diff --git a/cmake/external/snappy.cmake b/cmake/external/snappy.cmake
index 80282329c6ac65fbd1493a6838efca4bd9cadaad..af09ed4d5d6e21cc50aba5198a7e9ea56f49451a 100644
--- a/cmake/external/snappy.cmake
+++ b/cmake/external/snappy.cmake
@@ -47,8 +47,6 @@ ExternalProject_Add(
                      -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPY_INSTALL_DIR}/lib
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
-    BUILD_COMMAND   make -j8
-    INSTALL_COMMAND make install
 )
 
 add_library(snappy STATIC IMPORTED GLOBAL)
diff --git a/cmake/external/snappystream.cmake b/cmake/external/snappystream.cmake
index 20a96430823d07a07d4bb4602e7fc0cfe55c3bf2..6df636d7fa8757ade73892bda03a80ba9767472b 100644
--- a/cmake/external/snappystream.cmake
+++ b/cmake/external/snappystream.cmake
@@ -46,8 +46,6 @@ ExternalProject_Add(
                         -DCMAKE_INSTALL_PREFIX:PATH=${SNAPPYSTREAM_INSTALL_DIR}
                         -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPYSTREAM_INSTALL_DIR}/lib
                         -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
-        BUILD_COMMAND   make -j8
-        INSTALL_COMMAND make install
         DEPENDS snappy
 )
 
diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake
index 06a7ae56827d5afe857ed0a98092210917a52430..807a48a41f72f17944dc1be5b793b0ca7d70c527 100644
--- a/cmake/inference_lib.cmake
+++ b/cmake/inference_lib.cmake
@@ -148,4 +148,10 @@ copy(string_lib
   DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
 )
 
+set(module "pybind")
+copy(pybind_lib
+  SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h
+  DSTS ${dst_dir}/${module}
+)
+
 add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) 
diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt
index 8b1ca5e16548334ed0c9a6d31b88e0805304579e..d722eec1892206ac44c49e7a12d92be0c54df8c0 100644
--- a/paddle/CMakeLists.txt
+++ b/paddle/CMakeLists.txt
@@ -24,6 +24,6 @@ if(NOT WITH_FLUID_ONLY)
 endif()
 
 add_subdirectory(testing)
-if(NOT MOBILE_INFERENCE AND NOT RPI)
+if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API)
   add_subdirectory(fluid)
 endif()
diff --git a/paddle/fluid/framework/data_type.cc b/paddle/fluid/framework/data_type.cc
index b9c90cb0c32f337ba82ce1eaa5b43199540491ef..b6b93cf422a60c1d8e9cb8b477efd562f9fe4758 100644
--- a/paddle/fluid/framework/data_type.cc
+++ b/paddle/fluid/framework/data_type.cc
@@ -58,6 +58,7 @@ static DataTypeMap* InitDataTypeMap() {
   RegType(bool, proto::VarType::BOOL);
   RegType(size_t, proto::VarType::SIZE_T);
   RegType(int16_t, proto::VarType::INT16);
+  RegType(uint8_t, proto::VarType::UINT8);
 
 #undef RegType
   return retv;
diff --git a/paddle/fluid/framework/data_type.h b/paddle/fluid/framework/data_type.h
index 4b9f572ec5f1cda71c8b8dd8fae54b42e9f16f7a..491413db8c8d66fd907801131e89d9303bdef9f2 100644
--- a/paddle/fluid/framework/data_type.h
+++ b/paddle/fluid/framework/data_type.h
@@ -47,8 +47,14 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
     case proto::VarType::BOOL:
       visitor.template operator()<bool>();
       break;
+    case proto::VarType::UINT8:
+      visitor.template operator()<uint8_t>();
+      break;
+    case proto::VarType::INT16:
+      visitor.template operator()<int16_t>();
+      break;
     default:
-      PADDLE_THROW("Not supported");
+      PADDLE_THROW("Not supported %d", type);
   }
 }
 
diff --git a/paddle/fluid/framework/details/fetch_op_handle.cc b/paddle/fluid/framework/details/fetch_op_handle.cc
index b1c9dd0d15223f7d1bf6ea44144589f1de927e3e..224e8e1f6efd7a894591ac51c929517cae7539ce 100644
--- a/paddle/fluid/framework/details/fetch_op_handle.cc
+++ b/paddle/fluid/framework/details/fetch_op_handle.cc
@@ -48,17 +48,18 @@ void FetchOpHandle::RunImpl() {
   WaitInputVarGenerated(platform::CPUPlace());
 
   tensors_.resize(inputs_.size());
-  auto *var_handle = static_cast<VarHandle *>(inputs_[0]);
-  auto &var_name = var_handle->name_;
   platform::CPUPlace cpu;
   auto &scopes = *local_scopes_;
 
-  for (size_t i = 0; i < scopes.size(); ++i) {
-    auto &scope = scopes[i];
-    auto *var =
-        scope->FindVar(kLocalExecScopeName)->Get<Scope *>()->FindVar(var_name);
+  for (size_t i = 0; i < inputs_.size(); ++i) {
+    auto *var_handle = static_cast<VarHandle *>(inputs_[i]);
+    auto &scope = scopes.at(var_handle->scope_idx_);
+    auto *var = scope->FindVar(kLocalExecScopeName)
+                    ->Get<Scope *>()
+                    ->FindVar(var_handle->name_);
     PADDLE_ENFORCE_NOT_NULL(var, "Cannot find variable %s in execution scope",
-                            var_name);
+                            var_handle->name_);
+
     auto &t = var->Get<framework::LoDTensor>();
     if (platform::is_gpu_place(t.place())) {
 #ifdef PADDLE_WITH_CUDA
diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto
index d2558f111f49139b33f921f7260b41830279edc8..d35125fe8c3c8018c38650dc87b2b1474ded6058 100644
--- a/paddle/fluid/framework/framework.proto
+++ b/paddle/fluid/framework/framework.proto
@@ -103,6 +103,7 @@ message VarType {
     FP64 = 6;
     // Tensor<size_t> is used in C++.
     SIZE_T = 19;
+    UINT8 = 20;
 
     // Other types that may need additional descriptions
     LOD_TENSOR = 7;
diff --git a/paddle/fluid/framework/lod_tensor_test.cc b/paddle/fluid/framework/lod_tensor_test.cc
index 77e5ec4c7dd14b7ebb6d606b8c401ee714259d40..2ceffc93319359683e87e7fec2d18784c9bf02f3 100644
--- a/paddle/fluid/framework/lod_tensor_test.cc
+++ b/paddle/fluid/framework/lod_tensor_test.cc
@@ -228,11 +228,12 @@ TEST(LoD, CheckAbsLoD) {
   ASSERT_FALSE(CheckAbsLoD(abs_lod0));
 }
 
-TEST(LoDTensor, RecordIO) {
+template <typename T>
+static void TestRecordIO() {
   LoDTensor tensor;
-  int* tmp = tensor.mutable_data<int>(make_ddim({4, 5}), platform::CPUPlace());
+  T* tmp = tensor.mutable_data<T>(make_ddim({4, 5}), platform::CPUPlace());
   for (int i = 0; i < 20; ++i) {
-    tmp[i] = i;
+    tmp[i] = static_cast<T>(i);
   }
 
   std::stringstream* stream = new std::stringstream();
@@ -247,7 +248,7 @@ TEST(LoDTensor, RecordIO) {
 
   auto assert_tensor_ok = [](const LoDTensor& tensor) {
     for (int i = 0; i < 20; ++i) {
-      ASSERT_EQ(tensor.data<int>()[i], i);
+      ASSERT_EQ(tensor.data<T>()[i], static_cast<T>(i));
     }
   };
 
@@ -265,5 +266,13 @@ TEST(LoDTensor, RecordIO) {
   }
 }
 
+TEST(LoDTensor, RecordIO) {
+  TestRecordIO<int>();
+  TestRecordIO<int16_t>();
+  TestRecordIO<uint8_t>();
+  TestRecordIO<float>();
+  TestRecordIO<double>();
+}
+
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc
index d62ea387cc55c7399973b6f35bace491a49666dc..d39154c6f88d6d17c1719eb9a5b048211f4bb52b 100644
--- a/paddle/fluid/operators/math/math_function.cc
+++ b/paddle/fluid/operators/math/math_function.cc
@@ -38,7 +38,9 @@ template struct SetConstant<platform::CPUDeviceContext, bool>;
   template struct Transpose<platform::CPUDeviceContext, double, RANK>;     \
   template struct Transpose<platform::CPUDeviceContext, int, RANK>;        \
   template struct Transpose<platform::CPUDeviceContext, int64_t, RANK>;    \
-  template struct Transpose<platform::CPUDeviceContext, bool, RANK>;
+  template struct Transpose<platform::CPUDeviceContext, bool, RANK>;       \
+  template struct Transpose<platform::CPUDeviceContext, int16_t, RANK>;    \
+  template struct Transpose<platform::CPUDeviceContext, uint8_t, RANK>;
 
 DEFINE_CPU_TRANS(1);
 DEFINE_CPU_TRANS(2);
diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt
index 598fd4d419078a973647f2f8f20e8a12c8115a8b..79e3c26fef51b4d27520a8079de1074d72f89617 100644
--- a/paddle/fluid/platform/CMakeLists.txt
+++ b/paddle/fluid/platform/CMakeLists.txt
@@ -1,4 +1,4 @@
-proto_library(profiler_proto SRCS profiler.proto)
+proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto)
 py_proto_compile(profiler_py_proto SRCS profiler.proto)
 
 add_custom_target(profiler_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
index 5bef232cd8fc44ded89ac56a790c8db0955b390a..cb0624eac4f55cbd4739b1dd10a4256da67dd85f 100755
--- a/paddle/scripts/paddle_build.sh
+++ b/paddle/scripts/paddle_build.sh
@@ -504,6 +504,7 @@ function main() {
         ;;
       capi)
         cmake_gen ${PYTHON_ABI:-""}
+        build
         gen_capi_package
         ;;
       fluid_inference_lib)
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor.py b/python/paddle/fluid/tests/unittests/test_parallel_executor.py
index 6dc016487fd81a9292f94042a20b7356bc50abe1..056f9e1781997aa1586d972874b652d5b725fe3f 100644
--- a/python/paddle/fluid/tests/unittests/test_parallel_executor.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor.py
@@ -775,7 +775,7 @@ class TestCRFModel(unittest.TestCase):
         build_strategy = fluid.BuildStrategy()
         build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
         self.check_network_convergence(
-            is_sparse=False, build_strategy=build_strategy)
+            is_sparse=True, build_strategy=build_strategy)
 
     def test_update_dense_parameter_reduce(self):
         build_strategy = fluid.BuildStrategy()
@@ -849,8 +849,7 @@ class TestFetchOp(unittest.TestCase):
                     assert not math.isnan(np.sum(ret[i])) and \
                            not math.isinf(np.sum(ret[i]))
 
-    @unittest.skip("this test is buggy")
-    def test_feed(self):
+    def test_fetch_op(self):
         tst_reader = paddle.batch(flowers.test(use_xmap=False), batch_size=16)
         tst_reader_iter = tst_reader()