diff --git a/cmake/generic.cmake b/cmake/generic.cmake
index 6839abc1a7bce319ec63b8c87588a9e378aa71ea..11c1f677ae5b308558b54bf49caf168cf6023444 100644
--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@@ -296,7 +296,7 @@ function(go_library TARGET_NAME)
     COMMAND rm -rf ${PADDLE_IN_GOPATH}                                                                                                                                         
     COMMAND ln -sf ${CMAKE_SOURCE_DIR} ${PADDLE_IN_GOPATH}
     # Automatically get all dependencies specified in the source code                                                                                                                                 
-    COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} get -d ./..
+    COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} get -d ./...
     # Golang build source code
     COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} build ${BUILD_MODE}
     -o "${CMAKE_CURRENT_BINARY_DIR}/${LIB_NAME}"
diff --git a/paddle/framework/variable.h b/paddle/framework/variable.h
index b33e10e6820129a874f5355d14d8a3e990186025..72c4a7a2a1d1cf93a784f24e687727ee8481484c 100644
--- a/paddle/framework/variable.h
+++ b/paddle/framework/variable.h
@@ -25,21 +25,24 @@ class Variable {
  public:
   template <typename T>
   const T& Get() const {
-    PADDLE_ASSERT(holder_ != nullptr);
-    PADDLE_ASSERT(std::type_index(typeid(T)) ==
-                  std::type_index(holder_->Type()));
+    PADDLE_ASSERT(IsType<T>());
     return *static_cast<const T*>(holder_->Ptr());
   }
 
   template <typename T>
   T* GetMutable() {
-    if (holder_ == nullptr ||
-        std::type_index(typeid(T)) != std::type_index(holder_->Type())) {
+    if (!IsType<T>()) {
       holder_.reset(new PlaceholderImpl<T>(new T()));
     }
     return static_cast<T*>(holder_->Ptr());
   }
 
+  template <typename T>
+  bool IsType() const {
+    return holder_ != nullptr &&
+           std::type_index(typeid(T)) == std::type_index(holder_->Type());
+  }
+
  private:
   struct Placeholder {
     virtual ~Placeholder() {}
diff --git a/paddle/gserver/layers/Layer.cpp b/paddle/gserver/layers/Layer.cpp
index 125aaf947f3c9d976b117667d1d1b7700a029cc6..4b92b5d163ad107c0783beae45f8c936112fcccf 100644
--- a/paddle/gserver/layers/Layer.cpp
+++ b/paddle/gserver/layers/Layer.cpp
@@ -191,6 +191,11 @@ void Layer::addOutputArgument(int deviceId) {
 void Layer::copyOutputToOtherDevice() {
   for (size_t i = 0; i != outputOtherDevice_.size(); i++) {
     SetDevice device(outputOtherDevice_[i].deviceId);
+    // If outputOtherDevice_[i].value is a CpuMatrix,
+    // the copyFrom is a synchronous interface.
+    // If outputOtherDevice_[i].value is a GpuMatrix, since subsequent
+    // calculations are all on HPPL_STREAM_DEFAULT,
+    // copyFrom can be an asynchronous interface.
     outputOtherDevice_[i].value->copyFrom(*getOutputValue(),
                                           HPPL_STREAM_DEFAULT);
     outputOtherDevice_[i].sequenceStartPositions =
diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp
index c910146164ebfb0737583c72c48ce6dbc5b49939..4431d613f655c1d0c8da13bb5ac9225980c650ad 100644
--- a/paddle/math/Matrix.cpp
+++ b/paddle/math/Matrix.cpp
@@ -1565,6 +1565,8 @@ void CpuMatrix::copyFrom(const Matrix& src, hl_stream_t stream) {
                     const_cast<real*>(src.getData()),
                     sizeof(real) * elementCnt_,
                     stream);
+    // There is a need to add synchronization to ensure that the data is copied.
+    hl_stream_synchronize(stream);
   } else if (typeid(src) == typeid(CpuMatrix)) {
     memcpy(data_, src.getData(), sizeof(real) * elementCnt_);
   } else {
diff --git a/paddle/math/Matrix.h b/paddle/math/Matrix.h
index 748be850b4c902d1b48c1dafbb0d5ea2bf197e6e..7dfd593225065e18830b2b0c0ce854fe7a2d5178 100644
--- a/paddle/math/Matrix.h
+++ b/paddle/math/Matrix.h
@@ -239,7 +239,8 @@ public:
     LOG(FATAL) << "Not implemented";
   }
 
-  // asynchronous copy
+  // For GpuMatrix this is an asynchronous copy interface
+  // For CpuMatrix this is an synchronous copy interface
   virtual void copyFrom(const Matrix& src, hl_stream_t stream) {
     LOG(FATAL) << "Not implemented";
   }
diff --git a/paddle/math/Vector.cpp b/paddle/math/Vector.cpp
index c519ca500afb1dbfdff6e8d211786f4e18ccf1fd..eb87ee9bb7936d27c0c32a1a4b35ff49871c0a10 100644
--- a/paddle/math/Vector.cpp
+++ b/paddle/math/Vector.cpp
@@ -657,6 +657,8 @@ void CpuVectorT<T>::copyFrom(const VectorT<T>& src, hl_stream_t stream) {
                     (void*)src.getData(),
                     sizeof(T) * this->getSize(),
                     stream);
+    // There is a need to add synchronization to ensure that the data is copied.
+    hl_stream_synchronize(stream);
   } else {
     src.copyTo(this);
   }
diff --git a/paddle/math/Vector.h b/paddle/math/Vector.h
index 9af6e30c9e13895ad95653a787ec1c1ad77a248f..80b9775fccf10c57bb48145ef56165ec7c86d8b8 100644
--- a/paddle/math/Vector.h
+++ b/paddle/math/Vector.h
@@ -168,11 +168,11 @@ public:
   virtual void copyFrom(const VectorT<T>& src) = 0;
 
   /**
-   * If use_gpu, this function will push the copy-task to the specifed-stream
-   * and return immediately.
+   * If GpuVector, this function is an asynchronous interface,
+   * will push the copy-task to the specifed-stream and return immediately.
    *
-   * If not use GPU, this function is same as
-   * the copyFrom(const VectorT<T>& src), which use stream HPPL_STREAM_DEFAULT.
+   * If CpuVector, this function is an synchronous interface,
+   * same as the copyFrom(const VectorT<T>& src).
    */
   virtual void copyFrom(const VectorT<T>& src, hl_stream_t stream) = 0;
 
diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp
index 5a0dffe086c4e265d17c79dba435b66c0873e3c7..354f58df39365410ff9aec2576c768e58db9e0d2 100644
--- a/paddle/math/tests/test_matrixCompare.cpp
+++ b/paddle/math/tests/test_matrixCompare.cpp
@@ -1127,4 +1127,18 @@ TEST(Matrix, MaxOutFwdBwd) {
   }
 }
 
+TEST(CpuMatrix, copyFrom) {
+  const size_t height = 1000;
+  const size_t width = 1000;
+  CpuMatrix cpu(height, width);
+  GpuMatrix gpu(height, width);
+  CpuMatrix copy(height, width);
+
+  cpu.randomizeUniform();
+  gpu.copyFrom(cpu);
+  copy.copyFrom(gpu, HPPL_STREAM_DEFAULT);
+
+  TensorCheckEqual(cpu, copy);
+}
+
 #endif
diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh
index 2b48e4dc0f875be9a87797fa14885926999a5010..a182e5f4aef9de8c6f20681328d5ba6c0e6944ef 100644
--- a/paddle/scripts/docker/build.sh
+++ b/paddle/scripts/docker/build.sh
@@ -31,6 +31,7 @@ Configuring cmake in /paddle/build ...
       -DWITH_DOC=OFF
       -DWITH_GPU=${WITH_GPU:-OFF}
       -DWITH_AVX=${WITH_AVX:-OFF}
+      -DWITH_GOLANG=${WITH_GOLANG:-OFF}
       -DWITH_SWIG_PY=ON
       -DCUDNN_ROOT=/usr/
       -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF}
@@ -43,6 +44,7 @@ cmake .. \
       -DWITH_DOC=OFF \
       -DWITH_GPU=${WITH_GPU:-OFF} \
       -DWITH_AVX=${WITH_AVX:-OFF} \
+      -DWITH_GOLANG=${WITH_GOLANG:-OFF} \
       -DWITH_SWIG_PY=ON \
       -DCUDNN_ROOT=/usr/ \
       -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} \