_CXX_COMPILER)
-        SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_CXX_COMPILER_NAME}")
-    ELSE()
-        GET_FILENAME_COMPONENT(ANDROID_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
-    ENDIF()
-    IF(NOT EXISTS ${ANDROID_CXX_COMPILER})
-        MESSAGE(FATAL_ERROR "Cannot find CXX compiler: ${ANDROID_CXX_COMPILER}")
-    ENDIF()
-
-    SET(CMAKE_C_COMPILER ${ANDROID_C_COMPILER} CACHE PATH "C compiler" FORCE)
-    SET(CMAKE_CXX_COMPILER ${ANDROID_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
-
-    # Toolchain and ABI specific flags.
-    SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections")
-    SET(ANDROID_LINKER_FLAGS "-Wl,--gc-sections")
-
-    IF(ANDROID_ABI STREQUAL "armeabi")
-        LIST(APPEND ANDROID_COMPILER_FLAGS
-             -march=armv5te
-             -mtune=xscale
-             -msoft-float)
-    ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
-        LIST(APPEND ANDROID_COMPILER_FLAGS
-             -march=armv7-a
-             -mfloat-abi=softfp)
-        IF(ANDROID_ARM_NEON)
-            LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=neon)
-        ELSE()
-            LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=vfpv3-d16)
-        ENDIF()
-        LIST(APPEND ANDROID_LINKER_FLAGS -Wl,--fix-cortex-a8)
-    ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
-        LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
-    ENDIF()
-
-    IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
-        IF(ANDROID_ARM_MODE)
-            LIST(APPEND ANDROID_COMPILER_FLAGS -marm)
-        ELSE()
-            LIST(APPEND ANDROID_COMPILER_FLAGS -mthumb)
-        ENDIF()
-        IF(ANDROID_TOOLCHAIN STREQUAL clang)
-            # Disable integrated-as for better compatibility.
-            LIST(APPEND ANDROID_COMPILER_FLAGS -fno-integrated-as)
-        ENDIF()
-    ENDIF()
-
-    IF(ANDROID_TOOLCHAIN STREQUAL clang)
-        # CMake automatically forwards all compiler flags to the linker,
-        # and clang doesn't like having -Wa flags being used for linking.
-        # To prevent CMake from doing this would require meddling with
-        # the CMAKE_<LANG>_COMPILE_OBJECT rules, which would get quite messy.
-        LIST(APPEND ANDROID_LINKER_FLAGS -Qunused-arguments)
-    ENDIF()
-
-    STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}")
-    STRING(REPLACE ";" " " ANDROID_LINKER_FLAGS "${ANDROID_LINKER_FLAGS}")
-
-    SET(CMAKE_C_FLAGS "${ANDROID_COMPILER_FLAGS} ${CMAKE_C_FLAGS}"
-        CACHE STRING "C flags")
-    SET(CMAKE_CXX_FLAGS "${ANDROID_COMPILER_FLAGS} ${CMAKE_CXX_FLAGS}"
-        CACHE STRING "CXX flags")
-    SET(CMAKE_SHARED_LINKER_FLAGS "${ANDROID_LINKER_FLAGS} ${CMAKE_SHARED_LINKER_FLAGS}"
-        CACHE STRING "shared linker flags")
-
-    SET(CMAKE_POSITION_INDEPENDENT_CODE TRUE)
-    SET(CMAKE_EXE_LINKER_FLAGS "-pie -fPIE ${ANDROID_LINKER_FLAGS} ${CMAKE_EXE_LINKER_FLAGS}"
-        CACHE STRING "executable linker flags")
-
-    MESSAGE(STATUS "Android: Targeting API '${CMAKE_SYSTEM_VERSION}' "
-            "with architecture '${ANDROID_ARM_MODE_NAME}', "
-            "ABI '${ANDROID_ABI}', and processor '${CMAKE_SYSTEM_PROCESSOR}'")
-    MESSAGE(STATUS "System CMAKE_C_FLAGS: " ${CMAKE_C_FLAGS})
-    MESSAGE(STATUS "System CMAKE_CXX_FLAGS: " ${CMAKE_CXX_FLAGS})
-ELSE()
-    IF(ANDROID_STANDALONE_TOOLCHAIN)
-        SET(CMAKE_ANDROID_STANDALONE_TOOLCHAIN ${ANDROID_STANDALONE_TOOLCHAIN})
-    ENDIF()
-    SET(CMAKE_ANDROID_ARCH_ABI ${ANDROID_ABI})
-    IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
-        SET(CMAKE_ANDROID_ARM_MODE ${ANDROID_ARM_MODE})
-        IF(ANDROID_ABI STREQUAL "armeabi-v7a")
-            SET(CMAKE_ANDROID_ARM_NEON ${ANDROID_ARM_NEON})
-        ENDIF()
-    ENDIF()
-ENDIF()
diff --git a/cmake/cross_compiling/host.cmake b/cmake/cross_compiling/host.cmake
deleted file mode 100644
index f9c6b12136f488a9a6ab77b1ba673b6be75391b5..0000000000000000000000000000000000000000
--- a/cmake/cross_compiling/host.cmake
+++ /dev/null
@@ -1,49 +0,0 @@
-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
-# 
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-# 
-# http://www.apache.org/licenses/LICENSE-2.0
-# 
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# find host C compiler
-IF(HOST_C_COMPILER)
-    SET(HOST_C_COMPILER_NAME ${HOST_C_COMPILER})
-ELSEIF(NOT $ENV{CC} STREQUAL "")
-    SET(HOST_C_COMPILER_NAME $ENV{CC})
-ELSE()
-    SET(HOST_C_COMPILER_NAME cc)
-ENDIF()
-
-GET_FILENAME_COMPONENT(HOST_C_COMPILER_PATH ${HOST_C_COMPILER_NAME} PROGRAM)
-IF(NOT HOST_C_COMPILER_PATH OR NOT EXISTS ${HOST_C_COMPILER_PATH})
-    MESSAGE(FATAL_ERROR "Cannot find host C compiler, set host C compiler:\n"
-            "\tcmake .. -DHOST_C_COMPILER=...")
-ENDIF()
-
-# find host CXX compiler
-IF(HOST_CXX_COMPILER)
-    SET(HOST_CXX_COMPILER_NAME ${HOST_CXX_COMPILER})
-ELSEIF(NOT $ENV{CXX} STREQUAL "")
-    SET(HOST_CXX_COMPILER_NAME $ENV{CXX})
-ELSE()
-    SET(HOST_CXX_COMPILER_NAME c++)
-ENDIF()
-
-GET_FILENAME_COMPONENT(HOST_CXX_COMPILER_PATH ${HOST_CXX_COMPILER_NAME} PROGRAM)
-IF(NOT HOST_CXX_COMPILER_PATH OR NOT EXISTS ${HOST_CXX_COMPILER_PATH})
-    MESSAGE(FATAL_ERROR "Cannot find host CXX compiler, set host CXX compiler:\n"
-            "\tcmake .. -DHOST_CXX_COMPILER=...")
-ENDIF()
-
-SET(HOST_C_COMPILER ${HOST_C_COMPILER_PATH} CACHE PATH "Host C compiler")
-SET(HOST_CXX_COMPILER ${HOST_CXX_COMPILER_PATH} CACHE PATH "Host CXX compiler")
-
-MESSAGE(STATUS "Found host C compiler: " ${HOST_C_COMPILER})
-MESSAGE(STATUS "Found host CXX compiler: " ${HOST_CXX_COMPILER})
diff --git a/cmake/cross_compiling/ios.cmake b/cmake/cross_compiling/ios.cmake
deleted file mode 100644
index 10d389ec8ed57ac2b15dd925ef99c8aff4807b05..0000000000000000000000000000000000000000
--- a/cmake/cross_compiling/ios.cmake
+++ /dev/null
@@ -1,347 +0,0 @@
-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# This is a toolchain file for cross-compiling for iOS, and the
-# configuration largely refers to public toolchain file:
-#    https://raw.githubusercontent.com/leetal/ios-cmake/master/ios.toolchain.cmake
-# and
-#    https://github.com/cristeab/ios-cmake
-#
-# Supports options:
-# IOS_PLATFORM = OS (default) or SIMULATOR
-#   This decides if SDKS will be selected from the iPhoneOS.platform or iPhoneSimulator.platform folders
-#   OS - the default, used to build for iPhone and iPad physical devices, which have an arm arch.
-#   SIMULATOR - used to build for the Simulator platforms, which have an x86 arch.
-# IOS_ARCH
-#   The archectures wanted to support, such "arm64", "armv7;arm64"
-# IOS_DEPLOYMENT_TARGET
-#   The minimum iOS deployment version, such as "7.0"
-# IOS_ENABLE_BITCODE = ON (default) or OFF
-# IOS_USE_VECLIB_FOR_BLAS = OFF (default) or ON
-# IOS_DEVELOPER_ROOT = automatic(default) or /path/to/platform/Developer folder
-#   By default this location is automatcially chosen based on the IOS_PLATFORM value above.
-#   If set manually, it will override the default location and force the user of a particular Developer Platform
-# IOS_SDK_ROOT = automatic(default) or /path/to/platform/Developer/SDKs/SDK folder
-#   By default this location is automatcially chosen based on the IOS_DEVELOPER_ROOT value.
-#   In this case it will always be the most up-to-date SDK found in the IOS_DEVELOPER_ROOT path.
-#   If set manually, this will force the use of a specific SDK version
-
-# Macros:
-# set_xcode_property (TARGET XCODE_PROPERTY XCODE_VALUE)
-#  A convenience macro for setting xcode specific properties on targets
-#  example: set_xcode_property (myioslib IPHONEOS_DEPLOYMENT_TARGET "3.1")
-# find_host_package (PROGRAM ARGS)
-#  A macro used to find executable programs on the host system, not within the iOS environment.
-#  Thanks to the android-cmake project for providing the command
-
-if(NOT IOS)
-  return()
-endif()
-
-set(CMAKE_SYSTEM_NAME Darwin)
-
-# Get the Xcode version being used.
-execute_process(COMMAND xcodebuild -version
-                OUTPUT_VARIABLE XCODE_VERSION
-                RESULT_VARIABLE XCODE_VERSION_RESULT
-                ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
-if(NOT ${XCODE_VERSION_RESULT})
-  string(REGEX MATCH "Xcode [0-9\\.]+" XCODE_VERSION "${XCODE_VERSION}")
-  string(REGEX REPLACE "Xcode ([0-9\\.]+)" "\\1" XCODE_VERSION "${XCODE_VERSION}")
-  message(STATUS "Building with Xcode version: ${XCODE_VERSION}")
-else()
-  message(FATAL_ERROR "Cannot execute xcodebuild, please check whether xcode is installed.")
-endif()
-
-# Required as of cmake 2.8.10
-set(CMAKE_OSX_DEPLOYMENT_TARGET "" CACHE STRING "Force unset of the deployment target for iOS" FORCE)
-
-# Setup iOS platform unless specified manually with IOS_PLATFORM
-if(NOT DEFINED IOS_PLATFORM)
-  set(IOS_PLATFORM "OS")
-endif()
-set(IOS_PLATFORM ${IOS_PLATFORM} CACHE STRING "Type of iOS Platform")
-
-# Set the architecture for iOS
-if(NOT DEFINED IOS_ARCH)
-  if(IOS_PLATFORM STREQUAL "OS")
-    set(IOS_ARCH "armv7;armv7s;arm64")
-  elseif(IOS_PLATFORM STREQUAL "SIMULATOR")
-    set(IOS_ARCH "i386;x86_64")
-  endif()
-endif()
-set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string  "Build architecture for iOS")
-
-# Specify minimum iOS deployment version
-if(NOT DEFINED IOS_DEPLOYMENT_TARGET)
-  set(IOS_DEPLOYMENT_TARGET "7.0")
-endif()
-set(IOS_DEPLOYMENT_TARGET ${IOS_DEPLOYMENT_TARGET} CACHE STRING "Minimum iOS version")
-
-# Whether to enable bitcode
-if(NOT DEFINED IOS_ENABLE_BITCODE)
-  set(IOS_ENABLE_BITCODE ON)
-endif()
-set(IOS_ENABLE_BITCODE ${IOS_ENABLE_BITCODE} CACHE BOOL "Whether to enable bitcode")
-
-if(NOT DEFINED IOS_USE_VECLIB_FOR_BLAS)
-  set(IOS_USE_VECLIB_FOR_BLAS OFF)
-endif()
-set(IOS_USE_VECLIB_FOR_BLAS ${IOS_UES_VECLIB_FOR_BLAS} CACHE BOOL "Whether to use veclib")
-
-# Check the platform selection and setup for developer root
-if(${IOS_PLATFORM} STREQUAL "OS")
-  set(IOS_PLATFORM_LOCATION "iPhoneOS.platform")
-  set(XCODE_IOS_PLATFORM iphoneos)
-
-  # This causes the installers to properly locate the output libraries
-  set(CMAKE_XCODE_EFFECTIVE_PLATFORMS "-iphoneos")
-elseif(${IOS_PLATFORM} STREQUAL "SIMULATOR")
-  set(IOS_PLATFORM_LOCATION "iPhoneSimulator.platform")
-  set(XCODE_IOS_PLATFORM iphonesimulator)
-
-  # This causes the installers to properly locate the output libraries
-  set(CMAKE_XCODE_EFFECTIVE_PLATFORMS "-iphonesimulator")
-elseif(${IOS_PLATFORM} STREQUAL "WATCHOS")
-  set(IOS_PLATFORM_LOCATION "WatchOS.platform")
-  set(XCODE_IOS_PLATFORM watchos)
-
-  # This causes the installers to properly locate the output libraries
-  set(CMAKE_XCODE_EFFECTIVE_PLATFORMS "-watchos")
-else(${IOS_PLATFORM} STREQUAL "OS")
-  message(FATAL_ERROR "Unsupported IOS_PLATFORM value selected. Please set to\n"
-          "\t OS, SIMULATOR, or WATCHOS.")
-endif()
-
-# Check iOS developer toolchain
-if(NOT DEFINED IOS_DEVELOPER_ROOT)
-  # Setup iOS developer location
-  execute_process(COMMAND xcode-select -print-path
-                  OUTPUT_VARIABLE XCODE_DEVELOPER_DIR
-                  RESULT_VARIABLE XCODE_DEVELOPER_DIR_RESULT
-                  ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
-  # Xcode 4.3 changed the installation location, choose the most recent one available
-  if(${XCODE_VERSION} VERSION_LESS "4.3.0")
-    set(IOS_DEVELOPER_ROOT "/Developer/Platforms/${IOS_PLATFORM_LOCATION}/Developer")
-  else()
-    set(IOS_DEVELOPER_ROOT "${XCODE_DEVELOPER_DIR}/Platforms/${IOS_PLATFORM_LOCATION}/Developer")
-  endif()
-endif()
-if(EXISTS ${IOS_DEVELOPER_ROOT})
-  set(IOS_DEVELOPER_ROOT ${IOS_DEVELOPER_ROOT} CACHE PATH "Location of iOS Platform")
-else()
-  message(FATAL_ERROR "Invalid IOS_DEVELOPER_ROOT: ${IOS_DEVELOPER_ROOT} does not exist.")
-endif()
-
-# Check iOS SDK
-if(NOT DEFINED IOS_SDK_ROOT)
-  # Find and use the most recent iOS sdk
-  file(GLOB IOS_SDK_LISTS "${IOS_DEVELOPER_ROOT}/SDKs/*")
-  if(IOS_SDK_LISTS)
-    list(SORT IOS_SDK_LISTS)
-    list(REVERSE IOS_SDK_LISTS)
-    list(GET IOS_SDK_LISTS 0 IOS_SDK_ROOT)
-  else(IOS_SDK_LISTS)
-    message(FATAL_ERROR "No iOS SDK's found in default search path ${IOS_DEVELOPER_ROOT}."
-            " Please manually set IOS_SDK_ROOT or install the iOS SDK.")
-  endif(IOS_SDK_LISTS)
-endif()
-if(EXISTS ${IOS_SDK_ROOT})
-  set(IOS_SDK_ROOT ${IOS_SDK_ROOT} CACHE PATH "Location of the selected iOS SDK")
-  message(STATUS "iOS toolchain: ${IOS_SDK_ROOT}")
-else()
-  message(FATAL_ERROR "Invalid IOS_SDK_ROOT: ${IOS_SDK_ROOT} does not exist.")
-endif()
-
-# Set the sysroot default to the most recent SDK
-set(CMAKE_OSX_SYSROOT ${IOS_SDK_ROOT} CACHE PATH "Sysroot used for iOS support")
-
-# Get version of iOS SDK
-execute_process(COMMAND xcodebuild -sdk ${CMAKE_OSX_SYSROOT} -version SDKVersion
-                OUTPUT_VARIABLE IOS_SDK_VERSION
-                RESULT_VARIABLE IOS_SDK_VERSION_RESULT
-                ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
-if(${IOS_SDK_VERSION_RESULT})
-  string(REGEX MATCH "(([0-9]+)\\.)+([0-9]+)" IOS_SDK_VERSION "${IOS_SDK_ROOT}")
-endif()
-if(NOT IOS_SDK_VERSION)
-  message(WARNING "Cannot get SDK's version.")
-  set(IOS_SDK_VERSION 1)
-endif()
-set(CMAKE_SYSTEM_VERSION ${IOS_SDK_VERSION})
-
-# Find the C & C++ compilers for the specified SDK.
-if(NOT CMAKE_C_COMPILER)
-  # Default to use clang
-  execute_process(COMMAND xcrun -sdk ${CMAKE_OSX_SYSROOT} -find clang
-                  OUTPUT_VARIABLE IOS_C_COMPILER
-                  RESULT_VARIABLE IOS_C_COMPILER_RESULT
-                  ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
-  if(${IOS_C_COMPILER_RESULT})
-    get_filename_component(IOS_C_COMPILER clang PROGRAM)
-  endif()
-else(NOT CMAKE_C_COMPILER)
-  # User can set it in cmake command
-  get_filename_component(IOS_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
-endif(NOT CMAKE_C_COMPILER)
-if(NOT EXISTS ${IOS_C_COMPILER})
-  message(FATAL_ERROR "Cannot find C compiler: ${IOS_C_COMPILER}")
-endif()
-
-if(NOT CMAKE_CXX_COMPILER)
-  # Default to use clang++
-  execute_process(COMMAND xcrun -sdk ${CMAKE_OSX_SYSROOT} -find clang++
-                  OUTPUT_VARIABLE IOS_CXX_COMPILER
-                  RESULT_VARIABLE IOS_CXX_COMPILER_RESULT
-                  ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
-  if(${IOS_CXX_COMPILER_RESULT})
-    get_filename_component(IOS_CXX_COMPILER clang++ PROGRAM)
-  endif()
-else(NOT CMAKE_CXX_COMPILER)
-  # User can set it in cmake command
-  get_filename_component(IOS_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
-endif(NOT CMAKE_CXX_COMPILER)
-if(NOT EXISTS ${IOS_CXX_COMPILER})
-  message(FATAL_ERROR "Cannot find CXX compiler: ${IOS_CXX_COMPILER}")
-endif()
-
-set(CMAKE_C_COMPILER ${IOS_C_COMPILER} CACHE PATH "C compiler" FORCE)
-set(CMAKE_CXX_COMPILER ${IOS_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
-
-set(CMAKE_C_OSX_COMPATIBILITY_VERSION_FLAG "-compatibility_version ")
-set(CMAKE_C_OSX_CURRENT_VERSION_FLAG "-current_version ")
-set(CMAKE_CXX_OSX_COMPATIBILITY_VERSION_FLAG "${CMAKE_C_OSX_COMPATIBILITY_VERSION_FLAG}")
-set(CMAKE_CXX_OSX_CURRENT_VERSION_FLAG "${CMAKE_C_OSX_CURRENT_VERSION_FLAG}")
-
-# Set iOS specific C/C++ flags
-if(IOS_PLATFORM STREQUAL "OS")
-  if(XCODE_VERSION VERSION_LESS "7.0")
-    set(XCODE_IOS_PLATFORM_VERSION_FLAGS "-mios-version-min=${IOS_DEPLOYMENT_TARGET}")
-  else()
-    # Xcode 7.0+ uses flags we can build directly from XCODE_IOS_PLATFORM.
-    set(XCODE_IOS_PLATFORM_VERSION_FLAGS "-m${XCODE_IOS_PLATFORM}-version-min=${IOS_DEPLOYMENT_TARGET}")
-  endif()
-else()
-  set(XCODE_IOS_FLATFORM_VERSION_FLAGS "-mios-simulator-version-min=${IOS_DEPLOYMENT_TARGET}")
-endif()
-
-if(IOS_ENABLE_BITCODE)
-  set(XCODE_IOS_BITCODE_FLAGS "${IOS_COMPILER_FLAGS} -fembed-bitcode")
-else()
-  set(XCODE_IOS_BITCODE_FLAGS "")
-endif()
-
-set(IOS_COMPILER_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${XCODE_IOS_BITCODE_FLAGS}")
-
-# Hidden visibilty is required for cxx on iOS 
-set(CMAKE_C_FLAGS "${IOS_COMPILER_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags")
-set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility=hidden -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
-
-set(IOS_LINK_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} -Wl,-search_paths_first")
-
-if(IOS_USE_VECLIB_FOR_BLAS)
-  # Find vecLib for iOS
-  set(VECLIB_SEARCH_DIRS
-      ${IOS_SDK_ROOT}/System/Library/Frameworks/Accelerate.framework/Versions/Current/Frameworks
-      ${IOS_SDK_ROOT}/System/Library/Frameworks/Accelerate.framework/Frameworks
-      )
-  find_path(VECLIB_INC_DIR vecLib.h PATHS ${VECLIB_SEARCH_DIRS}/vecLib.framework/Headers)
-
-  include(FindPackageHandleStandardArgs)
-  find_package_handle_standard_args(vecLib DEFAULT_MSG VECLIB_INC_DIR)
-
-  if(VECLIB_FOUND)
-    if(VECLIB_INC_DIR MATCHES "^/System/Library/Frameworks/vecLib.framework.*")
-      set(IOS_LINK_FLAGS ${IOS_LINK_FLAGS} -lcblas "-framework vecLib")
-      message(STATUS "Found standalone vecLib.framework")
-    else()
-      set(IOS_LINK_FLAGS ${IOS_LINK_FLAGS} -lcblas "-framework Accelerate")
-      message(STATUS "Found vecLib as part of Accelerate.framework")
-    endif()
-
-  endif()
-endif()
-
-set(CMAKE_C_LINK_FLAGS "${IOS_LINK_FLAGS} ${CMAKE_C_LINK_FLAGS}")
-set(CMAKE_CXX_LINK_FLAGS "${IOS_LINK_FLAGS} ${CMAKE_CXX_LINK_FLAGS}")
-
-set(CMAKE_PLATFORM_HAS_INSTALLNAME 1)
-if(NOT IOS_ENABLE_BITCODE)
-  set(CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS "-dynamiclib -headerpad_max_install_names")
-  set(CMAKE_SHARED_MODULE_CREATE_C_FLAGS "-bundle -headerpad_max_install_names")
-else()
-  set(CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS "-dynamiclib")
-  set(CMAKE_SHARED_MODULE_CREATE_C_FLAGS "-bundle")
-endif()
-set(CMAKE_SHARED_MODULE_LOADER_C_FLAG "-Wl,-bundle_loader,")
-set(CMAKE_SHARED_MODULE_LOADER_CXX_FLAG "-Wl,-bundle_loader,")
-set(CMAKE_FIND_LIBRARY_SUFFIXES ".dylib" ".so" ".a")
-
-# hack: if a new cmake (which uses CMAKE_INSTALL_NAME_TOOL) runs on an old build tree
-# (where install_name_tool was hardcoded) and where CMAKE_INSTALL_NAME_TOOL isn't in the cache
-# and still cmake didn't fail in CMakeFindBinUtils.cmake (because it isn't rerun)
-# hardcode CMAKE_INSTALL_NAME_TOOL here to install_name_tool, so it behaves as it did before, Alex
-if(NOT DEFINED CMAKE_INSTALL_NAME_TOOL)
-  find_program(CMAKE_INSTALL_NAME_TOOL install_name_tool)
-endif()
-
-# Set the find root to the iOS developer roots and to user defined paths
-set(CMAKE_FIND_ROOT_PATH ${IOS_DEVELOPER_ROOT} ${IOS_SDK_ROOT} ${CMAKE_PREFIX_PATH}
-    CACHE string  "iOS find search path root")
-
-# default to searching for frameworks first
-set(CMAKE_FIND_FRAMEWORK FIRST)
-
-# set up the default search directories for frameworks
-set(CMAKE_SYSTEM_FRAMEWORK_PATH
-    ${IOS_SDK_ROOT}/System/Library/Frameworks
-    ${IOS_SDK_ROOT}/System/Library/PrivateFrameworks
-    ${IOS_SDK_ROOT}/Developer/Library/Frameworks
-    )
-
-# only search the iOS sdks, not the remainder of the host filesystem
-set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
-set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
-set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
-
-message(STATUS "iOS: Targeting iOS '${CMAKE_SYSTEM_VERSION}', "
-        "building for '${IOS_PLATFORM}' platform, with architecture '${CMAKE_OSX_ARCHITECTURES}'")
-message(STATUS "System CMAKE_C_FLAGS: ${CMAKE_C_FLAGS}")
-message(STATUS "System CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
-
-# Used in ExternalProject command
-string(REPLACE ";" "\\$<SEMICOLON>" EXTERNAL_IOS_ARCHITECTURES "${CMAKE_OSX_ARCHITECTURES}")
-set(EXTERNAL_OPTIONAL_ARGS
-    -DCMAKE_OSX_SYSROOT=${CMAKE_OSX_SYSROOT}
-    -DCMAKE_OSX_ARCHITECTURES=${EXTERNAL_IOS_ARCHITECTURES})
-
-# This little macro lets you set any XCode specific property
-macro(set_xcode_property TARGET XCODE_PROPERTY XCODE_VALUE)
-  set_property (TARGET ${TARGET} PROPERTY XCODE_ATTRIBUTE_${XCODE_PROPERTY} ${XCODE_VALUE})
-endmacro(set_xcode_property)
-
-# This macro lets you find executable programs on the host system
-macro(find_host_package)
-  set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
-  set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY NEVER)
-  set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE NEVER)
-  set(IOS FALSE)
-
-  find_package(${ARGN})
-
-  set(IOS TRUE)
-  set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM ONLY)
-  set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
-  set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
-endmacro(find_host_package)
diff --git a/cmake/cross_compiling/raspberry_pi.cmake b/cmake/cross_compiling/raspberry_pi.cmake
deleted file mode 100644
index 0425b2ae158b265fd6f8423b05190a8002f03f20..0000000000000000000000000000000000000000
--- a/cmake/cross_compiling/raspberry_pi.cmake
+++ /dev/null
@@ -1,84 +0,0 @@
-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
-# 
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-# 
-# http://www.apache.org/licenses/LICENSE-2.0
-# 
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# This is a toolchain file for cross-compiling for Raspberry Pi.
-# The supported variables are listed belows:
-#
-# RPI_TOOLCHAIN
-# RPI_ARM_NEON
-#
-# Also you can set CMAKE_C/CXX_COMPILER yourself, through cmake arguments.
-
-IF(NOT RPI)
-    return()
-ENDIF()
- 
-SET(CMAKE_SYSTEM_NAME Linux)
-SET(CMAKE_SYSTEM_VERSION 1)
-SET(CMAKE_SYSTEM_PROCESSOR arm)
-
-# check the exist of raspberry pi toolchain
-IF(NOT DEFINED RPI_TOOLCHAIN)
-    SET(RPI_TOOLCHAIN $ENV{RPI_TOOLCHAIN}
-        CACHE PATH "Folder holds the toolchain of Raspberr Pi")
-ENDIF()
-IF(NOT RPI_TOOLCHAIN)
-    MESSAGE(WARNING "It is recommended to set RPI_TOOLCHAIN to use toolchain.\n"
-            "To cross-compile for Raspberry Pi, you need to download the tools using:\n"
-            " git clone https://github.com/raspberrypi/tools\n")
-ENDIF()
-
-IF(NOT DEFINED RPI_ARM_NEON)
-    SET(RPI_ARM_NEON ON)
-ENDIF()
-
-IF(RPI_TOOLCHAIN)
-    SET(RPI_TOOLCHAIN_ROOT ${RPI_TOOLCHAIN})
-    IF(RPI_TOOLCHAIN_ROOT MATCHES "gcc-linaro-arm-linux-gnueabihf-raspbian(-x64)?$")
-        # gcc-linaro-arm-linux-gnueabihf-raspbian
-        # gcc-linaro-arm-linux-gnueabihf-raspbian-x64
-        SET(RPI_TOOLCHAIN_NAME arm-linux-gnueabihf)
-    ENDIF()
-    SET(RPI_TOOLCHAIN_PREFIX "${RPI_TOOLCHAIN_ROOT}/bin/${RPI_TOOLCHAIN_NAME}-")
-ENDIF()
-
-# C compiler
-IF(NOT CMAKE_C_COMPILER)
-    SET(RPI_C_COMPILER "${RPI_TOOLCHAIN_PREFIX}gcc")
-ELSE()
-    GET_FILENAME_COMPONENT(RPI_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
-ENDIF()
-IF(NOT EXISTS ${RPI_C_COMPILER})
-    MESSAGE(FATAL_ERROR "Cannot find C compiler: ${RPI_C_COMPILER}")
-ENDIF()
-
-# CXX compiler
-IF(NOT CMAKE_CXX_COMPILER)
-    SET(RPI_CXX_COMPILER "${RPI_TOOLCHAIN_PREFIX}g++")
-ELSE()
-    GET_FILENAME_COMPONENT(RPI_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
-ENDIF()
-IF(NOT EXISTS ${RPI_CXX_COMPILER})
-    MESSAGE(FATAL_ERROR "Cannot find CXX compiler: ${RPI_CXX_COMPILER}")
-ENDIF()
-
-SET(CMAKE_C_COMPILER ${RPI_C_COMPILER} CACHE PATH "C compiler" FORCE)
-SET(CMAKE_CXX_COMPILER ${RPI_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
-
-IF(RPI_ARM_NEON)
-    SET(RPI_C_FLAGS "${RPI_C_FLAGS} -mfpu=neon")
-ENDIF()
-
-SET(CMAKE_C_FLAGS "${RPI_C_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags")
-SET(CMAKE_CXX_FLAGS "${RPI_C_FLAGS} ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
diff --git a/cmake/cuda.cmake b/cmake/cuda.cmake
index 16432ce2b803f6d21bbf47200eda5404269b750f..ea46f6418edf1db70b2a308dd49cf2131cc89d3b 100644
--- a/cmake/cuda.cmake
+++ b/cmake/cuda.cmake
@@ -63,9 +63,7 @@ function(select_nvcc_arch_flags out_variable)
   # List of arch names
   set(archs_names "Kepler" "Maxwell" "Pascal" "Volta" "Turing" "All" "Manual")
   set(archs_name_default "All")
-  if(NOT CMAKE_CROSSCOMPILING)
-    list(APPEND archs_names "Auto")
-  endif()
+  list(APPEND archs_names "Auto")
 
   # set CUDA_ARCH_NAME strings (so it will be seen as dropbox in CMake-Gui)
   set(CUDA_ARCH_NAME ${archs_name_default} CACHE STRING "Select target NVIDIA GPU achitecture.")
diff --git a/cmake/external/cares.cmake b/cmake/external/cares.cmake
index a743b572a6c3f6f152d85909500e9dbb35c72a01..52507a6ae4aabe300cf8bf88d0946c45a2c0e79c 100644
--- a/cmake/external/cares.cmake
+++ b/cmake/external/cares.cmake
@@ -13,7 +13,7 @@
 # limitations under the License.
 #
 
-IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE)
+IF(NOT WITH_DISTRIBUTE)
     return()
 ENDIF()
 
diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake
index 95ca16f57f2704eaded85aa5f5c0546310fba3a7..f3ca74faea3629ddce053c49ef1e629f230fdc49 100644
--- a/cmake/external/gflags.cmake
+++ b/cmake/external/gflags.cmake
@@ -71,13 +71,3 @@ if (WIN32)
     set_property(GLOBAL PROPERTY OS_DEPENDENCY_MODULES shlwapi.lib)
   endif(HAVE_SHLWAPI)
 endif (WIN32)
-
-IF(WITH_C_API)
-  INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
-  IF(ANDROID)
-    INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib/${ANDROID_ABI})
-  ELSE()
-    INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib)
-  ENDIF()
-ENDIF()
-
diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake
index 8cd0455c16bf84909b735102e7fb1089744c4245..7a6a4523886824a67c82f9ce978de025ddb9c2cd 100644
--- a/cmake/external/glog.cmake
+++ b/cmake/external/glog.cmake
@@ -26,14 +26,8 @@ ENDIF(WIN32)
 
 INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR})
 
-IF(ANDROID AND ${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
-  # Using the unofficial glog for Android API < 21
-  SET(GLOG_REPOSITORY "https://github.com/Xreki/glog.git")
-  SET(GLOG_TAG "8a547150548b284382ccb6582408e9140ff2bea8")
-ELSE()
-  SET(GLOG_REPOSITORY "https://github.com/google/glog.git")
-  SET(GLOG_TAG "v0.3.5")
-ENDIF()
+SET(GLOG_REPOSITORY "https://github.com/google/glog.git")
+SET(GLOG_TAG "v0.3.5")
 
 ExternalProject_Add(
     extern_glog
@@ -78,12 +72,3 @@ ADD_DEPENDENCIES(glog extern_glog gflags)
 LINK_LIBRARIES(glog gflags)
 
 LIST(APPEND external_project_dependencies glog)
-
-IF(WITH_C_API)
-  INSTALL(DIRECTORY ${GLOG_INCLUDE_DIR} DESTINATION third_party/glog)
-  IF(ANDROID)
-    INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib/${ANDROID_ABI})
-  ELSE()
-    INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib)
-  ENDIF()
-ENDIF()
diff --git a/cmake/external/grpc.cmake b/cmake/external/grpc.cmake
index fd9835d023c67b76579913f2ec56c2444fea8c15..c5754da59bf2053931be413eb10c481adecbae6b 100644
--- a/cmake/external/grpc.cmake
+++ b/cmake/external/grpc.cmake
@@ -13,7 +13,7 @@
 # limitations under the License.
 #
 
-IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE)
+IF(NOT WITH_DISTRIBUTE)
     return()
 ENDIF()
 
diff --git a/cmake/external/gzstream.cmake b/cmake/external/gzstream.cmake
index 3e36ef7ae205bbf85f345d55456309cc05a58fbd..af7a8bfda6f7db12049203e7c9d54885884d8cf2 100644
--- a/cmake/external/gzstream.cmake
+++ b/cmake/external/gzstream.cmake
@@ -13,10 +13,6 @@
 # limitations under the License.
 #
 
-IF(MOBILE_INFERENCE)
-    return()
-ENDIF()
-
 include (ExternalProject)
 
 # NOTE: gzstream is needed when linking with ctr reader.
diff --git a/cmake/external/libxsmm.cmake b/cmake/external/libxsmm.cmake
index 530f7ebe2813fb2f00c6b5b4d1f7b2f04fe650b0..39f49d210a20d49a06c120361ecf0a5d07d1af28 100644
--- a/cmake/external/libxsmm.cmake
+++ b/cmake/external/libxsmm.cmake
@@ -19,8 +19,8 @@ IF(NOT WITH_LIBXSMM)
     return()
 ENDIF()
 
-IF(WIN32 OR APPLE OR ANDROID OR IOS)
-    MESSAGE(WARNING "Windows, Mac or Mobile are not supported with libxsmm in Paddle yet.")
+IF(WIN32 OR APPLE)
+    MESSAGE(WARNING "Windows, Mac are not supported with libxsmm in Paddle yet.")
     SET(WITH_LIBXSMM OFF CACHE STRING "Disable LIBXSMM" FORCE)
     return()
 ENDIF()
diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake
index 03f0dee85911bdaa0312b624114b7f4aef1fb723..6a7be73f09a278ab0fd29c7599a7781df3d29413 100644
--- a/cmake/external/mkldnn.cmake
+++ b/cmake/external/mkldnn.cmake
@@ -110,7 +110,3 @@ else(WIN32)
 endif(WIN32)
 ADD_CUSTOM_TARGET(mkldnn_shared_lib ALL DEPENDS ${MKLDNN_SHARED_LIB})
 ADD_DEPENDENCIES(mkldnn_shared_lib ${MKLDNN_PROJECT} mkldnn)
-IF(WITH_C_API)
-  INSTALL(FILES ${MKLDNN_SHARED_LIB} DESTINATION lib)
-ENDIF()
-
diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake
index 43322a257a02c3fd756078db6fe20b582826066a..2caff27357687018f29c1efc55b7b82c9dc3ccf6 100644
--- a/cmake/external/mklml.cmake
+++ b/cmake/external/mklml.cmake
@@ -74,7 +74,3 @@ ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
 ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
 LIST(APPEND external_project_dependencies mklml)
-
-IF(WITH_C_API)
-  INSTALL(FILES ${MKLML_LIB} ${MKLML_IOMP_LIB} DESTINATION lib)
-ENDIF()
diff --git a/cmake/external/nnpack.cmake b/cmake/external/nnpack.cmake
deleted file mode 100644
index d42bcb0f329041462bd8b568052fbb8226d18e4e..0000000000000000000000000000000000000000
--- a/cmake/external/nnpack.cmake
+++ /dev/null
@@ -1,30 +0,0 @@
-# Find the NNPACK library
-#  NNPACK_ROOT - where to find NNPACK include and library.
-#
-
-set(NNPACK_FOUND OFF)
-set(NNPACK_ROOT $ENV{NNPACK_ROOT} CACHE PATH "Folder contains NNPACK")
-find_path(NNPACK_INC_DIR nnpack.h PATHS ${NNPACK_ROOT}/include)
-find_library(NNPACK_LIB NAMES nnpack PATHS ${NNPACK_ROOT}/lib)
-find_library(PTHREADPOOL_LIB NAMES pthreadpool PATHS ${NNPACK_ROOT}/lib)
-find_library(NNPACK_UKERNELS_LIB NAMES nnpack_ukernels PATHS ${NNPACK_ROOT}/lib)
-find_library(NNPACK_CPUFEATURES_LIB NAMES cpufeatures PATHS ${NNPACK_ROOT}/lib)
-
-if(NNPACK_INC_DIR AND NNPACK_LIB AND PTHREADPOOL_LIB)
-  set(NNPACK_FOUND ON)
-  INCLUDE_DIRECTORIES(${NNPACK_INC_DIR})
-
-  set(NNPACK_LIBS)
-  list(APPEND NNPACK_LIBS ${NNPACK_LIB} ${PTHREADPOOL_LIB})
-  if (NNPACK_UKERNELS_LIB)
-    list(APPEND NNPACK_LIBS ${NNPACK_UKERNELS_LIB})
-  endif()
-  if (NNPACK_CPUFEATURES_LIB)
-    list(APPEND NNPACK_LIBS ${NNPACK_CPUFEATURES_LIB})
-  endif()
-  if(NOT ANDROID)
-    list(APPEND NNPACK_LIBS "rt")
-  endif()
-else()
-  message(FATAL_ERROR "Cannot find NNPACK in (${NNPACK_ROOT})")
-endif()
diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake
index aeb976b840e999a20e8cab11939cbb1f49a27850..b347a592929836a473ac764c0af1153b07d54258 100644
--- a/cmake/external/openblas.cmake
+++ b/cmake/external/openblas.cmake
@@ -40,38 +40,12 @@ IF(NOT ${CBLAS_FOUND})
     SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
     SET(OPENBLAS_COMMIT "v0.2.20")
 
-    IF(CMAKE_CROSSCOMPILING)
-        SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER})
-        GET_FILENAME_COMPONENT(CROSS_SUFFIX ${CMAKE_C_COMPILER} DIRECTORY)
-        SET(CROSS_SUFFIX ${CROSS_SUFFIX}/)
-        IF(ANDROID)
-            IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
-                # use softfp
-                SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0)
-            ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
-                SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0)
-            ENDIF()
-        ELSEIF(IOS)
-            IF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
-                SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}")
-                SET(OPENBLAS_CC "${OPENBLAS_CC} -arch arm64")
-                SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0 CROSS_SUFFIX=${CROSS_SUFFIX})
-            ELSE()
-                MESSAGE(FATAL_ERROR "OpenBLAS only support arm64 architectures on iOS. "
-                       "You can set IOS_USE_VECLIB_FOR_BLAS=ON or USE_EIGEN_FOR_BLAS=ON to use other blas library instead.")
-            ENDIF()
-        ELSEIF(RPI)
-            # use hardfp
-            SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 USE_THREAD=0)
-        ENDIF()
-    ELSE()
-        IF(APPLE)
-            SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
-        ENDIF()
-        SET(OPTIONAL_ARGS "")
-        IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$")
-            SET(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64)
-        ENDIF()
+    IF(APPLE)
+        SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
+    ENDIF()
+    SET(OPTIONAL_ARGS "")
+    IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$")
+        SET(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64)
     ENDIF()
 
     SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
@@ -92,25 +66,6 @@ IF(NOT ${CBLAS_FOUND})
     ELSE()
     ENDIF(NOT WIN32)
     SET(CBLAS_PROVIDER openblas)
-    IF(WITH_C_API)
-        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
-        # Because libopenblas.a is a symbolic link of another library, thus need to
-        # install the whole directory.
-        IF(ANDROID)
-            SET(TMP_INSTALL_DIR third_party/openblas/lib/${ANDROID_ABI})
-        ELSE()
-            SET(TMP_INSTALL_DIR third_party/openblas/lib)
-        ENDIF()
-        INSTALL(CODE "execute_process(
-            COMMAND ${CMAKE_COMMAND} -E copy_directory ${CBLAS_INSTALL_DIR}/lib
-                    ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
-            )"
-        )
-        INSTALL(CODE "MESSAGE(STATUS \"Installing: \"
-                \"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\"
-            )"
-        )
-    ENDIF()
 ENDIF(NOT ${CBLAS_FOUND})
 
 MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake
index e1e619e572b05e83fbe751af2e5391aafc494416..e05b7694ddf1e1652b00f156cde1a2d433c9fc46 100644
--- a/cmake/external/protobuf.cmake
+++ b/cmake/external/protobuf.cmake
@@ -204,15 +204,6 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
 
     SET(PROTOBUF_REPO "https://github.com/google/protobuf.git")
     SET(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546")
-    IF(MOBILE_INFERENCE)
-        # The reason why the official version is not used is described in
-        # https://github.com/PaddlePaddle/Paddle/issues/6114
-        SET(PROTOBUF_REPO "https://github.com/qingqing01/protobuf.git")
-        SET(PROTOBUF_TAG "v3.2.0")
-        IF(NOT BUILD_FOR_HOST)
-            SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} "-Dprotobuf_BUILD_PROTOC_BINARIES=OFF")
-        ENDIF()
-    ENDIF()
 
     ExternalProject_Add(
         ${TARGET_NAME}
@@ -240,19 +231,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
     )
 ENDFUNCTION()
 
-IF(NOT MOBILE_INFERENCE)
-    SET(PROTOBUF_VERSION 3.1)
-ELSE()
-    SET(PROTOBUF_VERSION 3.2)
-ENDIF()
-IF(CMAKE_CROSSCOMPILING)
-    build_protobuf(protobuf_host TRUE)
-    LIST(APPEND external_project_dependencies protobuf_host)
-
-    SET(PROTOBUF_PROTOC_EXECUTABLE ${protobuf_host_PROTOC_EXECUTABLE}
-        CACHE FILEPATH "protobuf executable." FORCE)
-ENDIF()
-
+SET(PROTOBUF_VERSION 3.1)
 
 IF(NOT PROTOBUF_FOUND)
     build_protobuf(extern_protobuf FALSE)
@@ -266,20 +245,7 @@ IF(NOT PROTOBUF_FOUND)
     SET(PROTOBUF_PROTOC_LIBRARY ${extern_protobuf_PROTOC_LIBRARY}
         CACHE FILEPATH "protoc library." FORCE)
 
-    IF(WITH_C_API)
-        INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
-        IF(ANDROID)
-            INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
-        ELSE()
-            INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib)
-        ENDIF()
-    ENDIF()
-
-    IF(CMAKE_CROSSCOMPILING)
-        PROMPT_PROTOBUF_LIB(protobuf_host extern_protobuf)
-    ELSE()
-        SET(PROTOBUF_PROTOC_EXECUTABLE ${extern_protobuf_PROTOC_EXECUTABLE}
-            CACHE FILEPATH "protobuf executable." FORCE)
-        PROMPT_PROTOBUF_LIB(extern_protobuf)
-    ENDIF()
+    SET(PROTOBUF_PROTOC_EXECUTABLE ${extern_protobuf_PROTOC_EXECUTABLE}
+        CACHE FILEPATH "protobuf executable." FORCE)
+    PROMPT_PROTOBUF_LIB(extern_protobuf)
 ENDIF(NOT PROTOBUF_FOUND)
diff --git a/cmake/external/pslib.cmake b/cmake/external/pslib.cmake
index 3b495d78e2c61f90418adbc5746792bc6e49d90b..b4ea268e5a48e29d00b0ec8b957b61a42553ec7e 100644
--- a/cmake/external/pslib.cmake
+++ b/cmake/external/pslib.cmake
@@ -71,7 +71,3 @@ ADD_LIBRARY(pslib SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET pslib PROPERTY IMPORTED_LOCATION ${PSLIB_LIB})
 ADD_DEPENDENCIES(pslib ${PSLIB_PROJECT})
 LIST(APPEND external_project_dependencies pslib)
-
-IF(WITH_C_API)
-  INSTALL(FILES ${PSLIB_LIB} ${PSLIB_IOMP_LIB} DESTINATION lib)
-ENDIF()
diff --git a/cmake/external/pslib_brpc.cmake b/cmake/external/pslib_brpc.cmake
index 7ff5a8aca187240108164900638f5a376e9fbc93..8b43f2ef5c999fc351543ba958c7cc4b0856625d 100644
--- a/cmake/external/pslib_brpc.cmake
+++ b/cmake/external/pslib_brpc.cmake
@@ -71,7 +71,3 @@ ADD_LIBRARY(pslib_brpc SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET pslib_brpc PROPERTY IMPORTED_LOCATION ${PSLIB_BRPC_LIB})
 ADD_DEPENDENCIES(pslib_brpc ${PSLIB_BRPC_PROJECT})
 LIST(APPEND external_project_dependencies pslib_brpc)
-
-IF(WITH_C_API)
-  INSTALL(FILES ${PSLIB_BRPC_LIB} ${PSLIB_BRPC_IOMP_LIB} DESTINATION lib)
-ENDIF()
diff --git a/cmake/external/snappy.cmake b/cmake/external/snappy.cmake
index f9d4cd97400a68e613e3dd5467191a0d42a9942e..27d075336d556528ffaf1929c34753494692f0a0 100644
--- a/cmake/external/snappy.cmake
+++ b/cmake/external/snappy.cmake
@@ -12,10 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-if(MOBILE_INFERENCE OR RPI)
-    return()
-endif()
-
 include (ExternalProject)
 
 # NOTE: snappy is needed when linking with recordio
diff --git a/cmake/external/snappystream.cmake b/cmake/external/snappystream.cmake
index 1ec79462c14e44f2d0abe6904383ebd91d94d35e..392f186b7ce3821f313ed6fc3dd5a97c2a7adebd 100644
--- a/cmake/external/snappystream.cmake
+++ b/cmake/external/snappystream.cmake
@@ -12,10 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-IF(MOBILE_INFERENCE OR RPI)
-    return()
-ENDIF()
-
 include (ExternalProject)
 
 set(SNAPPYSTREAM_SOURCES_DIR ${THIRD_PARTY_PATH}/snappy_stream)
diff --git a/cmake/external/swig.cmake b/cmake/external/swig.cmake
deleted file mode 100644
index de07703695eb14e76eedd3758d55cb98edd1e02b..0000000000000000000000000000000000000000
--- a/cmake/external/swig.cmake
+++ /dev/null
@@ -1,65 +0,0 @@
-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-IF(NOT WITH_SWIG_PY)
-    return()
-ENDIF()
-
-FIND_PACKAGE(SWIG)
-
-IF(NOT SWIG_FOUND)
-    # build swig as an external project
-    INCLUDE(ExternalProject)
-
-    SET(SWIG_SOURCES_DIR ${THIRD_PARTY_PATH}/swig)
-    SET(SWIG_INSTALL_DIR ${THIRD_PARTY_PATH}/install/swig)
-    SET(SWIG_TARGET_VERSION "3.0.2")
-    SET(SWIG_DOWNLOAD_SRC_MD5 "62f9b0d010cef36a13a010dc530d0d41")
-    SET(SWIG_DOWNLOAD_WIN_MD5 "3f18de4fc09ab9abb0d3be37c11fbc8f")
-
-    IF(WIN32)
-        # swig.exe available as pre-built binary on Windows:
-        ExternalProject_Add(swig
-            URL                 http://prdownloads.sourceforge.net/swig/swigwin-${SWIG_TARGET_VERSION}.zip
-            URL_MD5             ${SWIG_DOWNLOAD_WIN_MD5}
-            SOURCE_DIR          ${SWIG_SOURCES_DIR}
-            CONFIGURE_COMMAND   ""
-            BUILD_COMMAND       ""
-            INSTALL_COMMAND     ""
-            UPDATE_COMMAND      ""
-        )
-        SET(SWIG_DIR ${SWIG_SOURCES_DIR} CACHE FILEPATH "SWIG Directory" FORCE)
-        SET(SWIG_EXECUTABLE ${SWIG_SOURCES_DIR}/swig.exe  CACHE FILEPATH "SWIG Executable" FORCE)
-    ELSE(WIN32)
-        # swig uses bison find it by cmake and pass it down
-        FIND_PACKAGE(BISON)
-
-        # From SWIG configure
-        ExternalProject_Add(swig
-            GIT_REPOSITORY      https://github.com/swig/swig.git
-            GIT_TAG             rel-3.0.10
-            PREFIX              ${SWIG_SOURCES_DIR}
-            CONFIGURE_COMMAND   cd <SOURCE_DIR> && ./autogen.sh && ./configure
-                                --prefix=${SWIG_INSTALL_DIR} --without-pcre
-            BUILD_COMMAND       cd <SOURCE_DIR> && make
-            INSTALL_COMMAND     cd <SOURCE_DIR> && make install
-            UPDATE_COMMAND      ""
-        )
-
-        SET(SWIG_DIR ${SWIG_INSTALL_DIR}/share/swig/${SWIG_TARGET_VERSION})
-        SET(SWIG_EXECUTABLE ${SWIG_INSTALL_DIR}/bin/swig)
-    ENDIF(WIN32)
-
-    LIST(APPEND external_project_dependencies swig)
-ENDIF(NOT SWIG_FOUND)
diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake
index 7b937c93febdfa1d7d5f4c73fc2a5830322688e5..7a25aaf15f2c7f46d99394d82d69bc24e4f5cb2c 100644
--- a/cmake/external/warpctc.cmake
+++ b/cmake/external/warpctc.cmake
@@ -12,10 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-IF(MOBILE_INFERENCE)
-    return()
-ENDIF()
-
 INCLUDE(ExternalProject)
 
 SET(WARPCTC_SOURCES_DIR ${THIRD_PARTY_PATH}/warpctc)
diff --git a/cmake/external/xxhash.cmake b/cmake/external/xxhash.cmake
index c3e1212d8f8358e0148b5e00223414c9696686ee..a0f300c2e8bab9e7402f869eed1b4c2d1c579aab 100644
--- a/cmake/external/xxhash.cmake
+++ b/cmake/external/xxhash.cmake
@@ -73,12 +73,3 @@ include_directories(${XXHASH_INCLUDE_DIR})
 add_dependencies(xxhash extern_xxhash)
 
 LIST(APPEND external_project_dependencies xxhash)
-
-IF(WITH_C_API)
-  INSTALL(DIRECTORY ${XXHASH_INCLUDE_DIR} DESTINATION third_party/xxhash)
-  IF(ANDROID)
-    INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib/${ANDROID_ABI})
-  ELSE()
-    INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib)
-  ENDIF()
-ENDIF()
diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake
index d35073753725cd5772de3fc7a23af5ba69a65558..6c8d79c25e6a2655711fe4450e65600c9a584015 100644
--- a/cmake/external/zlib.cmake
+++ b/cmake/external/zlib.cmake
@@ -59,12 +59,3 @@ SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
 ADD_DEPENDENCIES(zlib extern_zlib)
 
 LIST(APPEND external_project_dependencies zlib)
-
-IF(WITH_C_API)
-  INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib)
-  IF(ANDROID)
-    INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib/${ANDROID_ABI})
-  ELSE()
-    INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib)
-  ENDIF()
-ENDIF()
diff --git a/cmake/flags.cmake b/cmake/flags.cmake
index c4472040cef870454c072c1b84a04e1ac592b476..9e6c47f016fe6dfd809c5b2bc88ff59d0a6b2b84 100644
--- a/cmake/flags.cmake
+++ b/cmake/flags.cmake
@@ -156,10 +156,8 @@ set(GPU_COMMON_FLAGS
 endif(NOT WIN32)
 
 if (APPLE)
-    if(NOT CMAKE_CROSSCOMPILING)
-        # On Mac OS X build fat binaries with x86_64 architectures by default.
-        set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
-    endif()
+    # On Mac OS X build fat binaries with x86_64 architectures by default.
+    set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
     # On Mac OS X register class specifier is deprecated and will cause warning error on latest clang 10.0
     set (COMMON_FLAGS -Wno-deprecated-register)
 endif(APPLE)
diff --git a/cmake/generic.cmake b/cmake/generic.cmake
index 3f1be11d85555671eebb1c2ba3a5642d64d7f2bf..1f4dbe0b49825aef9a236f7ae72c6bea168b2ec5 100644
--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@@ -90,11 +90,11 @@
 # including binary directory for generated headers.
 include_directories(${CMAKE_CURRENT_BINARY_DIR})
 
-if(NOT APPLE AND NOT ANDROID)
+if(NOT APPLE)
   find_package(Threads REQUIRED)
   link_libraries(${CMAKE_THREAD_LIBS_INIT})
   set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt")
-endif(NOT APPLE AND NOT ANDROID)
+endif(NOT APPLE)
 
 set_property(GLOBAL PROPERTY FLUID_MODULES "")
 # find all fluid modules is used for paddle fluid static library
@@ -655,12 +655,6 @@ function(paddle_protobuf_generate_cpp SRCS HDRS)
   set(${SRCS})
   set(${HDRS})
 
-  if (MOBILE_INFERENCE)
-      set(EXTRA_FLAG "lite:")
-  else()
-      set(EXTRA_FLAG "")
-  endif()
-
   foreach(FIL ${ARGN})
     get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
     get_filename_component(FIL_WE ${FIL} NAME_WE)
@@ -677,7 +671,7 @@ function(paddle_protobuf_generate_cpp SRCS HDRS)
       COMMAND ${CMAKE_COMMAND} -E make_directory "${CMAKE_CURRENT_BINARY_DIR}"
       COMMAND ${PROTOBUF_PROTOC_EXECUTABLE}
       -I${CMAKE_CURRENT_SOURCE_DIR}
-      --cpp_out "${EXTRA_FLAG}${CMAKE_CURRENT_BINARY_DIR}" ${ABS_FIL}
+      --cpp_out "${CMAKE_CURRENT_BINARY_DIR}" ${ABS_FIL}
       DEPENDS ${ABS_FIL} protoc
       COMMENT "Running C++ protocol buffer compiler on ${FIL}"
       VERBATIM )
diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake
index 3e11d332ff71098adf65e487a39351ae57427e9e..a7dce4dfdb530b13bea9df128694f0946714ccff 100644
--- a/cmake/inference_lib.cmake
+++ b/cmake/inference_lib.cmake
@@ -149,25 +149,23 @@ if (WITH_NGRAPH)
             )
 endif ()
 
-if (NOT MOBILE_INFERENCE AND NOT RPI)
-    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
-    copy(snappy_lib
-            SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
-            DSTS ${dst_dir} ${dst_dir}/lib
-            DEPS snappy)
+set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
+copy(snappy_lib
+        SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
+        DSTS ${dst_dir} ${dst_dir}/lib
+        DEPS snappy)
 
-    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
-    copy(snappystream_lib
-            SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
-            DSTS ${dst_dir} ${dst_dir}/lib
-            DEPS snappystream)
+set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
+copy(snappystream_lib
+        SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
+        DSTS ${dst_dir} ${dst_dir}/lib
+        DEPS snappystream)
 
-    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
-    copy(zlib_lib
-            SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
-            DSTS ${dst_dir} ${dst_dir}/lib
-            DEPS zlib)
-endif ()
+set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
+copy(zlib_lib
+        SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
+        DSTS ${dst_dir} ${dst_dir}/lib
+        DEPS zlib)
 
 # paddle fluid module
 set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
diff --git a/cmake/system.cmake b/cmake/system.cmake
index c91ef91127511da9ac8b9e11349f4a23aaedd613..65db05bebe957d740e391847d980e211b0e9e750 100644
--- a/cmake/system.cmake
+++ b/cmake/system.cmake
@@ -74,21 +74,6 @@ MARK_AS_ADVANCED(HOST_SYSTEM CPU_CORES)
 MESSAGE(STATUS "Found Paddle host system: ${HOST_SYSTEM}, version: ${HOST_SYSTEM_VERSION}")
 MESSAGE(STATUS "Found Paddle host system's CPU: ${CPU_CORES} cores")
 
-# configuration for cross-compiling
-IF(DEFINED CMAKE_SYSTEM_NAME)
-    INCLUDE(cross_compiling/host)
-    IF(${CMAKE_SYSTEM_NAME} STREQUAL "Android")
-        SET(ANDROID TRUE)
-        INCLUDE(cross_compiling/android)
-    ELSEIF(${CMAKE_SYSTEM_NAME} STREQUAL "RPi")
-        SET(RPI TRUE)
-        INCLUDE(cross_compiling/raspberry_pi)
-    ELSEIF(${CMAKE_SYSTEM_NAME} STREQUAL "iOS")
-        SET(IOS TRUE)
-        INCLUDE(cross_compiling/ios)
-    ENDIF()
-ENDIF()
-
 # external dependencies log output
 SET(EXTERNAL_PROJECT_LOG_ARGS
     LOG_DOWNLOAD    0     # Wrap download in script to log output
diff --git a/cmake/util.cmake b/cmake/util.cmake
index 0dc33ce385175d1e2dc454d41db467d4b9d9cf9a..02667dbce69ed159193ff88f38069dd08cdcf678 100644
--- a/cmake/util.cmake
+++ b/cmake/util.cmake
@@ -53,118 +53,3 @@ function(target_circle_link_libraries TARGET_NAME)
                 "-Wl,--end-group")
     endif()
 endfunction()
-
-# compile_cu_as_cpp
-# Make a cu file compiled as C++
-# Arguments: Source files
-macro(compile_cu_as_cpp)
-    foreach(s ${ARGN})
-        set_source_files_properties(${s} PROPERTIES LANGUAGE CXX)
-        set_source_files_properties(${s} PROPERTIES COMPILE_FLAGS "-x c++")
-    endforeach()
-endmacro()
-
-# link_paddle_exe
-# add paddle library for a paddle executable, such as trainer, pserver.
-#
-# It will handle WITH_PYTHON etc.
-function(link_paddle_exe TARGET_NAME)
-    if(WITH_RDMA)
-        generate_rdma_links()
-    endif()
-
-    if(MOBILE_INFERENCE)
-        target_circle_link_libraries(${TARGET_NAME}
-            ARCHIVE_START
-            paddle_gserver
-            paddle_function
-            ARCHIVE_END
-            paddle_math
-            paddle_utils
-            paddle_parameter
-            paddle_proto
-            paddle_cuda
-            ${EXTERNAL_LIBS}
-            ${CMAKE_THREAD_LIBS_INIT}
-            ${CMAKE_DL_LIBS}
-            ${RDMA_LD_FLAGS}
-            ${RDMA_LIBS})
-    else()
-        target_circle_link_libraries(${TARGET_NAME}
-            ARCHIVE_START
-            paddle_gserver
-            paddle_function
-            ARCHIVE_END
-            paddle_pserver
-            paddle_trainer_lib
-            paddle_network
-            paddle_math
-            paddle_utils
-            paddle_parameter
-            paddle_proto
-            paddle_cuda
-            paddle_optimizer
-            ${EXTERNAL_LIBS}
-            ${CMAKE_THREAD_LIBS_INIT}
-            ${CMAKE_DL_LIBS}
-            ${RDMA_LD_FLAGS}
-            ${RDMA_LIBS})
-    endif()
-
-    if(ANDROID)
-        target_link_libraries(${TARGET_NAME} log)
-    endif(ANDROID)
-
-    if(WITH_MKLML AND MKLML_LIB_DIR AND MKLML_IOMP_LIB)
-      target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
-    endif()
-
-    add_dependencies(${TARGET_NAME} ${external_project_dependencies})
-endfunction()
-
-# link_paddle_test
-# Link a paddle unittest for target
-# TARGET_NAME: the unittest target name
-# Rest Arguemnts: not used.
-function(link_paddle_test TARGET_NAME)
-    link_paddle_exe(${TARGET_NAME})
-    target_link_libraries(${TARGET_NAME}
-                          paddle_test_main
-                          paddle_test_util
-                          ${GTEST_LIBRARIES})
-endfunction()
-
-# add_unittest_without_exec
-#
-# create a paddle unittest. not specifically define how to run this unittest.
-# TARGET_NAME: the unittest target name, same as executable file name
-# Rest Arguments: the source files to compile this unittest.
-macro(add_unittest_without_exec TARGET_NAME)
-    add_executable(${TARGET_NAME} ${ARGN})
-    link_paddle_test(${TARGET_NAME})
-endmacro()
-
-# add_unittest
-# create a paddle unittest and just to execute this binary to make unittest.
-#
-# TARGET_NAME: the unittest target name, same as executable file name
-# Rest Arguments: the source files to compile this unittest.
-macro(add_unittest TARGET_NAME)
-    add_unittest_without_exec(${TARGET_NAME} ${ARGN})
-    add_test(${TARGET_NAME} ${TARGET_NAME})
-endmacro()
-
-# add_simple_unittest
-# create a paddle unittest with file name. It just compile ${TARGET_NAME}.cpp to
-# ${TARGET_NAME} and then execute it.
-macro(add_simple_unittest TARGET_NAME)
-    add_unittest(${TARGET_NAME} ${TARGET_NAME}.cpp)
-endmacro()
-
-# Creates C resources file from files in given resource file
-function(create_resources res_file output_file)
-  add_custom_command(
-    OUTPUT ${output_file}
-    COMMAND python ARGS ${PADDLE_SOURCE_DIR}/cmake/make_resource.py ${res_file} ${output_file}
-    DEPENDS ${res_file} ${PADDLE_SOURCE_DIR}/cmake/make_resource.py)
-endfunction()
diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc
index 756470c5b0bfb2ce2a90531442211bcb19172dae..ce5731a1f414e8ef6d8af22a3bb17109e82beb87 100644
--- a/paddle/fluid/framework/details/build_strategy.cc
+++ b/paddle/fluid/framework/details/build_strategy.cc
@@ -24,6 +24,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/details/sequential_execution_pass.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/ir/graph_to_program_pass.h"
 #include "paddle/fluid/framework/ir/graph_viz_pass.h"
 
 namespace paddle {
@@ -243,3 +244,4 @@ USE_PASS(sequential_execution_pass);
 USE_PASS(all_reduce_deps_pass);
 USE_PASS(modify_op_lock_and_record_event_pass);
 USE_PASS(lock_free_optimize_pass);
+USE_PASS(graph_to_program_pass);
diff --git a/paddle/fluid/framework/ir/pass.cc b/paddle/fluid/framework/ir/pass.cc
index 6cf405efe63d2bc284c4650771a747b27bb4a9f6..33ccee6aa0a94b8fd8308214d6144ae832d40bab 100644
--- a/paddle/fluid/framework/ir/pass.cc
+++ b/paddle/fluid/framework/ir/pass.cc
@@ -28,10 +28,14 @@ std::unique_ptr<Graph> Pass::Apply(std::unique_ptr<Graph> graph) const {
     PADDLE_ENFORCE(graph->Has(attr), "Required graph atrribute %s not set.",
                    attr);
   }
+  auto* native_graph = graph.get();
   auto applied_graph = ApplyImpl(std::move(graph));
   // TODO(panyx0718): Add more verifications.
   PADDLE_ENFORCE(!HasCircle(*applied_graph),
                  "Illegal Pass. Generated graph shouldn't has cycle.");
+  PADDLE_ENFORCE(applied_graph.get() == native_graph,
+                 "Pass::Apply() cannot delete the passed graph and shouldn't "
+                 "return a new graph.(For the need of pybind11)");
   applied_ = true;
   return applied_graph;
 }
diff --git a/paddle/fluid/operators/conv_fusion_op.cu.cc b/paddle/fluid/operators/conv_fusion_op.cu.cc
index d8b997cca613f660046106512fc03bf55f9b992d..f97ebecfdd90beade3bef824c04ad7b2763eb036 100644
--- a/paddle/fluid/operators/conv_fusion_op.cu.cc
+++ b/paddle/fluid/operators/conv_fusion_op.cu.cc
@@ -104,7 +104,9 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
     // ------------------- cudnn conv algorithm ---------------------
     cudnnConvolutionFwdAlgo_t algo;
     auto handle = dev_ctx.cudnn_handle();
-    auto workspace_handle = dev_ctx.cudnn_workspace_handle();
+
+    Tensor cudnn_workspace;
+    void* cudnn_workspace_ptr = nullptr;
 
     CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
         cudnn_conv_desc, CUDNN_DEFAULT_MATH));
@@ -118,19 +120,24 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
           workspace_size_limit, &algo));
       VLOG(3) << "cuDNN forward algo " << algo;
     } else {
+      cudnn_workspace =
+          ctx.AllocateTmpTensor<int8_t, platform::CUDADeviceContext>(
+              framework::make_ddim(
+                  {static_cast<int64_t>(workspace_size_limit)}),
+              dev_ctx);
+      cudnn_workspace_ptr = static_cast<void*>(cudnn_workspace.data<int8_t>());
+
       auto search_func = [&]() {
         int returned_algo_count;
         std::array<cudnnConvolutionFwdAlgoPerf_t, kNUM_CUDNN_FWD_ALGS>
             fwd_perf_stat;
-        auto cudnn_find_func = [&](void* cudnn_workspace) {
-          CUDNN_ENFORCE(
-              platform::dynload::cudnnFindConvolutionForwardAlgorithmEx(
-                  handle, cudnn_input_desc, input_data, cudnn_filter_desc,
-                  filter_data, cudnn_conv_desc, cudnn_output_desc, output_data,
-                  kNUM_CUDNN_FWD_ALGS, &returned_algo_count,
-                  fwd_perf_stat.data(), cudnn_workspace, workspace_size_limit));
-        };
-        workspace_handle.RunFunc(cudnn_find_func, workspace_size_limit);
+
+        CUDNN_ENFORCE(platform::dynload::cudnnFindConvolutionForwardAlgorithmEx(
+            handle, cudnn_input_desc, input_data, cudnn_filter_desc,
+            filter_data, cudnn_conv_desc, cudnn_output_desc, output_data,
+            kNUM_CUDNN_FWD_ALGS, &returned_algo_count, fwd_perf_stat.data(),
+            cudnn_workspace_ptr, workspace_size_limit));
+
         VLOG(3) << "Perf result: (algo: stat, time, memory)";
         for (int i = 0; i < returned_algo_count; ++i) {
           const auto& stat = fwd_perf_stat[i];
@@ -181,6 +188,15 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_LE(workspace_size_in_bytes, workspace_size_limit,
                       "workspace_size to be allocated exceeds the limit");
 
+    if (!cudnn_workspace_ptr) {
+      cudnn_workspace =
+          ctx.AllocateTmpTensor<int8_t, platform::CUDADeviceContext>(
+              framework::make_ddim(
+                  {static_cast<int64_t>(workspace_size_in_bytes)}),
+              dev_ctx);
+      cudnn_workspace_ptr = static_cast<void*>(cudnn_workspace.data<int8_t>());
+    }
+
     if ((activation == "identity") && (!residual)) {
       // Only the CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM algo is
       // enabled with CUDNN_ACTIVATION_IDENTITY in cuDNN lib.
@@ -188,13 +204,12 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
       // cudnnConvolutionForward and cudnnAddTensor
       // ------------- cudnn conv forward and bias add ---------------------
       ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
-      auto cudnn_func = [&](void* cudnn_workspace) {
-        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
-            handle, &alpha, cudnn_input_desc, input_data, cudnn_filter_desc,
-            filter_data, cudnn_conv_desc, algo, cudnn_workspace,
-            workspace_size_in_bytes, &beta, cudnn_output_desc, output_data));
-      };
-      workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
+
+      CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
+          handle, &alpha, cudnn_input_desc, input_data, cudnn_filter_desc,
+          filter_data, cudnn_conv_desc, algo, cudnn_workspace_ptr,
+          workspace_size_in_bytes, &beta, cudnn_output_desc, output_data));
+
       CUDNN_ENFORCE(platform::dynload::cudnnAddTensor(
           handle, &alpha, cudnn_bias_desc, bias_data, &alpha, cudnn_output_desc,
           output_data));
@@ -205,15 +220,13 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
       // ------------------- cudnn conv+bias+act forward --------------------
       ScalingParamType<T> alpha1 = 1.0f;
       ScalingParamType<T> alpha2 = residual ? 1.0f : 0.0f;
-      auto cudnn_func = [&](void* cudnn_workspace) {
-        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBiasActivationForward(
-            handle, &alpha1, cudnn_input_desc, input_data, cudnn_filter_desc,
-            filter_data, cudnn_conv_desc, algo, cudnn_workspace,
-            workspace_size_in_bytes, &alpha2, cudnn_output_desc, residual_data,
-            cudnn_bias_desc, bias_data, cudnn_act_desc, cudnn_output_desc,
-            output_data));
-      };
-      workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
+
+      CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBiasActivationForward(
+          handle, &alpha1, cudnn_input_desc, input_data, cudnn_filter_desc,
+          filter_data, cudnn_conv_desc, algo, cudnn_workspace_ptr,
+          workspace_size_in_bytes, &alpha2, cudnn_output_desc, residual_data,
+          cudnn_bias_desc, bias_data, cudnn_act_desc, cudnn_output_desc,
+          output_data));
     }
     std::vector<int> channels = ctx.Attr<std::vector<int>>("split_channels");
     if (channels.size()) {
diff --git a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc
index f44094ca6b7b7f23f2e7593ad79e4e2a6f0d3070..016cf8448c5e07fdedab8c5e4a7d0ae9e2ded1ee 100644
--- a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc
+++ b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc
@@ -104,16 +104,18 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
     int output_offset = output->numel() / output->dims()[0] / groups;
     int filter_offset = filter->numel() / groups;
     T alpha = 1.0f, beta = 0.0f;
-    auto workspace_handle = dev_ctx.cudnn_workspace_handle();
+
+    auto temp_allocation =
+        platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
+            workspace_size_in_bytes);
+    void* cudnn_workspace = temp_allocation->ptr();
+
     for (int g = 0; g < groups; g++) {
-      auto cudnn_func = [&](void* cudnn_workspace) {
-        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
-            handle, &alpha, cudnn_filter_desc, filter_data + filter_offset * g,
-            cudnn_input_desc, input_data + input_offset * g, cudnn_conv_desc,
-            algo, cudnn_workspace, workspace_size_in_bytes, &beta,
-            cudnn_output_desc, output_data + output_offset * g));
-      };
-      workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
+      CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
+          handle, &alpha, cudnn_filter_desc, filter_data + filter_offset * g,
+          cudnn_input_desc, input_data + input_offset * g, cudnn_conv_desc,
+          algo, cudnn_workspace, workspace_size_in_bytes, &beta,
+          cudnn_output_desc, output_data + output_offset * g));
     }
   }
 };
@@ -209,20 +211,22 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
         output_grad->numel() / output_grad->dims()[0] / groups;
     int filter_offset = filter->numel() / groups;
     T alpha = 1.0f, beta = 0.0f;
-    auto workspace_handle = dev_ctx.cudnn_workspace_handle();
+
+    auto temp_allocation =
+        platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
+            workspace_size_in_bytes);
+    void* cudnn_workspace = temp_allocation->ptr();
+
     if (input_grad) {
       T* input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
       // Because beta is zero, it is unnecessary to reset input_grad.
       for (int g = 0; g < groups; g++) {
-        auto cudnn_func = [&](void* cudnn_workspace) {
-          CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
-              handle, &alpha, cudnn_output_desc,
-              output_grad_data + output_grad_offset * g, cudnn_filter_desc,
-              filter_data + filter_offset * g, cudnn_conv_desc, data_algo,
-              cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc,
-              input_grad_data + input_offset * g));
-        };
-        workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
+        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
+            handle, &alpha, cudnn_output_desc,
+            output_grad_data + output_grad_offset * g, cudnn_filter_desc,
+            filter_data + filter_offset * g, cudnn_conv_desc, data_algo,
+            cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc,
+            input_grad_data + input_offset * g));
       }
     }
 
@@ -232,15 +236,12 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
       // Because beta is zero, it is unnecessary to reset filter_grad.
       // Gradient with respect to the filter
       for (int g = 0; g < groups; g++) {
-        auto cudnn_func = [&](void* cudnn_workspace) {
-          CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
-              handle, &alpha, cudnn_output_desc,
-              output_grad_data + output_grad_offset * g, cudnn_input_desc,
-              input_data + input_offset * g, cudnn_conv_desc, filter_algo,
-              cudnn_workspace, workspace_size_in_bytes, &beta,
-              cudnn_filter_desc, filter_grad_data + filter_offset * g));
-        };
-        workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
+        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
+            handle, &alpha, cudnn_output_desc,
+            output_grad_data + output_grad_offset * g, cudnn_input_desc,
+            input_data + input_offset * g, cudnn_conv_desc, filter_algo,
+            cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_filter_desc,
+            filter_grad_data + filter_offset * g));
       }
     }
   }
diff --git a/paddle/fluid/operators/distributed/collective_server_test.cc b/paddle/fluid/operators/distributed/collective_server_test.cc
index 5009058422b81d3187f7792bf7bf56db1d03f4d6..90f2f9fd65bf1b8c1edda6a2ebe0ce5288ddcb5d 100644
--- a/paddle/fluid/operators/distributed/collective_server_test.cc
+++ b/paddle/fluid/operators/distributed/collective_server_test.cc
@@ -20,6 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/tensor_util.h"
 
 #include "paddle/fluid/operators/distributed/collective_client.h"
 #include "paddle/fluid/operators/distributed/collective_server.h"
@@ -57,7 +58,7 @@ std::unique_ptr<framework::Scope> GenerateVars(platform::Place place) {
   auto* tensor = slr->mutable_value();
   auto* rows = slr->mutable_rows();
 
-  tensor->Resize(framework::make_ddim({20000, 1024}));
+  tensor->Resize(framework::make_ddim({3, 1024}));
   tensor->mutable_data<float>(place);
 
   paddle::operators::math::set_constant(ctx, tensor, 32.7);
@@ -80,6 +81,20 @@ void Gather(const std::vector<distributed::RemoteVar>& vars,
   std::vector<const framework::SelectedRows*> dst;
   client->Gather(vars, &dst, *dev_ctx, scope);
   std::cout << "dst:" << distributed::GetSelectedRowsInfo(*dst[0]);
+  dev_ctx->Wait();
+
+  ASSERT_EQ(dst[0]->value().dims(), framework::make_ddim({3, 1024}));
+  ASSERT_EQ(dst[0]->height(), 20000);
+  ASSERT_EQ(dst[0]->rows().size(), static_cast<size_t>(3));
+  for (int i = 0; i < 3; i++) {
+    ASSERT_EQ(dst[0]->rows()[i], i);
+  }
+
+  std::vector<float> vec;
+  TensorToVector(dst[0]->value(), *dev_ctx, &vec);
+  for (size_t i = 0; i < 3 * 1024; i++) {
+    ASSERT_FLOAT_EQ(vec[i], 32.7);
+  }
 }
 
 TEST(CollectiveServer, GPU) {
diff --git a/paddle/fluid/operators/fused/fusion_conv_inception_op.cu b/paddle/fluid/operators/fused/fusion_conv_inception_op.cu
index 6e13887866485bd114ebf12f4bdfa8d60fca6d01..c72a966c575d4a63471905b82643e96454f08187 100644
--- a/paddle/fluid/operators/fused/fusion_conv_inception_op.cu
+++ b/paddle/fluid/operators/fused/fusion_conv_inception_op.cu
@@ -216,18 +216,19 @@ class CUDNNConvInceptionFusionOpKernel : public framework::OpKernel<T> {
     out_datas.push_back(
         static_cast<void*>(output_data + (oc0 + oc1 + oc2) * h * w));
 
+    auto temp_allocation =
+        platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
+            workspace_size_in_bytes);
+    void* cudnn_workspace = temp_allocation->ptr();
+
     for (int i = 0; i < 4; ++i) {
-      auto func = [&](void* cudnn_workspace) {
-        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBiasActivationForward(
-            handle, &alpha, in_desc[i], in_datas[i], filter_desc[i],
-            static_cast<const void*>(filters[i]->data<T>()), conv_desc[i],
-            algo[i], cudnn_workspace, workspace_size_in_bytes, &beta,
-            out_desc[i], out_datas[i], bias_desc[i],
-            static_cast<const void*>(bias[i]->data<T>()), cudnn_act_desc,
-            out_desc[i], out_datas[i]));
-      };
-      auto workspace_handle = dev_ctx.cudnn_workspace_handle();
-      workspace_handle.RunFunc(func, workspace_size_in_bytes);
+      CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBiasActivationForward(
+          handle, &alpha, in_desc[i], in_datas[i], filter_desc[i],
+          static_cast<const void*>(filters[i]->data<T>()), conv_desc[i],
+          algo[i], cudnn_workspace, workspace_size_in_bytes, &beta, out_desc[i],
+          out_datas[i], bias_desc[i],
+          static_cast<const void*>(bias[i]->data<T>()), cudnn_act_desc,
+          out_desc[i], out_datas[i]));
     }
 
     cudnnTensorDescriptor_t x_desc;
diff --git a/paddle/fluid/operators/warpctc_cudnn_op.cu.cc b/paddle/fluid/operators/warpctc_cudnn_op.cu.cc
index a764d59410c90535dbda0b3f11e89ae9bf578c04..5e16a209e712a143e1083e171f88002817aef838 100644
--- a/paddle/fluid/operators/warpctc_cudnn_op.cu.cc
+++ b/paddle/fluid/operators/warpctc_cudnn_op.cu.cc
@@ -144,17 +144,19 @@ class CudnnCTCKernel : public framework::OpKernel<T> {
         CUDNN_CTC_LOSS_ALGO_DETERMINISTIC, cu_ctcloss_desc, &workspace_size));
 
     T* loss_data = loss->mutable_data<T>(loss_dims, ctx.GetPlace());
-
-    auto workspace_handle = dev_ctx.cudnn_workspace_handle();
-    auto cudnn_func = [&](void* cudnn_workspace) {
-      CUDNN_ENFORCE(platform::dynload::cudnnCTCLoss(
-          handle, cu_logits_desc, warpctc_logits_data, warpctc_label_data,
-          warpctc_label_lengths.data(), warpctc_logits_lengths.data(),
-          loss_data, cu_grad_desc, warpctc_grad_data,
-          CUDNN_CTC_LOSS_ALGO_DETERMINISTIC, cu_ctcloss_desc, cudnn_workspace,
-          workspace_size));
-    };
-    workspace_handle.RunFunc(cudnn_func, workspace_size);
+    math::SetConstant<DeviceContext, T>()(
+        ctx.template device_context<DeviceContext>(), loss, static_cast<T>(0));
+
+    auto temp_allocation =
+        platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
+            workspace_size);
+    void* cudnn_workspace = temp_allocation->ptr();
+
+    CUDNN_ENFORCE(platform::dynload::cudnnCTCLoss(
+        handle, cu_logits_desc, warpctc_logits_data, warpctc_label_data,
+        warpctc_label_lengths.data(), warpctc_logits_lengths.data(), loss_data,
+        cu_grad_desc, warpctc_grad_data, CUDNN_CTC_LOSS_ALGO_DETERMINISTIC,
+        cu_ctcloss_desc, cudnn_workspace, workspace_size));
   }
 };
 
diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt
index 67b2386813eeafca5ac627edf79cad7e02445b8c..803ea6b26087884ad79c6bf80238953a012eaddc 100644
--- a/paddle/fluid/pybind/CMakeLists.txt
+++ b/paddle/fluid/pybind/CMakeLists.txt
@@ -18,9 +18,9 @@ if(WITH_PYTHON)
       SRCS ${PYBIND_SRCS}
       DEPS ${PYBIND_DEPS}
       ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS})
-    if(NOT APPLE AND NOT ANDROID AND NOT WIN32)
+    if(NOT APPLE AND NOT WIN32)
       target_link_libraries(paddle_pybind rt)
-    endif(NOT APPLE AND NOT ANDROID AND NOT WIN32)
+    endif(NOT APPLE AND NOT WIN32)
   endif(WITH_AMD_GPU)
 
   get_property (os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
diff --git a/paddle/fluid/pybind/ir.cc b/paddle/fluid/pybind/ir.cc
index d32fe58f8695a5c14f276ef038416f5c47f3400f..24059140ab20e24917b93a5f60936b1087797ff9 100644
--- a/paddle/fluid/pybind/ir.cc
+++ b/paddle/fluid/pybind/ir.cc
@@ -15,7 +15,9 @@
 #include "paddle/fluid/pybind/ir.h"
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 #include "paddle/fluid/framework/ir/node.h"
 #include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/framework/var_desc.h"
@@ -24,6 +26,7 @@
 namespace py = pybind11;
 using paddle::framework::ir::Graph;
 using paddle::framework::ir::Node;
+using paddle::framework::ir::GraphSafeRemoveNodes;
 using paddle::framework::OpDesc;
 using paddle::framework::ProgramDesc;
 using paddle::framework::VarDesc;
@@ -32,6 +35,7 @@ using pybind11::return_value_policy;
 namespace paddle {
 namespace pybind {
 void BindGraph(py::module *m) {
+  m->def("graph_safe_remove_nodes", GraphSafeRemoveNodes);
   py::class_<Graph, std::shared_ptr<Graph>>(
       *m, "Graph",
       "The graph is a Directed Acyclic Single Static Assignment Graph, see "
@@ -42,6 +46,8 @@ void BindGraph(py::module *m) {
       .def("get_float", &Graph::Get<float>)
       .def("get_double", &Graph::Get<double>)
       .def("get_string", &Graph::Get<std::string>)
+      .def("get_program", &Graph::Get<ProgramDesc>)
+      .def("get_marked_nodes", &Graph::Get<std::unordered_set<const Node *>>)
       .def("set", [](Graph &self, const std::string &attr_name,
                      int attr) { return self.Set(attr_name, new int(attr)); })
       .def("set",
@@ -57,6 +63,17 @@ void BindGraph(py::module *m) {
            [](Graph &self, const std::string &attr_name, double attr) {
              return self.Set(attr_name, new double(attr));
            })
+      .def("set",
+           [](Graph &self, const std::string &attr_name,
+              const ProgramDesc &attr) {
+             return self.Set(attr_name, new ProgramDesc(attr));
+           })
+      .def("set",
+           [](Graph &self, const std::string &attr_name,
+              const std::unordered_set<const Node *> &attr) {
+             return self.Set(attr_name,
+                             new std::unordered_set<const Node *>(attr));
+           })
       .def("erase", &Graph::Erase)
       .def("nodes", &Graph::Nodes, return_value_policy::reference)
       .def("create_var_node",
@@ -85,12 +102,52 @@ void BindNode(py::module *m) {
   py::class_<Node> node(*m, "Node");
   node.def("name", &Node::Name)
       .def("node_type", &Node::NodeType)
-      .def("var", &Node::Var)
-      .def("op", &Node::Op)
+      .def("var", &Node::Var, return_value_policy::reference)
+      .def("op", &Node::Op, return_value_policy::reference)
       .def("id", &Node::id)
       .def("is_op", &Node::IsOp)
       .def("is_var", &Node::IsVar)
       .def("is_ctrl_var", &Node::IsCtrlVar)
+      .def("inputs_remove",
+           [](Node &self, int node_id) {
+             for (auto it = self.inputs.begin(); it != self.inputs.end();
+                  it++) {
+               if ((*it)->id() == node_id) {
+                 self.inputs.erase(it);
+               }
+             }
+           })
+      .def("inputs_remove",
+           [](Node &self, Node &node) {
+             for (auto it = self.inputs.begin(); it != self.inputs.end();
+                  it++) {
+               if (*it == &node) {
+                 self.inputs.erase(it);
+               }
+             }
+           })
+      .def("inputs_append",
+           [](Node &self, Node &node) { self.inputs.push_back(&node); })
+      .def("outputs_remove",
+           [](Node &self, int node_id) {
+             for (auto it = self.outputs.begin(); it != self.outputs.end();
+                  it++) {
+               if ((*it)->id() == node_id) {
+                 self.outputs.erase(it);
+               }
+             }
+           })
+      .def("outputs_remove",
+           [](Node &self, Node &node) {
+             for (auto it = self.outputs.begin(); it != self.outputs.end();
+                  it++) {
+               if (*it == &node) {
+                 self.outputs.erase(it);
+               }
+             }
+           })
+      .def("outputs_append",
+           [](Node &self, Node &node) { self.outputs.push_back(&node); })
       .def_readwrite("inputs", &Node::inputs)
       .def_readwrite("outputs", &Node::outputs);
 
diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc
index 4b218fb3a2af0933ea1e87abe20e7e031c32f721..e729be4a95a58510f1e0162af4216feaa400d971 100644
--- a/paddle/fluid/pybind/protobuf.cc
+++ b/paddle/fluid/pybind/protobuf.cc
@@ -228,7 +228,7 @@ void BindBlockDesc(pybind11::module *m) {
 
 void BindVarDsec(pybind11::module *m) {
   pybind11::class_<pd::VarDesc> var_desc(*m, "VarDesc", "");
-  var_desc
+  var_desc.def(pybind11::init<const std::string &>())
       .def("name", &pd::VarDesc::Name, pybind11::return_value_policy::reference)
       .def("set_name", &pd::VarDesc::SetName)
       .def("set_shape", &pd::VarDesc::SetShape)
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 3ea38266775ed59b3f3e2a09a6cf5c2a653cdbc0..d84ee06492d52c93e0b774c6e08218aca8a14916 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -788,21 +788,33 @@ All parameter, weight, gradient are variables in Paddle.
   m.def("disable_profiler", platform::DisableProfiler);
   m.def("is_profiler_enabled", platform::IsProfileEnabled);
   m.def("reset_profiler", platform::ResetProfiler);
+  m.def("get_pass", [](const py::bytes &binary_str) {
+    std::string pass_type(binary_str);
+    auto pass = framework::ir::PassRegistry::Instance().Get(pass_type);
+    return std::shared_ptr<framework::ir::Pass>(std::move(pass));
+  });
 
   py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
   pass.def(py::init())
+      .def("has", &ir::Pass::Has)
+      .def("set",
+           [](ir::Pass &self, const std::string &attr_name,
+              const ProgramDesc &attr) {
+             return self.Set(attr_name, new ProgramDesc(attr));
+           })
       .def(
-          "set_str",
+          "set",
           [](ir::Pass &self, const std::string &name, const std::string &attr) {
             self.Set<std::string>(name, new std::string(attr));
           })
-      .def("set_int", [](ir::Pass &self, const std::string &name,
-                         int val) { self.Set<const int>(name, new int(val)); })
+      .def("set", [](ir::Pass &self, const std::string &name,
+                     int val) { self.Set<const int>(name, new int(val)); })
+      .def("get_program", &ir::Pass::Get<ProgramDesc>)
       .def("type", &ir::Pass::Type)
       .def("apply", [](ir::Pass &self, std::shared_ptr<ir::Graph> graph) {
         std::unique_ptr<ir::Graph> origin_graph(graph.get());
         auto optim_graph = self.Apply(std::move(origin_graph));
-        graph.reset(optim_graph.release());
+        optim_graph.release();
       });
 
   py::class_<ir::PassBuilder, std::shared_ptr<ir::PassBuilder>> pb(
diff --git a/paddle/py_paddle/.gitignore b/paddle/py_paddle/.gitignore
deleted file mode 100644
index 80d1f76fbc05627e21e334af55d63a4a534434c6..0000000000000000000000000000000000000000
--- a/paddle/py_paddle/.gitignore
+++ /dev/null
@@ -1,2 +0,0 @@
-swig_paddle.py
-_swig_paddle.so
diff --git a/paddle/py_paddle/dataprovider_converter.py b/paddle/py_paddle/dataprovider_converter.py
deleted file mode 100644
index 43614b9779d21795f1f274589ea93639e923ce75..0000000000000000000000000000000000000000
--- a/paddle/py_paddle/dataprovider_converter.py
+++ /dev/null
@@ -1,309 +0,0 @@
-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import paddle.trainer.PyDataProvider2 as dp2
-import collections
-import swig_paddle
-import numpy
-import itertools
-from functools import reduce
-
-__all__ = ['DataProviderConverter']
-
-
-class IScanner(object):
-    """
-    The scanner will scan Python object two passes, then convert it to Paddle's
-    argument.
-
-    In the first pass, `pre_scan` will be invoked by every data instance, and
-    then invoke `finish_pre_scan` to arguments. And the second pass do the same
-    thing except the functions changed to `scan`, `finish_scan`.
-
-    During the first pass, a scanner may count the shape of input matrix and
-    allocate memory for this argument. Then fill the data into this  argument
-    in second pass.
-    """
-
-    def __init__(self, input_type, pos):
-        self.input_type = input_type
-        if not isinstance(self.input_type, dp2.InputType):
-            raise ValueError("input type should be dataprovider2.InputType")
-        self.pos = pos
-        # data_in_gpu is used to indicate whether to create argument on GPU
-        # or not in GPU mode. Now if using one thread (trainer_count=1),
-        # trainer uses NeuralNetwork which needs to create argument on GPU
-        # before calling forward function. So, set data_in_gpu to True.
-        # Otherwise, trainer uses MultiGradientMachine which will transfer
-        # data from CPU to GPU in the forward function, set data_in_gpu to
-        # False in this case.
-        self.data_in_gpu = swig_paddle.isUsingGpu(
-        ) and swig_paddle.getTrainerCount() == 1
-
-    def pre_scan(self, dat):
-        """
-        First pass scan method. During this method, the scanner could count the
-        data number, and get the total memory size this batch would use.
-
-        :param dat: The python object.
-        """
-        pass
-
-    def finish_pre_scan(self, argument):
-        """
-        Finish first scan pass. Allocate the memory.
-
-        :param argument: Output arguments object.
-        :type argument: swig_paddle.Arguments
-        :param dat: Output arguments object.
-        :type dat: The Python object, numpy.array or List.
-        :return:
-        """
-        pass
-
-    def scan(self, dat):
-        """
-        Second pass scan method. Copy the data to arguments.
-
-        :param dat: The python object.
-        """
-        pass
-
-    def finish_scan(self, argument):
-        """
-        Finish second pass. Finalize the resources, etc.
-
-        :param argument: Output arguments object.
-        :type argument: swig_paddle.Arguments
-        """
-        pass
-
-
-class DenseScanner(IScanner):
-    """
-    :type __mat__: numpy.ndarray
-    """
-
-    def __init__(self, input_type, pos):
-        IScanner.__init__(self, input_type, pos)
-        self.__mat__ = None
-        self.__shape__ = None
-        self.__height__ = 0
-        self.__dim__ = 0
-
-    def pre_scan(self, dat):
-        self.__height__ += 1
-        if self.__shape__ is None:
-            self.__shape__ = numpy.array(dat).shape
-            if len(self.__shape__) > 3:
-                raise ValueError(
-                    "The dimension of input cannot be greater than 3.")
-            if len(self.__shape__) == 0:
-                raise ValueError(
-                    "The input should be a vector, please check your input data."
-                )
-            self.__dim__ = reduce(lambda x, y: x * y, self.__shape__)
-            if len(self.__shape__) == 1 and self.__dim__ != self.input_type.dim:
-                raise ValueError(
-                    "The data size must be equal to it in data layer.")
-        else:
-            if self.__shape__ != numpy.array(dat).shape:
-                raise ValueError(
-                    "The data shape must be same in one mini-batch.")
-
-    def finish_pre_scan(self, argument):
-        self.__mat__ = numpy.ndarray(
-            shape=(self.__height__, self.__dim__), dtype=numpy.float32)
-        self.__height__ = 0
-
-    def scan(self, dat):
-        # It's better to use NumPy array for speed.
-        dat = numpy.array(dat)
-        dat = dat.flatten()
-        self.__mat__[self.__height__] = dat
-        self.__height__ += 1
-
-    def finish_scan(self, argument):
-        assert isinstance(argument, swig_paddle.Arguments)
-        if self.__mat__.dtype != numpy.float32:
-            self.__mat__ = self.__mat__.astype(numpy.float32)
-        m = swig_paddle.Matrix.createDenseFromNumpy(self.__mat__, True,
-                                                    self.data_in_gpu)
-        argument.setSlotValue(self.pos, m)
-        if len(self.__shape__) > 1:
-            # The last-two dimenstions are the frame height and width.
-            # For example, the layout is CHW for 3-D feature of image.
-            # The H and W are the frame height and width.
-            h, w = self.__shape__[-2:]
-            argument.setSlotFrameHeight(self.pos, h)
-            argument.setSlotFrameWidth(self.pos, w)
-        self.__shape__ = None
-
-
-class SparseBinaryScanner(IScanner):
-    def __init__(self, input_type, pos):
-        IScanner.__init__(self, input_type, pos)
-        self.__rows__ = [0]
-        self.__cols__ = []
-        self.__height__ = 0
-        self.__value__ = []
-
-    def scan(self, dat):
-        self.extend_cols(dat)
-        self.__rows__.append(len(self.__cols__))
-        self.__height__ += 1
-
-    def extend_cols(self, dat):
-        self.__cols__.extend(dat)
-
-    def finish_scan(self, argument):
-        assert isinstance(argument, swig_paddle.Arguments)
-        m = swig_paddle.Matrix.createSparse(
-            self.__height__,
-            self.input_type.dim,
-            len(self.__cols__),
-            len(self.__value__) == 0,
-            False,  # trans
-            False)  # TODO supoort GPU
-        assert isinstance(m, swig_paddle.Matrix)
-        m.sparseCopyFrom(self.__rows__, self.__cols__, self.__value__)
-        argument.setSlotValue(self.pos, m)
-
-
-class SparseFloatScanner(SparseBinaryScanner):
-    def __init__(self, input_type, pos):
-        SparseBinaryScanner.__init__(self, input_type, pos)
-
-    def extend_cols(self, dat):
-        self.__cols__.extend((x[0] for x in dat))
-        self.__value__.extend((x[1] for x in dat))
-
-
-class IndexScanner(IScanner):
-    def __init__(self, input_type, pos):
-        IScanner.__init__(self, input_type, pos)
-        self.__ids__ = None
-        self.__idx__ = 0
-
-    def pre_scan(self, dat):
-        self.__idx__ += 1
-
-    def finish_pre_scan(self, argument):
-        self.__ids__ = [0] * self.__idx__
-        self.__idx__ = 0
-
-    def scan(self, dat):
-        self.__ids__[self.__idx__] = dat
-        self.__idx__ += 1
-
-    def finish_scan(self, argument):
-        ids = swig_paddle.IVector.create(self.__ids__, self.data_in_gpu)
-        assert isinstance(argument, swig_paddle.Arguments)
-        argument.setSlotIds(self.pos, ids)
-
-
-class SequenceScanner(IScanner):
-    def __init__(self, input_type, pos, inner_scanner, setter):
-        IScanner.__init__(self, input_type, pos)
-        self.__seq__ = [0]
-        self.__inner_scanner__ = inner_scanner
-        self.__setter__ = setter
-
-    def pre_scan(self, dat):
-        for each in dat:
-            self.__inner_scanner__.pre_scan(each)
-
-    def finish_pre_scan(self, argument):
-        self.__inner_scanner__.finish_pre_scan(argument)
-
-    def scan(self, dat):
-        self.__seq__.append(self.__seq__[-1] + self.get_size(dat))
-        for each in dat:
-            self.__inner_scanner__.scan(each)
-
-    def finish_scan(self, argument):
-        seq = swig_paddle.IVector.create(self.__seq__, False)
-        self.__setter__(argument, self.pos, seq)
-        self.__inner_scanner__.finish_scan(argument)
-
-    def get_size(self, dat):
-        if isinstance(self.__inner_scanner__, SequenceScanner):
-            return sum(self.__inner_scanner__.get_size(item) for item in dat)
-        else:
-            return len(dat)
-
-
-class DataProviderConverter(object):
-    def __init__(self, input_types):
-        self.input_types = input_types
-        assert isinstance(self.input_types, collections.Sequence)
-        for each in self.input_types:
-            assert isinstance(each, dp2.InputType)
-
-    def convert(self, dat, argument=None):
-        if argument is None:
-            argument = swig_paddle.Arguments.createArguments(0)
-        assert isinstance(argument, swig_paddle.Arguments)
-        argument.resize(len(self.input_types))
-
-        scanners = [
-            DataProviderConverter.create_scanner(i, each_type)
-            for i, each_type in enumerate(self.input_types)
-        ]
-
-        for each_sample in dat:
-            for each_step, scanner in itertools.izip(each_sample, scanners):
-                scanner.pre_scan(each_step)
-
-        for scanner in scanners:
-            scanner.finish_pre_scan(argument)
-
-        for each_sample in dat:
-            for each_step, scanner in itertools.izip(each_sample, scanners):
-                scanner.scan(each_step)
-
-        for scanner in scanners:
-            scanner.finish_scan(argument)
-
-        return argument
-
-    def __call__(self, dat, argument=None):
-        return self.convert(dat, argument)
-
-    @staticmethod
-    def create_scanner(i, each):
-        assert isinstance(each, dp2.InputType)
-        retv = None
-        if each.type == dp2.DataType.Dense:
-            retv = DenseScanner(each, i)
-        elif each.type == dp2.DataType.Index:
-            retv = IndexScanner(each, i)
-        elif each.type == dp2.DataType.SparseNonValue:
-            retv = SparseBinaryScanner(each, i)
-        elif each.type == dp2.DataType.SparseValue:
-            retv = SparseFloatScanner(each, i)
-        assert retv is not None
-
-        if each.seq_type == dp2.SequenceType.SUB_SEQUENCE:
-            retv = SequenceScanner(
-                each, i, retv,
-                lambda a, p, seq: a.setSlotSubSequenceStartPositions(p, seq))
-
-        if each.seq_type in [
-                dp2.SequenceType.SUB_SEQUENCE, dp2.SequenceType.SEQUENCE
-        ]:
-            retv = SequenceScanner(
-                each, i, retv,
-                lambda a, p, seq: a.setSlotSequenceStartPositions(p, seq))
-        return retv
diff --git a/paddle/py_paddle/util.py b/paddle/py_paddle/util.py
deleted file mode 100644
index 3ae8dbf964c68c6f01ba30cb3ac69fb6c2f08c30..0000000000000000000000000000000000000000
--- a/paddle/py_paddle/util.py
+++ /dev/null
@@ -1,578 +0,0 @@
-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""
-Some Useful method for py_paddle.
-"""
-
-import swig_paddle
-import os
-import paddle.trainer.PyDataProviderWrapper
-import paddle.proto.ParameterConfig_pb2
-import paddle.proto.ModelConfig_pb2
-import paddle.proto.TrainerConfig_pb2
-import weakref
-import numpy
-import struct
-import sys
-import copy
-
-
-def initializePaddle(*args):
-    """
-    To initialize paddle process.
-    :param args: Command line options, such as --use_gpu=0, etc.
-    :return: Nothing.
-    """
-    old_argv = copy.deepcopy(sys.argv)
-    old_pypath = os.getenv("PYTHONPATH")
-    pypath = os.path.abspath(os.path.join(os.path.dirname(__file__), '..'))
-    if old_pypath is not None:
-        pypath = os.pathsep.join([pypath, old_pypath])
-        os.putenv("PYTHONPATH", pypath)
-    args = [""] + list(args)  # argv[0] is command name, it is not important.
-    swig_paddle.__initPaddle__(args)
-    sys.argv = old_argv
-
-
-def __monkeypatch_init_paddle__():
-    swig_paddle.__initPaddle__ = swig_paddle.initPaddle
-    swig_paddle.initPaddle = initializePaddle
-
-
-class __ParameterCallbackWrapper__(swig_paddle.UpdateCallback):
-    """
-    Wrap the python callable object to paddle.UpdateCallback.
-
-    INTERNAL USE ONLY.
-    """
-
-    def __init__(self, callback):
-        swig_paddle.UpdateCallback.__init__(self)
-        self.callback = callback
-
-    def apply(self, param):
-        self.callback(param)
-
-    @staticmethod
-    def wrap(callback):
-        """
-        Cast the python callable object/paddle.UpdateCallback to
-        swig_paddle.UpdateCallback.__disown__
-        :param callback: callable or swig_paddle.UpdateCallback object.
-        """
-        if isinstance(callback, swig_paddle.UpdateCallback):
-            return callback.__disown__()
-        elif isinstance(callback, weakref.ProxyType):
-            raise RuntimeError("Should not pass __disown__ object")
-        else:
-            return __ParameterCallbackWrapper__(callback).__disown__()
-
-
-def __arguments_to_numpy__(i, arg):
-    assert isinstance(arg, swig_paddle.Arguments)
-    value = arg.getSlotValue(i)
-    ids = arg.getSlotIds(i)
-    prob = arg.getSlotIn(i)
-    if value is not None:
-        assert isinstance(value, swig_paddle.Matrix)
-        value = value.copyToNumpyMat()
-    if ids is not None:
-        assert isinstance(ids, swig_paddle.IVector)
-        ids = ids.copyToNumpyArray()
-    if prob is not None:
-        assert isinstance(prob, swig_paddle.Matrix)
-        prob = prob.copyToNumpyMat()
-    return {"value": value, "id": ids, "prob": prob}
-
-
-def __monkeypatch_gradient_machine__():
-    """
-    Add some class methods to GradientMachine.
-    This method should be only used internally.
-    """
-    swig_paddle.GradientMachine.loadFromConfigFile = \
-        staticmethod(loadGradientMachine)
-
-    def __matrix_to_numpy__(m):
-        if isinstance(m, swig_paddle.Matrix):
-            return m.copyToNumpyMat()
-        elif isinstance(m, swig_paddle.IVector):
-            return m.copyToNumpyArra()
-        else:
-            raise RuntimeError("Input arg should be matrix or vecotr.")
-
-    def createFromConfigProto(protoObj,
-                              createMode=swig_paddle.CREATE_MODE_NORMAL,
-                              paramTypes=[
-                                  swig_paddle.PARAMETER_VALUE,
-                                  swig_paddle.PARAMETER_GRADIENT,
-                                  swig_paddle.PARAMETER_MOMENTUM
-                              ]):
-        """
-        Create Gradient Machine From Proto object.
-        :param protoObj: Model config
-        :type protoObj: proto.ModelConfig_pb2.ModelConfig
-        :param createMode: Create Mode, default is normal.
-        :type createMode: int
-        :param paramTypes: the gradient machine parameter type.
-        :type paramTypes: list of int
-        :return: paddle.GradientMachine
-        """
-        assert isinstance(protoObj, paddle.proto.ModelConfig)
-        return swig_paddle.GradientMachine.createByConfigProtoStr(
-            protoObj.SerializeToString(), createMode, paramTypes)
-
-    swig_paddle.GradientMachine.createFromConfigProto = \
-        staticmethod(createFromConfigProto)
-
-    def forwardTest(self, inArgs):
-        """
-        forwardTest. forward gradient machine in test mode, and return a numpy
-        matrix dict.
-
-        :param inArgs: The input arguments
-        :type inArgs: paddle.Arguments
-        :return: A dictionary with keys ['id', 'value'], each value is a
-                 numpy.ndarray.
-        """
-        outArgs = swig_paddle.Arguments.createArguments(0)
-        self.forward(inArgs, outArgs, swig_paddle.PASS_TEST)
-        return [
-            __arguments_to_numpy__(i, outArgs)
-            for i in xrange(outArgs.getSlotNum())
-        ]
-
-    swig_paddle.GradientMachine.forwardTest = forwardTest
-
-    # Monkey patching backward
-    swig_paddle.GradientMachine.__backward__ = swig_paddle.GradientMachine.backward
-
-    def backward(self, callback):
-        """
-        GradientMachine Backward
-        :param callback: a callback which parameter is (paddle.Parameter) or
-                         a paddle.UpdateCallback object.
-        """
-        self.__backward__(__ParameterCallbackWrapper__.wrap(callback))
-
-    swig_paddle.GradientMachine.backward = backward
-
-    # Monkey patching forwardBackward.
-    swig_paddle.GradientMachine.__forwardBackward__ = \
-        swig_paddle.GradientMachine.forwardBackward
-
-    def forwardBackward(self,
-                        inArgs,
-                        outArgs,
-                        passType,
-                        callback=swig_paddle.UpdateCallback()):
-        """
-        GradientMachine forward backward.
-        :param inArgs: Input Arguments for GradientMachine.
-        :type inArgs: paddle.Arguments
-        :param outArgs: Output Arguments for GradientMachine.
-        :type outArgs: paddle.Arguments
-        :param passType: gradient machine's pass type.
-        :type passType: paddle.PassType
-        :param callback: a callable object with arguments (paddle.Parameter) or
-                         a paddle.UpdateCallback it will be called when
-                         backward
-        """
-        self.__forwardBackward__(inArgs, outArgs, passType,
-                                 __ParameterCallbackWrapper__.wrap(callback))
-
-    swig_paddle.GradientMachine.forwardBackward = forwardBackward
-
-    def getParameters(self):
-        return (self.getParameter(i) for i in xrange(self.getParameterSize()))
-
-    swig_paddle.GradientMachine.getParameters = getParameters
-
-    def getNonStaticParameters(self):
-        return (self.getNonStaticParameter(i)
-                for i in xrange(self.getNonStaticParameterSize()))
-
-    swig_paddle.GradientMachine.getNonStaticParameters = getNonStaticParameters
-
-    def getLayerOutputs(self, layerNames):
-        """
-        getLayerOutputs. get outputs of layers and return a numpy matrix dict.
-        :param layerNames: layer names.
-        :type layerNames: string or list.
-        """
-        if isinstance(layerNames, basestring):
-            layerNames = [layerNames]
-        elif not isinstance(layerNames, list):
-            raise RuntimeError("Input args shuld be string or a sting list.")
-
-        output = dict()
-        for name in layerNames:
-            output[name] = __arguments_to_numpy__(0, self.getLayerOutput(name))
-        return output
-
-    swig_paddle.GradientMachine.getLayerOutputs = getLayerOutputs
-
-
-def loadGradientMachine(config_filename, model_dir=None):
-    """
-    Load a gradient machine from config file name/path.
-    :param config_filename: The trainer config file name/path
-    :param model_dir: The model parameter directory. None if same as the
-    directory of config_filename
-    :return: GradientMachine with some enhance methods.
-    :rtype: paddle.GradientMachine
-    """
-    trainer_config = swig_paddle.TrainerConfig.createFromTrainerConfigFile(
-        config_filename)
-    assert isinstance(trainer_config, swig_paddle.TrainerConfig)
-    model_conf = trainer_config.getModelConfig()
-    network = swig_paddle.GradientMachine.createByModelConfig(model_conf)
-    assert isinstance(network, swig_paddle.GradientMachine)
-    if model_dir is None:
-        model_dir = os.path.dirname(config_filename)
-    network.loadParameters(model_dir)
-    return network
-
-
-def loadParameterFile(fn):
-    """
-    Load Paddle Parameter file to numpy.ndarray
-    :param fn: file name or file like object.
-    :type fn: str or file like object.
-    :return: numpy array
-    :rtype: numpy.ndarray
-    :raise: paddle.UnsupportError when parameter format is wrong.
-    """
-    if isinstance(fn, str):
-        with open(fn, 'rb') as f:
-            return loadParameterFile(f)
-    elif hasattr(fn, 'read'):  # File like object
-        version, = struct.unpack('i', fn.read(4))
-        if version != 0:
-            raise swig_paddle.UnsupportError()
-        value_length, = struct.unpack("I", fn.read(4))
-        if value_length != 4 and value_length != 8:
-            raise swig_paddle.UnsupportError()
-        dtype = 'float32' if value_length == 4 else 'float64'
-        param_size, = struct.unpack("L", fn.read(8))
-        value = numpy.fromfile(fn, dtype)
-        if len(value) != param_size:
-            raise swig_paddle.UnsupportError()
-        return value
-    else:
-        raise swig_paddle.UnsupportError()
-
-
-class DataProviderWrapperConverter(object):
-    """
-    A class convert DataFormat from PyDataProvider Wrapper to
-    py_paddle.paddle.Arguemnts.
-    """
-
-    class DenseValueConverter(object):
-        """
-        Internal class
-        """
-
-        def __init__(self, header_def):
-            self.__dim__ = header_def.dim
-            self.buf = []
-
-        def append(self, other):
-            assert len(other) == self.__dim__
-            self.buf += other
-
-        def __call__(self, slot_idx, arg):
-            mat = swig_paddle.Matrix.createDense(self.buf,
-                                                 len(self.buf) / self.__dim__,
-                                                 self.__dim__)
-            arg.setSlotValue(slot_idx, mat)
-
-    class IdValueConverter(object):
-        """
-        Internal class
-        """
-
-        def __init__(self, *args):
-            self.buf = []
-
-        def append(self, other):
-            assert isinstance(other, int)
-            self.buf.append(other)
-
-        def __call__(self, slot_idx, arg):
-            arg.setSlotIds(slot_idx, swig_paddle.IVector.create(self.buf))
-
-    class SparseNonValueConverter(object):
-        """
-        Internal class
-        """
-
-        def __init__(self, slot_def):
-            self.indices = [0]
-            self.cols = []
-            self.dim = slot_def.dim
-
-        def append(self, other):
-            self.indices.append(self.indices[-1] + len(other))
-            self.cols += other
-
-        def __call__(self, slot_idx, arg):
-            mat = swig_paddle.Matrix.createSparse(
-                len(self.indices) - 1, self.dim, len(self.cols), True)
-            assert isinstance(mat, swig_paddle.Matrix)
-            mat.sparseCopyFrom(self.indices, self.cols)
-            self.putIntoArg(slot_idx, arg, mat)
-
-        def putIntoArg(self, slot_idx, arg, mat):
-            arg.setSlotValue(slot_idx, mat)
-
-    class SparseValueConverter(SparseNonValueConverter):
-        """
-        Internal class
-        """
-
-        def __init__(self, slot_def):
-            super(DataProviderWrapperConverter.SparseValueConverter,
-                  self).__init__(slot_def)
-            self.values = []
-
-        def append(self, other):
-            super(DataProviderWrapperConverter.SparseValueConverter,
-                  self).append(map(lambda x: x[0], other))
-            self.values += map(lambda x: x[1], other)
-
-        def __call__(self, slot_idx, arg):
-            mat = swig_paddle.Matrix.createSparse(
-                len(self.indices) - 1, self.dim, len(self.cols), False)
-            assert isinstance(mat, swig_paddle.Matrix)
-            mat.sparseCopyFrom(self.indices, self.cols, self.values)
-            self.putIntoArg(slot_idx, arg, mat)
-
-    __SLOT_VALUE_CONVERTER_MAP__ = {
-        paddle.trainer.PyDataProviderWrapper.DenseSlot: DenseValueConverter,
-        paddle.trainer.PyDataProviderWrapper.IndexSlot: IdValueConverter,
-        paddle.trainer.PyDataProviderWrapper.SparseNonValueSlot:
-        SparseNonValueConverter,
-        paddle.trainer.PyDataProviderWrapper.SparseValueSlot:
-        SparseValueConverter
-    }
-
-    def __init__(self, use_seq, header):
-        """
-        Ctor
-        :param use_seq: True if use sequence.
-        :param header:  List of slots type,
-                       trainer.PyDataProviderWrapper.SlotType
-        """
-        self.__use_seq__ = use_seq
-        self.__header__ = header
-
-    def convert(self, wrapper_data, argument=None):
-        """
-        Convert PyDataProviderWrapper format to paddle.Argument
-        :param wrapper_data: PyDataProviderWrapper yield's data list.
-        :param argument: The output paddle.Arguments.
-                        If it is not None, it will assign data in this
-                        arguments, else it will create new arguments.
-        :return: arguments that contains data.
-        :rtype: paddle.Arguments
-        """
-        if argument is None:
-            argument = swig_paddle.Arguments.createArguments(0)
-        assert isinstance(argument, swig_paddle.Arguments)
-        argument.resize(len(self.__header__))
-
-        values = map(
-            lambda x: DataProviderWrapperConverter.__SLOT_VALUE_CONVERTER_MAP__[x.__class__](x),
-            self.__header__)
-
-        if self.__use_seq__:
-            seq_dim = [[] for _ in xrange(self.__header__.__len__())]
-            seq_start_pos = [[0] for _ in xrange(self.__header__.__len__())]
-
-            for each_sample in wrapper_data:
-                for slot_idx, sequence in enumerate(each_sample):
-                    for raw_data in sequence:
-                        values[slot_idx].append(raw_data)
-                    seq_start_pos[slot_idx].append(seq_start_pos[slot_idx][-1] +
-                                                   len(sequence))
-                    seq_dim[slot_idx].append(len(sequence))
-
-            for slot_idx in xrange(len(self.__header__)):
-                argument.setSlotSequenceDim(
-                    slot_idx, swig_paddle.IVector.create(seq_dim[slot_idx]))
-                argument.setSlotSequenceStartPositions(
-                    slot_idx,
-                    swig_paddle.IVector.create(seq_start_pos[slot_idx]))
-        else:
-            for each_sample in wrapper_data:
-                for raw_data, value in zip(each_sample, values):
-                    value.append(raw_data)
-
-        for i, v in enumerate(values):
-            v(i, argument)
-
-        return argument
-
-    def __call__(self, wrapper_data, argument=None):
-        """
-        Invoke self.convert. See documents in self.convert.
-        """
-        return self.convert(wrapper_data, argument)
-
-
-def __monkey_patch_protobuf_objects__():
-    def ParameterConfig_toProto(self):
-        """
-        Convert paddle.ParameterConfig to
-        proto.ParameterConfig_pb2.ParameterConfig
-
-        :return: proto.ParameterConfig_pb2.ParameterConfig object.
-        """
-        param_conf = paddle.proto.ParameterConfig_pb2.ParameterConfig()
-        param_conf.ParseFromString(self.toProtoString())
-        return param_conf
-
-    swig_paddle.ParameterConfig.toProto = ParameterConfig_toProto
-
-    def OptimizationConfig_toProto(self):
-        """
-        Convert paddle.OptimizationConfig to
-        proto.TrainerConfig_pb2.OptimizationConfig
-
-        :return: proto.TrainerConfig_pb2.OptimizationConfig
-        """
-        opt_conf = proto.TrainerConfig_pb2.OptimizationConfig()
-        opt_conf.ParseFromString(self.toProtoString())
-        return opt_conf
-
-    swig_paddle.OptimizationConfig.toProto = OptimizationConfig_toProto
-
-    def OptimizationConfig_createFromProto(protoObj):
-        """
-        Create a new paddle.OptimizationConfig from
-        proto.TrainerConfig_pb2.OptimizationConfig
-
-        :param protoObj: proto.TrainerConfig_pb2.OptimizationConfig
-        :return: paddle.OptimizationConfig
-        """
-
-        assert isinstance(protoObj, paddle.proto.OptimizationConfig)
-        return swig_paddle.OptimizationConfig.createFromProtoString(
-            protoObj.SerializeToString())
-
-    swig_paddle.OptimizationConfig.createFromProto = staticmethod(
-        OptimizationConfig_createFromProto)
-
-    def TrainerConfig_createFromProto(protoObj):
-        """
-        Create a new paddle.TrainerConfig from
-        proto.OptimizationConfig
-
-        :param protoObj: proto.TrainerConfig
-        :return: paddle.TrainerConfig
-        """
-        assert isinstance(protoObj, paddle.proto.TrainerConfig)
-        return swig_paddle.TrainerConfig.createFromProtoString(
-            protoObj.SerializeToString())
-
-    swig_paddle.TrainerConfig.createFromProto = staticmethod(
-        TrainerConfig_createFromProto)
-
-
-def __monkey_patch_parameter__():
-    def getBufs(self):
-        """
-        get all parameter vectors.
-        NOTE: the return value is a generator. Maybe you need to cast to
-        list or tuple or something else.
-
-        :return: generator of all parameter vectors.
-        :rtype: generator
-        """
-        return (self.getBuf(i) for i in xrange(swig_paddle.NUM_PARAMETER_TYPES))
-
-    swig_paddle.Parameter.getBufs = getBufs
-
-
-def __monkey_patch_trainer__():
-    swig_paddle.Trainer.__create__ = staticmethod(swig_paddle.Trainer.create)
-
-    def Trainer_create(config, model=None):
-        """
-        Create a trainer for model with TrainerCOnfig trainer_config
-        trainer_config.model_config will be ignored when model is supplied.
-        Trainer.trainOneBatch() and Trainer.forwardOneBatch() can be used only
-        when trainer_config.data_config is set.
-
-        A typical usage for Trainer is:
-        .. code-block:: python
-           trainer = Trainer.create(trainer_config, model)
-           for p in xrange(num_passes)
-               while True:
-                   data = get_next_batch(batch_size)
-                   if not data:
-                       break
-                   trainer.trainOneDataBatch(batch_size, data)
-               trainer.finishTrainPass()
-           trainer.finishTrain()
-
-        The trainer will take care of logging, model saving, distributed
-        training, etc.
-
-        :param config: trainer configuration
-        :type config: paddle.proto.TrainerConfig
-        :param model: the model to be trained
-        :type model: swig_paddle.GradientMachine
-        :return: a trainer
-        :rtype swig_paddle.Trainer
-
-        """
-        assert isinstance(config, paddle.proto.TrainerConfig)
-        if model is not None:
-            assert isinstance(model, swig_paddle.GradientMachine)
-        return swig_paddle.Trainer.__create__(
-            swig_paddle.TrainerConfig.createFromProto(config), model)
-
-    swig_paddle.Trainer.create = staticmethod(Trainer_create)
-
-    swig_paddle.Trainer.__getForwardOutput__ = \
-        swig_paddle.Trainer.getForwardOutput
-
-    def getForwardOutput(self):
-        """
-        Get the netword outputs from the previous trainOneBatch(),
-        trainOneDataBatch(), testOneDataPatch(), or forwardOneBatch() call.
-
-        :return: list of dictionary with keys ['id', 'value'], each value is a
-                 numpy.ndarray.
-        """
-        outArgs = self.__getForwardOutput__()
-        return [
-            __arguments_to_numpy__(i, outArgs)
-            for i in xrange(outArgs.getSlotNum())
-        ]
-
-    swig_paddle.Trainer.getForwardOutput = getForwardOutput
-
-
-def monkeypatches():
-    patches = [
-        __monkeypatch_init_paddle__, __monkeypatch_gradient_machine__,
-        __monkey_patch_protobuf_objects__, __monkey_patch_parameter__,
-        __monkey_patch_trainer__
-    ]
-    for patch in patches:
-        patch()
diff --git a/paddle/scripts/README.md b/paddle/scripts/README.md
index 9e8b135c1bc7fc05d88fe6f3bed17dd3b48e9615..6c608fce3cdad38f3109e563be3ffbe2f73e5390 100644
--- a/paddle/scripts/README.md
+++ b/paddle/scripts/README.md
@@ -40,7 +40,6 @@ The lastest pre-built build environment images are:
 | Image | Tag |
 | ----- | --- |
 | paddlepaddle/paddle | latest-dev |
-| paddlepaddle/paddle | latest-dev-android |
 
 ### Start Build
 
@@ -68,8 +67,6 @@ Users can specify the following Docker build arguments with either "ON" or "OFF"
 | `WITH_TESTING` | OFF | Build unit tests binaries. |
 | `WITH_MKL` | ON | Build with [Intel® MKL](https://software.intel.com/en-us/mkl) and [Intel® MKL-DNN](https://github.com/01org/mkl-dnn) support. |
 | `WITH_GOLANG` | OFF | Build fault-tolerant parameter server written in go. |
-| `WITH_SWIG_PY` | ON | Build with SWIG python API support. |
-| `WITH_C_API` | OFF | Build capi libraries for inference. |
 | `WITH_PYTHON` | ON | Build with python support. Turn this off if build is only for capi. |
 | `WITH_STYLE_CHECK` | ON | Check the code style when building. |
 | `PYTHON_ABI` | "" | Build for different python ABI support, can be cp27-cp27m or cp27-cp27mu |
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
index f58e392684d619e145b07ac61d2adfe175443bb6..cda04451f5e6a9a27c9352626f3f08e886fca04b 100755
--- a/paddle/scripts/paddle_build.sh
+++ b/paddle/scripts/paddle_build.sh
@@ -33,7 +33,6 @@ function print_usage() {
     ${BLUE}gen_doc_lib${NONE}: generate paddle documents library
     ${BLUE}html${NONE}: convert C++ source code into HTML
     ${BLUE}dockerfile${NONE}: generate paddle release dockerfile
-    ${BLUE}capi${NONE}: generate paddle CAPI package
     ${BLUE}fluid_inference_lib${NONE}: deploy fluid inference library
     ${BLUE}check_style${NONE}: run code style check
     ${BLUE}cicheck${NONE}: run CI tasks
@@ -180,9 +179,7 @@ function cmake_gen() {
         -DWITH_AVX=${WITH_AVX:-OFF}
         -DWITH_GOLANG=${WITH_GOLANG:-OFF}
         -DCUDA_ARCH_NAME=${CUDA_ARCH_NAME:-All}
-        -DWITH_C_API=${WITH_C_API:-OFF}
         -DWITH_PYTHON=${WITH_PYTHON:-ON}
-        -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON}
         -DCUDNN_ROOT=/usr/
         -DWITH_TESTING=${WITH_TESTING:-ON}
         -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake
@@ -216,8 +213,6 @@ EOF
         -DWITH_AVX=${WITH_AVX:-OFF} \
         -DWITH_GOLANG=${WITH_GOLANG:-OFF} \
         -DCUDA_ARCH_NAME=${CUDA_ARCH_NAME:-All} \
-        -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \
-        -DWITH_C_API=${WITH_C_API:-OFF} \
         -DWITH_PYTHON=${WITH_PYTHON:-ON} \
         -DCUDNN_ROOT=/usr/ \
         -DWITH_TESTING=${WITH_TESTING:-ON} \
@@ -706,59 +701,43 @@ EOF
 EOF
 }
 
-function gen_capi_package() {
-    if [[ ${WITH_C_API} == "ON" ]]; then
-        capi_install_prefix=${INSTALL_PREFIX:-/paddle/build}/capi_output
-        rm -rf $capi_install_prefix
-        make DESTDIR="$capi_install_prefix" install
-        cd $capi_install_prefix/
-        ls | egrep -v "^Found.*item$" | xargs tar -czf ${PADDLE_ROOT}/build/paddle.tgz
-    fi
-}
-
 function gen_fluid_lib() {
     mkdir -p ${PADDLE_ROOT}/build
     cd ${PADDLE_ROOT}/build
-    if [[ ${WITH_C_API:-OFF} == "OFF" ]] ; then
-        cat <<EOF
+    cat <<EOF
     ========================================
     Generating fluid library for train and inference ...
     ========================================
 EOF
-        cmake .. -DWITH_DISTRIBUTE=OFF -DON_INFER=ON
-        make -j `nproc` fluid_lib_dist
-        make -j `nproc` inference_lib_dist
-      fi
+    cmake .. -DWITH_DISTRIBUTE=OFF -DON_INFER=ON
+    make -j `nproc` fluid_lib_dist
+    make -j `nproc` inference_lib_dist
 }
 
 function tar_fluid_lib() {
-    if [[ ${WITH_C_API:-OFF} == "OFF" ]] ; then
-        cat <<EOF
+    cat <<EOF
     ========================================
     Taring fluid library for train and inference ...
     ========================================
 EOF
-        cd ${PADDLE_ROOT}/build
-        cp -r fluid_install_dir fluid
-        tar -czf fluid.tgz fluid
-        cp -r fluid_inference_install_dir fluid_inference
-        tar -czf fluid_inference.tgz fluid_inference
-      fi
+    cd ${PADDLE_ROOT}/build
+    cp -r fluid_install_dir fluid
+    tar -czf fluid.tgz fluid
+    cp -r fluid_inference_install_dir fluid_inference
+    tar -czf fluid_inference.tgz fluid_inference
 }
 
 function test_fluid_lib() {
-    if [[ ${WITH_C_API:-OFF} == "OFF" ]] ; then
-        cat <<EOF
+    cat <<EOF
     ========================================
     Testing fluid library for inference ...
     ========================================
 EOF
-        cd ${PADDLE_ROOT}/paddle/fluid/inference/api/demo_ci
-        ./run.sh ${PADDLE_ROOT} ${WITH_MKL:-ON} ${WITH_GPU:-OFF} ${INFERENCE_DEMO_INSTALL_DIR} \
-                 ${TENSORRT_INCLUDE_DIR:-/usr/local/TensorRT/include} \
-                 ${TENSORRT_LIB_DIR:-/usr/local/TensorRT/lib}
-        ./clean.sh
-      fi
+    cd ${PADDLE_ROOT}/paddle/fluid/inference/api/demo_ci
+    ./run.sh ${PADDLE_ROOT} ${WITH_MKL:-ON} ${WITH_GPU:-OFF} ${INFERENCE_DEMO_INSTALL_DIR} \
+             ${TENSORRT_INCLUDE_DIR:-/usr/local/TensorRT/include} \
+             ${TENSORRT_LIB_DIR:-/usr/local/TensorRT/lib}
+    ./clean.sh
 }
 
 function main() {
@@ -791,11 +770,6 @@ function main() {
       dockerfile)
         gen_dockerfile ${PYTHON_ABI:-""}
         ;;
-      capi)
-        cmake_gen ${PYTHON_ABI:-""}
-        build
-        gen_capi_package
-        ;;
       fluid_inference_lib)
         cmake_gen ${PYTHON_ABI:-""}
         gen_fluid_lib
@@ -810,7 +784,6 @@ function main() {
         build
         assert_api_not_changed ${PYTHON_ABI:-""}
         run_test
-        gen_capi_package
         gen_fluid_lib
         test_fluid_lib
         assert_api_spec_approvals
@@ -820,7 +793,6 @@ function main() {
         assert_api_spec_approvals
         ;;
       test_inference)
-        gen_capi_package
         gen_fluid_lib
         test_fluid_lib
         ;;
diff --git a/paddle/scripts/paddle_docker_build.sh b/paddle/scripts/paddle_docker_build.sh
index 174c2a12f007b282a5182c0aec9b0a6bec9e55fa..91ca8907c751ea706959a4eff29293a261359a41 100755
--- a/paddle/scripts/paddle_docker_build.sh
+++ b/paddle/scripts/paddle_docker_build.sh
@@ -28,7 +28,6 @@ function start_build_docker() {
         -e WITH_AVX=ON \
         -e WITH_GOLANG=OFF \
         -e WITH_TESTING=ON \
-        -e WITH_C_API=OFF \
         -e WITH_COVERAGE=ON \
         -e COVERALLS_UPLOAD=ON \
         -e WITH_DEB=OFF \
@@ -67,9 +66,6 @@ function main() {
     DOCKER_REPO="paddlepaddle/paddle"
     VERSION="latest-dev"
     PADDLE_ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}")/../../" && pwd )"
-    if [ "$1" == "build_android" ]; then
-        VERSION="latest-dev-android"
-    fi
     IMG=${DOCKER_REPO}:${VERSION}
     start_build_docker $@
 }
diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt
index 37ad77549c88166186d9bf354d6bfd4c6bef152c..59e695e6fcb66cbaed1bcc9e861df81b5f73c1ed 100644
--- a/python/CMakeLists.txt
+++ b/python/CMakeLists.txt
@@ -39,7 +39,6 @@ add_custom_target(copy_paddle_pybind ALL DEPENDS ${FLUID_CORE})
 IF(WIN32)
     add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
             COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_SOURCE_DIR}/python/paddle ${PADDLE_BINARY_DIR}/python/paddle/
-            COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_SOURCE_DIR}/paddle/py_paddle ${PADDLE_BINARY_DIR}/python/
             COMMAND ${CMAKE_COMMAND} -E env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
             COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
             COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python
@@ -48,7 +47,6 @@ ELSE(WIN32)
 	add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
 		COMMAND touch stub.cc
 		COMMAND cp -r ${PADDLE_SOURCE_DIR}/python/paddle ${PADDLE_BINARY_DIR}/python
-		COMMAND cp -r ${PADDLE_SOURCE_DIR}/paddle/py_paddle ${PADDLE_BINARY_DIR}/python/
 		COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
 		COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
 		COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python
diff --git a/python/paddle/fluid/contrib/slim/graph/graph.py b/python/paddle/fluid/contrib/slim/graph/graph.py
index 7d6b0702035d49189c0919f976ea3c0c52663547..f38d9783413a01cd1005a014c0aba5ecf5cc79c2 100644
--- a/python/paddle/fluid/contrib/slim/graph/graph.py
+++ b/python/paddle/fluid/contrib/slim/graph/graph.py
@@ -11,8 +11,12 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
+from __future__ import print_function
+import os
+import subprocess
 from ....framework import Program
+from ....framework import Block
+from .... import core
 
 __all__ = ['Graph', 'ImitationGraph', 'IRGraph']
 
diff --git a/paddle/py_paddle/__init__.py b/python/paddle/fluid/contrib/slim/quantization/__init__.py
similarity index 62%
rename from paddle/py_paddle/__init__.py
rename to python/paddle/fluid/contrib/slim/quantization/__init__.py
index 5504d1d50c523315036bfaaf6641c5216269a5e5..6c26475f48855674d97abf5778a631646734fcf8 100644
--- a/paddle/py_paddle/__init__.py
+++ b/python/paddle/fluid/contrib/slim/quantization/__init__.py
@@ -1,4 +1,4 @@
-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,13 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from util import DataProviderWrapperConverter
-from dataprovider_converter import DataProviderConverter
+from __future__ import print_function
 
-__all__ = [
-    'paddle',
-    'DataProviderConverter',
-    'DataProviderWrapperConverter',  # for deprecated usage.
-    'loadParameterFile'
-]
-util.monkeypatches()
+from . import quantization_pass
+from .quantization_pass import *
+
+__all__ = quantization_pass.__all__
diff --git a/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py
new file mode 100644
index 0000000000000000000000000000000000000000..266a106bc507104c0a8db1c882b55ac59e88195e
--- /dev/null
+++ b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py
@@ -0,0 +1,318 @@
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import collections
+from .... import core
+from ....framework import IrGraph
+from ....framework import Program
+from ....framework import Variable
+from ....initializer import Constant
+from .... import unique_name
+
+__all__ = ['QuantizationTransformPass']
+
+
+class QuantizationTransformPass(object):
+    def __init__(self,
+                 scope=None,
+                 program_exe=None,
+                 weight_bits=8,
+                 activation_bits=8,
+                 activation_quantize_type='abs_max',
+                 weight_quantize_type='abs_max',
+                 window_size=10000):
+        """
+        Convert and rewrite the IrGraph according to weight and
+        activation quantization type.
+        Args:
+            weight_bits (int): quantization bit number for weights,
+                the bias is not quantized.
+            activation_bits (int): quantization bit number for activation.
+            activation_quantize_type (str): quantization type for activation,
+                now support 'abs_max', 'range_abs_max'. If use 'abs_max' mode,
+                the quantization scale will be calculated dynamically each step
+                in both training and testing period. If use 'range_abs_max',
+                a static quantization scale will be calculated during training
+                and used in inference.
+            weight_quantize_type (str): quantization type for weights,
+                support 'abs_max'. The 'range_abs_max' usually is not used for
+                weight, since weights are fixed once the model is well trained.
+            window_size (int): the window size for 'range_abs_max' quantization.
+        Examples:
+        .. code-block:: python
+            # The original graph will be rewrite.
+            import paddle.fluid as fluid
+            from paddle.fluid.contrib.slim.quantization \
+                import QuantizationTransformPass
+            from paddle.fluid.contrib.slim.graph import IrGraph
+            from paddle.fluid import core
+
+            graph = IrGraph(core.Graph(program.desc), for_test=False)
+            exe = fluid.Executor(fluid.CPUPlace())
+            transform_pass = QuantizationTransformPass(fluid.global_scope(),
+            exe)
+            transform_pass.apply(graph)
+        """
+        self._scope = scope
+        self._program_exe = program_exe
+        self._weight_bits = weight_bits
+        self._activation_bits = activation_bits
+
+        quant_type = ['abs_max', 'range_abs_max']
+        if activation_quantize_type not in quant_type:
+            raise ValueError(
+                "Unknown activation_quantize_type : '%s'. It can only be ",
+                "'abs_max' or 'range_abs_max'.", str(activation_quantize_type))
+        if weight_quantize_type not in quant_type:
+            raise ValueError(
+                "Unknown weight_quantize_type: '%s'. It can only be ",
+                "'abs_max' or 'range_abs_max'.", str(weight_quantize_type))
+
+        self._activation_quantize_type = activation_quantize_type
+        self._weight_quantize_type = weight_quantize_type
+        self._window_size = window_size
+
+        self._need_initialized = collections.OrderedDict()
+        self._quantizable_ops = ['conv2d', 'depthwise_conv2d', 'mul']
+        self._quantizable_grad_ops = [
+            '%s_grad' % (op) for op in self._quantizable_ops
+        ]
+        self._fake_quant_op_types = [
+            'fake_quantize_abs_max', 'fake_quantize_range_abs_max'
+        ]
+        self._fake_dequant_op_types = ['fake_dequantize_max_abs']
+        self._is_test = None
+        self._global_step = None
+
+    def apply(self, graph):
+        assert isinstance(graph,
+                          IrGraph), 'graph must be the instance of IrGraph.'
+        self._need_initialized.clear()
+        self._is_test = graph.is_test()
+        # marked the variable which has been dequantized.
+        dequantized_vars = collections.OrderedDict()
+        params = [p.name() for p in graph.all_parameters()]
+
+        def _transform_forward(graph, op):
+            for var_node in op.inputs:
+                if var_node.name() in dequantized_vars:
+                    dequant_var_node = dequantized_vars[var_node.name()]
+                else:
+                    quant_bits = self._weight_bits if var_node.name() in params \
+                    else self._activation_bits
+                    quant_type = self._weight_quantize_type if var_node.name() \
+                        in params else self._activation_quantize_type
+                    quant_var_node, scale_var_node = self._insert_quant_op(
+                        graph, var_node, quant_bits, quant_type)
+                    dequant_var_node = self._insert_dequant_op(
+                        graph, quant_var_node, scale_var_node, quant_bits)
+                    dequantized_vars[var_node.name()] = dequant_var_node
+                graph.update_input_link(var_node, dequant_var_node, op)
+
+        def _transform_backward(graph, op):
+            no_dequanted_input_vars = True
+            for var_node in op.inputs:
+                if var_node.name() in dequantized_vars:
+                    dequant_var_node = dequantized_vars[var_node.name()]
+                    graph.update_input_link(var_node, dequant_var_node, op)
+                    no_dequanted_input_vars = False
+            if no_dequanted_input_vars:
+                raise ValueError("There is no dequanted inputs for op %s." %
+                                 (op.name()))
+
+        if not self._is_test:
+            self._create_global_step(graph)
+        ops = graph.all_ops()
+        # The process of _transform_forward and _transform_backward is needed in two for loops.
+        # The loop for transforming the forward graph:
+        for op in ops:
+            if op.name() in self._quantizable_ops:
+                _transform_forward(graph, op)
+        # The loop for renaming the inputs of backward op.
+        for op in ops:
+            if op.name() in self._quantizable_grad_ops:
+                _transform_backward(graph, op)
+
+        if len(self._need_initialized) > 0:
+            assert self._scope is not None, \
+            'The scope cannot be set None when activation_quantize_type equals to range_abs_max.'
+            assert self._program_exe is not None, \
+            'The program_exe cannot be set None when activation_quantize_type equals to range_abs_max.'
+            init_program = Program()
+            for var_desc, initializer in self._need_initialized.iteritems():
+                var = Variable(init_program.global_block())
+                var._set_desc(var_desc)
+                initializer(var, init_program.global_block())
+            self._program_exe.run(program=init_program, scope=self._scope)
+
+        return graph
+
+    def _create_global_step(self, graph):
+        if self._weight_quantize_type == 'range_abs_max' or \
+                self._activation_quantize_type == 'range_abs_max':
+            counter_name = '@STEP_COUNTER@'
+            for node in graph.all_vars():
+                if node.name() == counter_name:
+                    self._global_step = node
+            if self._global_step is None:
+                global_step_in = graph.create_param_node(
+                    name=counter_name,
+                    var_type=core.VarDesc.VarType.LOD_TENSOR,
+                    shape=[1],
+                    var_dtype=core.VarDesc.VarType.INT64)
+                self._need_initialized[global_step_in.var()] = \
+                    Constant(value=0, force_cpu=True)
+                global_step_out = graph.create_var_node_from_desc(
+                    global_step_in.var())
+                increment_op = graph.create_op_node(
+                    op_type='increment',
+                    attrs={'step': 1.0},
+                    inputs={'X': global_step_in},
+                    outputs={'Out': global_step_out})
+                graph.link_to(global_step_in, increment_op)
+                graph.link_to(increment_op, global_step_out)
+                self._global_step = global_step_out
+
+    def _insert_quant_op(self, graph, var_node, quant_bits, quant_type):
+        """
+        Insert fake_quantize_op in the graph.
+        """
+        if quant_type == 'abs_max':
+            return self._insert_quant_abs_max_op(graph, var_node, quant_bits)
+        elif quant_type == 'range_abs_max':
+            return self._insert_quant_range_abs_max_op(graph, var_node,
+                                                       quant_bits)
+
+    def _insert_quant_abs_max_op(self, graph, var_node, quant_bits):
+        """
+        Insert fake_quantize_abs_max op in the graph.
+        """
+        assert var_node.is_var(), '{} is not a var'.format(var_node.name())
+
+        quant_var_node = graph.create_var_node(
+            name=self._quantized_var_name(var_node.name()),
+            var_type=var_node.var().type(),
+            shape=var_node.var().shape(),
+            var_dtype=var_node.var().dtype())
+        scale_var_node = graph.create_var_node(
+            name=self._quantized_scale_name(var_node.name()),
+            var_type=var_node.var().type(),
+            shape=var_node.var().shape(),
+            var_dtype=var_node.var().dtype())
+        quant_op_node = graph.create_op_node(
+            op_type='fake_quantize_abs_max',
+            attrs={'bit_length': quant_bits},
+            inputs={'X': var_node},
+            outputs={'Out': quant_var_node,
+                     'OutScale': scale_var_node})
+        graph.link_to(var_node, quant_op_node)
+        graph.link_to(quant_op_node, quant_var_node)
+        graph.link_to(quant_op_node, scale_var_node)
+        return quant_var_node, scale_var_node
+
+    def _insert_quant_range_abs_max_op(self, graph, var_node, quant_bits):
+        """
+        Insert fake_quantize_range_abs_max on the graph.
+        """
+        assert var_node.is_var(), '{} is not a var'.format(var_node.name())
+
+        quant_var_node = graph.create_var_node(
+            name=self._quantized_var_name(var_node.name()),
+            var_type=var_node.var().type(),
+            shape=var_node.var().shape(),
+            var_dtype=var_node.var().dtype())
+
+        scale_in_node = graph.create_param_node(
+            name=self._quantized_scale_name(var_node.name()),
+            var_type=core.VarDesc.VarType.LOD_TENSOR,
+            shape=[1],
+            var_dtype=var_node.var().dtype())
+        self._need_initialized[scale_in_node.var()] = Constant(value=0.001)
+
+        scale_out_node = graph.create_var_node_from_desc(scale_in_node.var())
+        inputs = {'X': var_node, 'InScale': scale_in_node}
+        outputs = {'Out': quant_var_node, 'OutScale': scale_out_node}
+
+        if not self._is_test:
+            # The name of scales_var_node maybe 'scales_0', 'scales_1', etc.
+            scales_node = graph.create_param_node(
+                name=unique_name.generate('scales'),
+                var_type=core.VarDesc.VarType.LOD_TENSOR,
+                shape=[self._window_size],
+                var_dtype=var_node.var().dtype())
+            self._need_initialized[scales_node.var()] = Constant(value=0)
+            inputs['Iter'] = self._global_step
+            outputs['OutScales'] = scales_node
+        attrs = {
+            'window_size': self._window_size,
+            'bit_length': quant_bits,
+            'is_test': self._is_test
+        }
+        quant_op_node = graph.create_op_node(
+            op_type='fake_quantize_range_abs_max',
+            attrs=attrs,
+            inputs=inputs,
+            outputs=outputs)
+
+        graph.link_to(var_node, quant_op_node)
+        graph.link_to(scale_in_node, quant_op_node)
+        graph.link_to(quant_op_node, quant_var_node)
+        graph.link_to(quant_op_node, scale_out_node)
+
+        if not self._is_test:
+            graph.link_to(self._global_step, quant_op_node)
+            graph.link_to(quant_op_node, scales_node)
+
+        return quant_var_node, scale_out_node
+
+    def _insert_dequant_op(self, graph, var_node, scale_var_node, quant_bits):
+        """
+        Insert fake_dequantize_op in the graph.
+        """
+        assert var_node.is_var(), '{} is not a var'.format(var_node.name())
+
+        dequant_var_node = graph.create_var_node(
+            name=self._dequantized_var_name(var_node.name()),
+            var_type=var_node.var().type(),
+            shape=var_node.var().shape(),
+            var_dtype=var_node.var().dtype())
+        max_range = (1 << (quant_bits - 1)) - 1
+        dequant_op_node = graph.create_op_node(
+            op_type='fake_dequantize_max_abs',
+            attrs={'max_range': float(max_range)},
+            inputs={'X': var_node,
+                    'Scale': scale_var_node},
+            outputs={'Out': dequant_var_node})
+        graph.link_to(var_node, dequant_op_node)
+        graph.link_to(scale_var_node, dequant_op_node)
+        graph.link_to(dequant_op_node, dequant_var_node)
+        return dequant_var_node
+
+    def _quantized_var_name(self, var_name):
+        """
+        Return quantized variable name for the input `var_name`.
+        """
+        return "%s.quantized" % (var_name)
+
+    def _dequantized_var_name(self, var_name):
+        """
+        Return dequantized variable name for the input `var_name`.
+        """
+        return "%s.dequantized" % (var_name)
+
+    def _quantized_scale_name(self, var_name):
+        """
+        Return the scale name of quantized variable for the input `var_name`.
+        """
+        return "%s.scale" % (var_name)
diff --git a/python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py b/python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py
new file mode 100644
index 0000000000000000000000000000000000000000..1bd4b95d6b90b7f16d507061190f0b463f6c4cc5
--- /dev/null
+++ b/python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py
@@ -0,0 +1,175 @@
+#   copyright (c) 2018 paddlepaddle authors. all rights reserved.
+#
+# licensed under the apache license, version 2.0 (the "license");
+# you may not use this file except in compliance with the license.
+# you may obtain a copy of the license at
+#
+#     http://www.apache.org/licenses/license-2.0
+#
+# unless required by applicable law or agreed to in writing, software
+# distributed under the license is distributed on an "as is" basis,
+# without warranties or conditions of any kind, either express or implied.
+# see the license for the specific language governing permissions and
+# limitations under the license.
+
+import unittest
+import random
+import numpy as np
+import paddle.fluid as fluid
+import six
+from paddle.fluid.framework import Program
+from paddle.fluid.framework import IrGraph
+from paddle.fluid.contrib.slim.quantization import QuantizationTransformPass
+from paddle.fluid import core
+
+
+def linear_fc(num):
+    data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32')
+    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+    hidden = data
+    for _ in six.moves.xrange(num):
+        hidden = fluid.layers.fc(hidden, size=128, act='relu')
+    loss = fluid.layers.cross_entropy(input=hidden, label=label)
+    loss = fluid.layers.mean(loss)
+    return loss
+
+
+def residual_block(num):
+    def conv_bn_layer(input,
+                      ch_out,
+                      filter_size,
+                      stride,
+                      padding,
+                      act='relu',
+                      bias_attr=False):
+        tmp = fluid.layers.conv2d(
+            input=input,
+            filter_size=filter_size,
+            num_filters=ch_out,
+            stride=stride,
+            padding=padding,
+            act=None,
+            bias_attr=bias_attr)
+        return fluid.layers.batch_norm(input=tmp, act=act)
+
+    data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32')
+    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+    hidden = data
+    for _ in six.moves.xrange(num):
+        conv = conv_bn_layer(hidden, 16, 3, 1, 1, act=None, bias_attr=True)
+        short = conv_bn_layer(hidden, 16, 1, 1, 0, act=None)
+        hidden = fluid.layers.elementwise_add(x=conv, y=short, act='relu')
+    fc = fluid.layers.fc(input=hidden, size=10)
+    loss = fluid.layers.cross_entropy(input=fc, label=label)
+    loss = fluid.layers.mean(loss)
+    return loss
+
+
+class TestQuantizationTransformPass(unittest.TestCase):
+    def setUp(self):
+        self.quantizable_op_and_inputs = {
+            'conv2d': ['Input', 'Filter'],
+            'depthwise_conv2d': ['Input', 'Filter'],
+            'mul': ['X', 'Y']
+        }
+        self.quantizable_grad_op_inputs = {
+            'conv2d_grad': ['Input', 'Filter'],
+            'depthwise_conv2d_grad': ['Input', 'Filter'],
+            'mul_grad': ['X', 'Y']
+        }
+
+    def check_program(self, transform_pass, program):
+        quantized_ops = set()
+        for block in program.blocks:
+            for op in block.ops:
+                # check forward
+                if op.type in self.quantizable_op_and_inputs:
+                    for arg_name in op.input_arg_names:
+                        self.assertTrue(
+                            arg_name.endswith('.quantized.dequantized'))
+                        quantized_ops.add(arg_name)
+
+            for op in block.ops:
+                # check backward
+                if op.type in self.quantizable_grad_op_inputs:
+                    for pname in self.quantizable_grad_op_inputs[op.type]:
+                        arg_name = op.input(pname)[0]
+                        self.assertTrue(
+                            arg_name.endswith('.quantized.dequantized'))
+                        self.assertTrue(arg_name in quantized_ops)
+
+    def linear_fc_quant(self, quant_type):
+        main = fluid.Program()
+        startup = fluid.Program()
+        with fluid.program_guard(main, startup):
+            loss = linear_fc(3)
+            opt = fluid.optimizer.Adam(learning_rate=0.001)
+            opt.minimize(loss)
+        exe = fluid.Executor(fluid.CPUPlace())
+        graph = IrGraph(core.Graph(main.desc), for_test=False)
+        transform_pass = QuantizationTransformPass(
+            scope=fluid.global_scope(),
+            program_exe=exe,
+            activation_quantize_type=quant_type)
+        transform_pass.apply(graph)
+        marked_nodes = set()
+        for op in graph.all_ops():
+            if op.name().find('quantize') > -1:
+                marked_nodes.add(op)
+        graph.draw('.', 'quantize_fc_' + quant_type, marked_nodes)
+        program = graph.to_program()
+        self.check_program(transform_pass, program)
+        val_graph = IrGraph(core.Graph(program.desc), for_test=False)
+        val_marked_nodes = set()
+        for op in val_graph.all_ops():
+            if op.name().find('quantize') > -1:
+                val_marked_nodes.add(op)
+        val_graph.draw('.', 'val_fc_' + quant_type, val_marked_nodes)
+
+    def test_linear_fc_quant_abs_max(self):
+        self.act_quant_op_type = 'fake_quantize_abs_max'
+        self.linear_fc_quant('abs_max')
+
+    def test_linear_fc_quant_range_abs_max(self):
+        self.act_quant_op_type = 'fake_quantize_range_abs_max'
+        self.linear_fc_quant('range_abs_max')
+
+    def residual_block_quant(self, quant_type):
+        main = fluid.Program()
+        startup = fluid.Program()
+        with fluid.program_guard(main, startup):
+            loss = residual_block(2)
+            opt = fluid.optimizer.Adam(learning_rate=0.001)
+            opt.minimize(loss)
+        exe = fluid.Executor(fluid.CPUPlace())
+        graph = IrGraph(core.Graph(main.desc), for_test=False)
+        transform_pass = QuantizationTransformPass(
+            scope=fluid.global_scope(),
+            program_exe=exe,
+            activation_quantize_type=quant_type)
+        transform_pass.apply(graph)
+        marked_nodes = set()
+        for op in graph.all_ops():
+            if op.name().find('quantize') > -1:
+                marked_nodes.add(op)
+        graph.draw('.', 'quantize_residual_' + quant_type, marked_nodes)
+        program = graph.to_program()
+        self.check_program(transform_pass, program)
+        val_graph = IrGraph(core.Graph(program.desc), for_test=False)
+        val_marked_nodes = set()
+        for op in val_graph.all_ops():
+            if op.name().find('quantize') > -1:
+                val_marked_nodes.add(op)
+        val_graph.draw('.', 'val_residual_' + quant_type, val_marked_nodes)
+
+    def test_residual_block_abs_max(self):
+        self.act_quant_op_type = 'fake_quantize_abs_max'
+        self.residual_block_quant('abs_max')
+
+    def test_residual_block_range_abs_max(self):
+        self.act_quant_op_type = 'fake_quantize_range_abs_max'
+        self.residual_block_quant('range_abs_max')
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py
index 20aa6054fe4b7d6a1b1454292954237bdfbe045e..d3ff14a17955990bff851e95bd61fbc370ea7aa5 100644
--- a/python/paddle/fluid/executor.py
+++ b/python/paddle/fluid/executor.py
@@ -305,7 +305,9 @@ class Executor(object):
     def __init__(self, place):
         self.place = place
         self.program_caches = dict()
-        self.executor = None
+        p = core.Place()
+        p.set_place(self.place)
+        self._default_executor = core.Executor(p)
         self._closed = False
 
     def _get_program_cache(self, program_cache_key):
@@ -397,12 +399,13 @@ class Executor(object):
             >>> ...
             >>> exe.close()
         """
-        if not self._closed and self.executor:
-            self.executor.close()
+        if not self._closed:
+            self._default_executor.close()
             self._closed = True
 
     def _run_parallel(self, program, scope, feed, fetch_list, fetch_var_name,
                       return_numpy):
+        exe = program._executor
         if isinstance(feed, dict):
             feed_tensor_dict = dict()
             for feed_name in feed:
@@ -414,8 +417,7 @@ class Executor(object):
                     feed_tensor.set(feed[feed_name], core.CPUPlace())
                 feed_tensor_dict[feed_name] = feed_tensor
 
-            self.executor.feed_and_split_tensor_into_local_scopes(
-                feed_tensor_dict)
+            exe.feed_and_split_tensor_into_local_scopes(feed_tensor_dict)
         elif isinstance(feed, list) or isinstance(feed, tuple):
             if len(feed) != len(program._places):
                 raise ValueError(
@@ -436,10 +438,10 @@ class Executor(object):
                         tensor = tmp
                     res_dict[feed_name] = tensor
                 res.append(res_dict)
-            self.executor.feed_tensors_into_local_scopes(res)
+            exe.feed_tensors_into_local_scopes(res)
 
         fetch_var_names = list(map(_to_name_str, fetch_list))
-        self.executor.run(fetch_var_names, fetch_var_name)
+        exe.run(fetch_var_names, fetch_var_name)
         arr = scope.find_var(fetch_var_name).get_lod_tensor_array()
 
         if return_numpy:
@@ -511,12 +513,9 @@ class Executor(object):
         compiled = isinstance(program, compiler.CompiledProgram)
         # For backward compatibility, run directly.
         if not compiled:
-            if not self.executor:
-                p = core.Place()
-                p.set_place(self.place)
-                self.executor = core.Executor(p)
             return self._run(
                 program,
+                self._default_executor,
                 feed=feed,
                 fetch_list=fetch_list,
                 feed_var_name=feed_var_name,
@@ -526,7 +525,6 @@ class Executor(object):
                 use_program_cache=use_program_cache)
 
         program._compile(scope, self.place)
-        self.executor = program._executor
         if program._is_data_parallel:
             return self._run_parallel(
                 program,
@@ -536,12 +534,13 @@ class Executor(object):
                 fetch_var_name=fetch_var_name,
                 return_numpy=return_numpy)
         elif program._is_inference:
-            return self._run_inference(program, feed)
+            return self._run_inference(program._executor, feed)
         else:
             # TODO(panyx0718): Can compile program to optimize executor
             # performance.
             return self._run(
                 program._program,
+                self._default_executor,
                 feed=feed,
                 fetch_list=fetch_list,
                 feed_var_name=feed_var_name,
@@ -550,8 +549,8 @@ class Executor(object):
                 return_numpy=return_numpy,
                 use_program_cache=use_program_cache)
 
-    def _run(self, program, feed, fetch_list, feed_var_name, fetch_var_name,
-             scope, return_numpy, use_program_cache):
+    def _run(self, program, exe, feed, fetch_list, feed_var_name,
+             fetch_var_name, scope, return_numpy, use_program_cache):
 
         if feed is None:
             feed = {}
@@ -589,11 +588,11 @@ class Executor(object):
                 fetch_var_name=fetch_var_name)
 
         self._feed_data(program, feed, feed_var_name, scope)
-        self.executor.run(program.desc, scope, 0, True, True)
+        exe.run(program.desc, scope, 0, True, True)
         outs = self._fetch_data(fetch_list, fetch_var_name, scope)
         if return_numpy:
             outs = as_numpy(outs)
         return outs
 
-    def _run_inference(self, program, feed):
-        return self.executor.run(feed)
+    def _run_inference(self, exe, feed):
+        return exe.run(feed)
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index 77239f2e8ee5243611352d275c0d595670e24185..fc5e471ae3041b0245c873d85efa8b49f3a43678 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -23,6 +23,7 @@ import traceback
 import six
 
 import numpy as np
+import subprocess
 
 from .. import compat as cpt
 from .proto import framework_pb2
@@ -1512,6 +1513,154 @@ class Block(object):
         return ret_var
 
 
+class IrGraph(object):
+    """
+    IrGraph uses core.Graph as the delegation to accomplish the manipulation.
+    """
+
+    def __init__(self, graph, for_test=False):
+        """
+        Construct the IrGraph using core.Graph.
+        Args:
+            graph(core.Graph): C++ Graph.
+            for_test(bool): True for the test graph and false for the train graph.
+        """
+        assert isinstance(
+            graph, core.Graph), 'graph must be the instance of core.Graph.'
+        self.graph = graph
+        self._for_test = for_test
+
+    def is_test(self):
+        return self._for_test
+
+    def all_parameters(self):
+        param_nodes = set()
+        for node in self.graph.nodes():
+            if node.is_var() and node.var() is not None and node.var(
+            ).persistable():
+                param_nodes.add(node)
+        return param_nodes
+
+    def all_vars(self):
+        return {node for node in self.graph.nodes() if node.is_var()}
+
+    def all_ops(self):
+        return {node for node in self.graph.nodes() if node.is_op()}
+
+    def create_param_node(self, name, var_type, shape, var_dtype):
+        var_desc = core.VarDesc(name)
+        var_desc.set_type(var_type)
+        var_desc.set_shape(shape)
+        var_desc.set_dtype(var_dtype)
+        var_desc.set_persistable(True)
+        return self.graph.create_var_node(var_desc)
+
+    def create_var_node(self, name, var_type, shape, var_dtype):
+        var_desc = core.VarDesc(name)
+        var_desc.set_type(var_type)
+        var_desc.set_shape(shape)
+        var_desc.set_dtype(var_dtype)
+        return self.graph.create_var_node(var_desc)
+
+    def create_var_node_from_desc(self, var_desc):
+        return self.graph.create_var_node(var_desc)
+
+    def create_op_node(self, op_type, attrs, inputs, outputs):
+        op_desc = core.OpDesc()
+        op_desc.set_type(op_type)
+        for attr, value in attrs.iteritems():
+            self._update_desc_attr(op_desc, attr, value)
+        for input_name, var_nodes in inputs.iteritems():
+            if not isinstance(var_nodes, list):
+                var_nodes = [var_nodes]
+            op_desc.set_input(input_name,
+                              [var_node.name() for var_node in var_nodes])
+        for output_name, var_nodes in outputs.iteritems():
+            if not isinstance(var_nodes, list):
+                var_nodes = [var_nodes]
+            op_desc.set_output(output_name,
+                               [var_node.name() for var_node in var_nodes])
+        return self.graph.create_op_node(op_desc)
+
+    def create_op_node_from_desc(self, op_desc):
+        return self.graph.create_op_node(op_desc)
+
+    def update_input_link(self, old_input_node, new_input_node, op_node):
+        assert old_input_node in self.graph.nodes() and new_input_node in self.graph.nodes() and \
+            op_node in self.graph.nodes(), 'Th three arguments must be in the graph nodes.'
+        old_input_node.outputs_remove(op_node)
+        op_node.inputs_remove(old_input_node)
+        new_input_node.outputs_append(op_node)
+        op_node.inputs_append(new_input_node)
+        op_node.op()._rename_input(old_input_node.name(), new_input_node.name())
+
+    def link_to(self, node_in, node_out):
+        assert node_in in self.graph.nodes() and node_out in self.graph.nodes(), \
+            'Th two arguments must be in the graph nodes.'
+        node_in.outputs_append(node_out)
+        node_out.inputs_append(node_in)
+
+    def safe_remove_nodes(self, remove_nodes):
+        if not isinstance(remove_nodes, set):
+            remove_nodes = set(remove_nodes)
+        core.graph_safe_remove_nodes(self.graph, remove_nodes)
+
+    def draw(self, save_path, name, marked_nodes=None):
+        def _convert_to_pdf(dot_file_path):
+            pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf'
+            exited_code = subprocess.call('dot -Tpdf ' + dot_file_path \
+                            + ' -o ' + pdf_save_path, shell=True)
+            if exited_code != 0:
+                print('The dot command is needed for creating pdf files.')
+                print('The {} is saved as the dot filetype.'.format(
+                    dot_file_path))
+
+        remove_ctr_vars = set()
+        ops_num = 0
+        for node in self.graph.nodes():
+            if node.is_ctrl_var():
+                remove_ctr_vars.add(node)
+            elif node.is_op():
+                ops_num += 1
+        print('Total ops num = {}.'.format(ops_num))
+        self.safe_remove_nodes(remove_ctr_vars)
+        if marked_nodes is not None:
+            if not isinstance(marked_nodes, set):
+                marked_nodes = set(marked_nodes)
+            marked_nodes = marked_nodes - remove_ctr_vars
+            if self.graph.has('__graphviz__marked_node__'):
+                self.graph.erase('__graphviz__marked_node__')
+            self.graph.set('__graphviz__marked_node__', marked_nodes)
+        viz_dot_path = os.path.join(save_path, name) + '.dot'
+        viz_pass = core.get_pass('graph_viz_pass')
+        viz_pass.set('graph_viz_path', viz_dot_path)
+        viz_pass.apply(self.graph)
+        _convert_to_pdf(viz_dot_path)
+
+    def to_program(self):
+        convert_pass = core.get_pass('graph_to_program_pass')
+        convert_pass.set('program', Program().desc)
+        convert_pass.apply(self.graph)
+        desc = convert_pass.get_program('program')
+        program = Program._construct_from_desc(desc)
+        return program
+
+    def _update_desc_attr(self, desc, name, val):
+        """
+        Update the value of desc's attribute by attribute's name.
+        """
+        if isinstance(val, Block):
+            desc.set_block_attr(name, val.desc)
+        elif isinstance(val, list) and val and all(
+                isinstance(v, Block) for v in val):
+            desc.set_blocks_attr(name, [v.desc for v in val])
+        elif isinstance(val, core.BlockDesc) or \
+                isinstance(val, core.ProgramDesc):
+            desc.set_serialized_attr(name, val.serialize_to_string())
+        else:
+            desc._set_attr(name, val)
+
+
 class Program(object):
     """
     Python Program. Beneath it is a ProgramDesc, which is used for
@@ -1936,6 +2085,23 @@ class Program(object):
         p._sync_with_cpp()
         return p
 
+    @staticmethod
+    def _construct_from_desc(desc):
+        """
+        Construct a program from program desc.
+
+        Args:
+            desc(core.ProgramDesc): The program desc for constructing.
+
+        Returns:
+            Program: A program.
+        """
+        p = Program()
+        p.desc = desc
+        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
+        p._sync_with_cpp()
+        return p
+
     @property
     def random_seed(self):
         """
diff --git a/python/paddle/fluid/tests/book/test_word2vec.py b/python/paddle/fluid/tests/book/test_word2vec.py
index 48cb7789276479d905b1c78f6ae98c5346de14b7..487a29c8391231471737a25d521770ebbca18673 100644
--- a/python/paddle/fluid/tests/book/test_word2vec.py
+++ b/python/paddle/fluid/tests/book/test_word2vec.py
@@ -220,9 +220,7 @@ def infer(use_cuda, save_dirname=None):
         np_data = np.array(results[0])
         infer_out = infer_outputs[0].data.float_data()
         for a, b in zip(np_data[0], infer_out):
-            g_a = float("{:.6g}".format(a))
-            g_b = float("{:.6g}".format(b))
-            assert g_a == g_b
+            assert np.isclose(a, b), "a: {}, b: {}".format(a, b)
 
 
 def main(use_cuda, is_sparse, is_parallel):
diff --git a/python/paddle/fluid/tests/unittests/test_dist_base.py b/python/paddle/fluid/tests/unittests/test_dist_base.py
index 03be3fe84b022a68f358bb9ab9194020aa9f9e26..2789a07beefa02964b78e3e5dccebff9976f64c4 100644
--- a/python/paddle/fluid/tests/unittests/test_dist_base.py
+++ b/python/paddle/fluid/tests/unittests/test_dist_base.py
@@ -123,7 +123,7 @@ class TestDistRunnerBase(object):
             pass_builder = build_stra._finalize_strategy_and_create_passes()
             mypass = pass_builder.insert_pass(
                 len(pass_builder.all_passes()) - 3, "multi_batch_merge_pass")
-            mypass.set_int("num_repeats", args.batch_merge_repeat)
+            mypass.set("num_repeats", args.batch_merge_repeat)
 
         if args.update_method == "nccl2":
             build_stra.num_trainers = len(args.endpoints.split(","))
diff --git a/python/paddle/fluid/tests/unittests/test_pass_builder.py b/python/paddle/fluid/tests/unittests/test_pass_builder.py
index 8c9e489e02839e25cfabe14c16bfd91a908bd734..7e1c2572f08598b8b600517e4a82b48ca71cc20d 100644
--- a/python/paddle/fluid/tests/unittests/test_pass_builder.py
+++ b/python/paddle/fluid/tests/unittests/test_pass_builder.py
@@ -111,7 +111,7 @@ class TestPassBuilder(unittest.TestCase):
 
         pass_builder.remove_pass(len(pass_builder.all_passes()) - 1)
         self.assertEqual(origin_len + 1, len(pass_builder.all_passes()))
-        viz_pass.set_str("graph_viz_path", "/tmp/test_viz_pass")
+        viz_pass.set("graph_viz_path", "/tmp/test_viz_pass")
 
         self.check_network_convergence(
             use_cuda=core.is_compiled_with_cuda(),
diff --git a/python/setup.py.in b/python/setup.py.in
index e00c88b3a6e49bd806bcc6125fa8dc0f69928227..fb4b273a0676fcbcb4402eaf54ddf73d37a2754f 100644
--- a/python/setup.py.in
+++ b/python/setup.py.in
@@ -113,6 +113,7 @@ packages=['paddle',
           'paddle.fluid.contrib.slim.core',
           'paddle.fluid.contrib.slim.graph',
           'paddle.fluid.contrib.slim.prune',
+          'paddle.fluid.contrib.slim.quantization',
           'paddle.fluid.contrib.utils',
           'paddle.fluid.transpiler',
           'paddle.fluid.transpiler.details']