[Cherry-pick] Several bugs fix (#44991)

* fix device context init error (#43910) * Fix core so name mismatch error (#43977) * fix core avx soname error * remove print info * add clip_extra (#44008) * fix tensor stream error in custom op (#44500) * fix custom op attr names size error (#44938)

[Cherry-pick] Several bugs fix (#44991)
* fix device context init error (#43910) * Fix core so name mismatch error (#43977) * fix core avx soname error * remove print info * add clip_extra (#44008) * fix tensor stream error in custom op (#44500) * fix custom op attr names size error (#44938)
e00aa903 · Chen Weihang · GitHub · c5f4a9cc · e00aa903 · e00aa903
7 changed file
--- a/paddle/fluid/platform/device_context.h
+++ b/paddle/fluid/platform/device_context.h
@@ -58,7 +58,7 @@ limitations under the License. */
 #endif

 #ifdef PADDLE_WITH_MKLDNN
-#include "dnnl.hpp"
+#include "dnnl.hpp"  // NOLINT
 #include "paddle/fluid/framework/data_layout.h"
 #endif

@@ -902,6 +902,8 @@ class DeviceContextPool {
    return *pool;
  }

+  static bool IsInitialized() { return pool != nullptr; }
+
  static void SetPool(DeviceContextPool* dev_pool) { pool = dev_pool; }

  /*! \brief  Return handle of single device context. */

--- a/paddle/fluid/pybind/eager_functions.cc
+++ b/paddle/fluid/pybind/eager_functions.cc
@@ -305,8 +305,7 @@ static std::vector<paddle::any> CastAttrsToTragetType(
                        attrs_names.size(), src.size()));
  for (size_t i = 0; i < src.size(); i++) {
    size_t end = attrs_names[i].find(": ");
-    std::string type_name =
-        attrs_names[i].substr(end + 2, attrs_names.size() - end - 2);
+    std::string type_name = attrs_names[i].substr(end + 2);
    if (type_name == "int") {
      if (src[i].type() == typeid(bool)) {
        res.emplace_back(static_cast<int>(paddle::any_cast<bool>(src[i])));

--- a/paddle/phi/api/lib/CMakeLists.txt
+++ b/paddle/phi/api/lib/CMakeLists.txt
 add_subdirectory(utils)

-if (WITH_GPU)
-  nv_library(phi_tensor_raw SRCS tensor.cc DEPS tensor_base dense_tensor phi_api_utils phi_enforce)
-elseif (WITH_ROCM)
-  hip_library(phi_tensor_raw SRCS tensor.cc DEPS tensor_base dense_tensor phi_api_utils phi_enforce)
+if(WITH_GPU)
+  nv_library(
+    phi_tensor_raw
+    SRCS tensor.cc
+    DEPS tensor_base dense_tensor phi_api_utils phi_enforce context_pool)
+elseif(WITH_ROCM)
+  hip_library(
+    phi_tensor_raw
+    SRCS tensor.cc
+    DEPS tensor_base dense_tensor phi_api_utils phi_enforce context_pool)
 else()
-  cc_library(phi_tensor_raw SRCS tensor.cc DEPS tensor_base dense_tensor phi_api_utils phi_enforce)
+  cc_library(
+    phi_tensor_raw
+    SRCS tensor.cc
+    DEPS tensor_base dense_tensor phi_api_utils phi_enforce context_pool)
 endif()

 set(api_gen_base ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/api_base.py)
@@ -19,51 +28,71 @@ set(api_header_file_tmp ${api_header_file}.tmp)
 set(api_source_file_tmp ${api_source_file}.tmp)

 # backward api file
-set(bw_api_gen_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/backward_api_gen.py)
-set(bw_api_yaml_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/backward.yaml)
-set(bw_api_header_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/backward/backward_api.h)
+set(bw_api_gen_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/backward_api_gen.py)
+set(bw_api_yaml_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/backward.yaml)
+set(bw_api_header_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/backward/backward_api.h)
 set(bw_api_source_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/backward_api.cc)
 set(bw_api_header_file_tmp ${bw_api_header_file}.tmp)
 set(bw_api_source_file_tmp ${bw_api_source_file}.tmp)

 # dygraph(intermediate) api file
-set(im_api_gen_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/intermediate_api_gen.py)
-set(dygraph_api_header_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/dygraph_api.h)
-set(dygraph_api_source_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/dygraph_api.cc)
+set(im_api_gen_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/intermediate_api_gen.py)
+set(dygraph_api_header_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/dygraph_api.h)
+set(dygraph_api_source_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/dygraph_api.cc)
 set(dygraph_api_header_file_tmp ${dygraph_api_header_file}.tmp)
 set(dygraph_api_source_file_tmp ${dygraph_api_source_file}.tmp)

 # sparse api file
-set(sparse_api_gen_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/sparse_api_gen.py)
-set(sparse_api_yaml_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/sparse_api.yaml)
-set(sparse_api_header_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/include/sparse_api.h)
+set(sparse_api_gen_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/sparse_api_gen.py)
+set(sparse_api_yaml_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/sparse_api.yaml)
+set(sparse_api_header_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/include/sparse_api.h)
 set(sparse_api_source_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/sparse_api.cc)
 set(sparse_api_header_file_tmp ${sparse_api_header_file}.tmp)
 set(sparse_api_source_file_tmp ${sparse_api_source_file}.tmp)

 # sparse bw api file
-set(sparse_bw_api_gen_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/sparse_bw_api_gen.py)
-set(sparse_bw_api_yaml_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/sparse_bw_api.yaml)
-set(sparse_bw_api_header_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/backward/sparse_bw_api.h)
-set(sparse_bw_api_source_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/sparse_bw_api.cc)
+set(sparse_bw_api_gen_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/sparse_bw_api_gen.py)
+set(sparse_bw_api_yaml_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/sparse_bw_api.yaml)
+set(sparse_bw_api_header_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/backward/sparse_bw_api.h)
+set(sparse_bw_api_source_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/sparse_bw_api.cc)
 set(sparse_bw_api_header_file_tmp ${sparse_bw_api_header_file}.tmp)
 set(sparse_bw_api_source_file_tmp ${sparse_bw_api_source_file}.tmp)

 # strings api file
-set(strings_api_gen_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/strings_api_gen.py)
-set(strings_api_yaml_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/strings_api.yaml)
-set(strings_api_header_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/include/strings_api.h)
-set(strings_api_source_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/strings_api.cc)
+set(strings_api_gen_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/strings_api_gen.py)
+set(strings_api_yaml_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/strings_api.yaml)
+set(strings_api_header_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/include/strings_api.h)
+set(strings_api_source_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/strings_api.cc)
 set(strings_api_header_file_tmp ${strings_api_header_file}.tmp)
 set(strings_api_source_file_tmp ${strings_api_source_file}.tmp)

 # wrapped infermeta file
-set(wrapped_infermeta_gen_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/wrapped_infermeta_gen.py)
+set(wrapped_infermeta_gen_file
+    ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/wrapped_infermeta_gen.py)
 set(api_yaml_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/api.yaml)
-set(wrapped_infermeta_header_file ${CMAKE_SOURCE_DIR}/paddle/phi/infermeta/generated.h)
-set(wrapped_infermeta_source_file ${CMAKE_SOURCE_DIR}/paddle/phi/infermeta/generated.cc)
+set(wrapped_infermeta_header_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/infermeta/generated.h)
+set(wrapped_infermeta_source_file
+    ${CMAKE_SOURCE_DIR}/paddle/phi/infermeta/generated.cc)

-if (NOT PYTHON_EXECUTABLE)
+if(NOT PYTHON_EXECUTABLE)
  find_package(PythonInterp REQUIRED)
 endif()

@@ -71,25 +100,30 @@ endif()
 add_custom_command(
  OUTPUT ${api_header_file} ${api_source_file}
  COMMAND ${PYTHON_EXECUTABLE} -m pip install pyyaml
-  COMMAND ${PYTHON_EXECUTABLE} ${api_gen_file}
-                 --api_yaml_path ${api_yaml_file}
-                 --api_header_path ${api_header_file_tmp}
-                 --api_source_path ${api_source_file_tmp}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${api_header_file_tmp} ${api_header_file}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${api_source_file_tmp} ${api_source_file}
+  COMMAND
+    ${PYTHON_EXECUTABLE} ${api_gen_file} --api_yaml_path ${api_yaml_file}
+    --api_header_path ${api_header_file_tmp} --api_source_path
+    ${api_source_file_tmp}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${api_header_file_tmp}
+          ${api_header_file}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${api_source_file_tmp}
+          ${api_source_file}
  COMMENT "copy_if_different ${api_header_file} ${api_source_file}"
  DEPENDS ${api_yaml_file} ${api_gen_file} ${api_gen_base}
  VERBATIM)

 # generate backward api
 add_custom_command(
-  OUTPUT ${bw_api_header_file} ${bw_api_source_file} ${bw_api_header_file_tmp} ${bw_api_source_file_tmp}
-  COMMAND ${PYTHON_EXECUTABLE} ${bw_api_gen_file}
-                 --backward_yaml_path ${bw_api_yaml_file}
-                 --backward_header_path ${bw_api_header_file_tmp}
-                 --backward_source_path ${bw_api_source_file_tmp}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${bw_api_header_file_tmp} ${bw_api_header_file}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${bw_api_source_file_tmp} ${bw_api_source_file}
+  OUTPUT ${bw_api_header_file} ${bw_api_source_file} ${bw_api_header_file_tmp}
+         ${bw_api_source_file_tmp}
+  COMMAND
+    ${PYTHON_EXECUTABLE} ${bw_api_gen_file} --backward_yaml_path
+    ${bw_api_yaml_file} --backward_header_path ${bw_api_header_file_tmp}
+    --backward_source_path ${bw_api_source_file_tmp}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${bw_api_header_file_tmp}
+          ${bw_api_header_file}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${bw_api_source_file_tmp}
+          ${bw_api_source_file}
  COMMENT "copy_if_different ${bw_api_header_file} ${bw_api_source_file}"
  DEPENDS ${bw_api_yaml_file} ${bw_api_gen_file} ${api_gen_base}
  VERBATIM)
@@ -97,81 +131,173 @@ add_custom_command(
 # generate sparse api
 add_custom_command(
  OUTPUT ${sparse_api_header_file} ${sparse_api_source_file}
-  COMMAND ${PYTHON_EXECUTABLE} ${sparse_api_gen_file}
-                 --api_yaml_path ${sparse_api_yaml_file}
-                 --api_header_path ${sparse_api_header_file_tmp}
-                 --api_source_path ${sparse_api_source_file_tmp}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${sparse_api_header_file_tmp} ${sparse_api_header_file}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${sparse_api_source_file_tmp} ${sparse_api_source_file}
-  COMMENT "copy_if_different ${sparse_api_header_file} ${sparse_sparse_api_source_file}"
-  DEPENDS ${sparse_api_yaml_file} ${sparse_api_gen_file} ${api_gen_base} ${api_gen_file}
+  COMMAND
+    ${PYTHON_EXECUTABLE} ${sparse_api_gen_file} --api_yaml_path
+    ${sparse_api_yaml_file} --api_header_path ${sparse_api_header_file_tmp}
+    --api_source_path ${sparse_api_source_file_tmp}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${sparse_api_header_file_tmp}
+          ${sparse_api_header_file}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${sparse_api_source_file_tmp}
+          ${sparse_api_source_file}
+  COMMENT
+    "copy_if_different ${sparse_api_header_file} ${sparse_sparse_api_source_file}"
+  DEPENDS ${sparse_api_yaml_file} ${sparse_api_gen_file} ${api_gen_base}
+          ${api_gen_file}
  VERBATIM)

 # generate backward sparse api
 add_custom_command(
  OUTPUT ${sparse_bw_api_header_file} ${sparse_bw_api_source_file}
-  COMMAND ${PYTHON_EXECUTABLE} ${sparse_bw_api_gen_file}
-                 --api_yaml_path ${sparse_bw_api_yaml_file}
-                 --api_header_path ${sparse_bw_api_header_file_tmp}
-                 --api_source_path ${sparse_bw_api_source_file_tmp}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${sparse_bw_api_header_file_tmp} ${sparse_bw_api_header_file}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${sparse_bw_api_source_file_tmp} ${sparse_bw_api_source_file}
-  COMMENT "copy_if_different ${sparse_bw_api_header_file} ${sparse_bw_sparse_api_source_file}"
-  DEPENDS ${sparse_bw_api_yaml_file} ${sparse_bw_api_gen_file} ${api_gen_base} ${api_gen_file} ${sparse_api_gen_file} ${bw_api_gen_file}
+  COMMAND
+    ${PYTHON_EXECUTABLE} ${sparse_bw_api_gen_file} --api_yaml_path
+    ${sparse_bw_api_yaml_file} --api_header_path
+    ${sparse_bw_api_header_file_tmp} --api_source_path
+    ${sparse_bw_api_source_file_tmp}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${sparse_bw_api_header_file_tmp}
+          ${sparse_bw_api_header_file}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${sparse_bw_api_source_file_tmp}
+          ${sparse_bw_api_source_file}
+  COMMENT
+    "copy_if_different ${sparse_bw_api_header_file} ${sparse_bw_sparse_api_source_file}"
+  DEPENDS ${sparse_bw_api_yaml_file} ${sparse_bw_api_gen_file} ${api_gen_base}
+          ${api_gen_file} ${sparse_api_gen_file} ${bw_api_gen_file}
  VERBATIM)

 # generate strings api
 add_custom_command(
  OUTPUT ${strings_api_header_file} ${strings_api_source_file}
-  COMMAND ${PYTHON_EXECUTABLE} ${strings_api_gen_file}
-                 --api_yaml_path ${strings_api_yaml_file}
-                 --api_header_path ${strings_api_header_file_tmp}
-                 --api_source_path ${strings_api_source_file_tmp}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${strings_api_header_file_tmp} ${strings_api_header_file}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${strings_api_source_file_tmp} ${strings_api_source_file}
-  COMMENT "copy_if_different ${strings_api_header_file} ${strings_strings_api_source_file}"
-  DEPENDS ${strings_api_yaml_file} ${strings_api_gen_file} ${api_gen_base} ${api_gen_file}
+  COMMAND
+    ${PYTHON_EXECUTABLE} ${strings_api_gen_file} --api_yaml_path
+    ${strings_api_yaml_file} --api_header_path ${strings_api_header_file_tmp}
+    --api_source_path ${strings_api_source_file_tmp}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${strings_api_header_file_tmp}
+          ${strings_api_header_file}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${strings_api_source_file_tmp}
+          ${strings_api_source_file}
+  COMMENT
+    "copy_if_different ${strings_api_header_file} ${strings_strings_api_source_file}"
+  DEPENDS ${strings_api_yaml_file} ${strings_api_gen_file} ${api_gen_base}
+          ${api_gen_file}
  VERBATIM)

 # generate dygraph(intermediate) api
 add_custom_command(
  OUTPUT ${dygraph_api_header_file} ${dygraph_api_source_file}
-  COMMAND ${PYTHON_EXECUTABLE} ${im_api_gen_file}
-                 --api_yaml_path ${api_yaml_file}
-                 --sparse_api_yaml_path ${sparse_api_yaml_file}
-                 --dygraph_api_header_path ${dygraph_api_header_file_tmp}
-                 --dygraph_api_source_path ${dygraph_api_source_file_tmp}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${dygraph_api_header_file_tmp} ${dygraph_api_header_file}
-  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${dygraph_api_source_file_tmp} ${dygraph_api_source_file}
-  DEPENDS ${api_yaml_file} ${sparse_api_yaml_file} ${im_api_gen_file} ${api_gen_base} ${api_gen_file}
+  COMMAND
+    ${PYTHON_EXECUTABLE} ${im_api_gen_file} --api_yaml_path ${api_yaml_file}
+    --sparse_api_yaml_path ${sparse_api_yaml_file} --dygraph_api_header_path
+    ${dygraph_api_header_file_tmp} --dygraph_api_source_path
+    ${dygraph_api_source_file_tmp}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${dygraph_api_header_file_tmp}
+          ${dygraph_api_header_file}
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different ${dygraph_api_source_file_tmp}
+          ${dygraph_api_source_file}
+  DEPENDS ${api_yaml_file} ${sparse_api_yaml_file} ${im_api_gen_file}
+          ${api_gen_base} ${api_gen_file}
  VERBATIM)

 # generate wrapped infermeta
 add_custom_command(
  OUTPUT ${wrapped_infermeta_header_file} ${wrapped_infermeta_source_file}
-  COMMAND ${PYTHON_EXECUTABLE} ${wrapped_infermeta_gen_file}
-                 --api_yaml_path ${api_yaml_file}
-                 --wrapped_infermeta_header_path ${wrapped_infermeta_header_file}
-                 --wrapped_infermeta_source_path ${wrapped_infermeta_source_file}
+  COMMAND
+    ${PYTHON_EXECUTABLE} ${wrapped_infermeta_gen_file} --api_yaml_path
+    ${api_yaml_file} --wrapped_infermeta_header_path
+    ${wrapped_infermeta_header_file} --wrapped_infermeta_source_path
+    ${wrapped_infermeta_source_file}
  DEPENDS ${api_yaml_file} ${wrapped_infermeta_gen_file} ${api_gen_base}
  VERBATIM)

-cc_library(op_meta_info SRCS op_meta_info.cc DEPS phi_tensor_raw)
-cc_library(wrapped_infermeta SRCS ${wrapped_infermeta_source_file} DEPS phi)
-cc_library(context_pool SRCS context_pool.cc DEPS phi_context phi_enforce place)
-
-cc_library(kernel_dispatch SRCS kernel_dispatch.cc DEPS phi_tensor_raw phi_context kernel_factory context_pool)
-cc_library(api_gen_utils SRCS api_gen_utils.cc DEPS phi_tensor_raw selected_rows sparse_csr_tensor sparse_coo_tensor)
-cc_library(phi_data_transform SRCS data_transform.cc DEPS phi_tensor_raw transfer_layout_kernel cast_kernel data_device_transform)
-cc_library(api_custom_impl SRCS api_custom_impl.cc DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils backward_infermeta phi_data_transform)
-cc_library(sparse_api_custom_impl SRCS sparse_api_custom_impl.cc DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils phi_data_transform tensor_copy)
-
-cc_library(phi_function_api SRCS ${api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils phi_data_transform api_custom_impl)
-cc_library(phi_bw_function_api SRCS ${bw_api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils backward_infermeta phi_data_transform phi_function_api api_custom_impl global_utils)
-cc_library(sparse_api SRCS ${sparse_api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils sparse_api_custom_impl)
-cc_library(sparse_bw_api SRCS ${sparse_bw_api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils sparse_api sparse_api_custom_impl)
-cc_library(phi_dygraph_api SRCS ${dygraph_api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils phi_data_transform phi_function_api sparse_api)
-cc_library(strings_api SRCS ${strings_api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils)
-cc_library(phi_tensor SRCS tensor_method.cc DEPS phi_tensor_raw phi_function_api api_gen_utils kernel_dispatch infermeta sparse_api strings_api)
-cc_library(tensor_copy SRCS tensor_copy.cc DEPS phi_tensor_raw copy_kernel kernel_dispatch api_gen_utils)
-cc_library(api_scalar SRCS scalar.cc DEPS tensor_copy)
+cc_library(
+  op_meta_info
+  SRCS op_meta_info.cc
+  DEPS phi_tensor_raw)
+cc_library(
+  wrapped_infermeta
+  SRCS ${wrapped_infermeta_source_file}
+  DEPS phi)
+cc_library(
+  context_pool
+  SRCS context_pool.cc
+  DEPS phi_context phi_enforce place init)
+
+cc_library(
+  kernel_dispatch
+  SRCS kernel_dispatch.cc
+  DEPS phi_tensor_raw phi_context kernel_factory context_pool)
+cc_library(
+  api_gen_utils
+  SRCS api_gen_utils.cc
+  DEPS phi_tensor_raw selected_rows sparse_csr_tensor sparse_coo_tensor)
+cc_library(
+  phi_data_transform
+  SRCS data_transform.cc
+  DEPS phi_tensor_raw transfer_layout_kernel cast_kernel data_device_transform)
+cc_library(
+  api_custom_impl
+  SRCS api_custom_impl.cc
+  DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils backward_infermeta
+       phi_data_transform)
+cc_library(
+  sparse_api_custom_impl
+  SRCS sparse_api_custom_impl.cc
+  DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils phi_data_transform
+       tensor_copy)
+
+cc_library(
+  phi_function_api
+  SRCS ${api_source_file}
+  DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils phi_data_transform
+       api_custom_impl)
+cc_library(
+  phi_bw_function_api
+  SRCS ${bw_api_source_file}
+  DEPS phi_tensor_raw
+       phi
+       kernel_dispatch
+       api_gen_utils
+       backward_infermeta
+       phi_data_transform
+       phi_function_api
+       api_custom_impl
+       global_utils)
+cc_library(
+  sparse_api
+  SRCS ${sparse_api_source_file}
+  DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils sparse_api_custom_impl)
+cc_library(
+  sparse_bw_api
+  SRCS ${sparse_bw_api_source_file}
+  DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils sparse_api
+       sparse_api_custom_impl)
+cc_library(
+  phi_dygraph_api
+  SRCS ${dygraph_api_source_file}
+  DEPS phi_tensor_raw
+       phi
+       kernel_dispatch
+       api_gen_utils
+       phi_data_transform
+       phi_function_api
+       sparse_api)
+cc_library(
+  strings_api
+  SRCS ${strings_api_source_file}
+  DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils)
+cc_library(
+  phi_tensor
+  SRCS tensor_method.cc
+  DEPS phi_tensor_raw
+       phi_function_api
+       api_gen_utils
+       kernel_dispatch
+       infermeta
+       sparse_api
+       strings_api)
+cc_library(
+  tensor_copy
+  SRCS tensor_copy.cc
+  DEPS phi_tensor_raw copy_kernel kernel_dispatch api_gen_utils)
+cc_library(
+  api_scalar
+  SRCS scalar.cc
+  DEPS tensor_copy)
--- a/paddle/phi/api/lib/context_pool.cc
+++ b/paddle/phi/api/lib/context_pool.cc
@@ -17,6 +17,8 @@ limitations under the License. */
 #include "paddle/phi/backends/all_context.h"
 #include "paddle/phi/core/enforce.h"

+#include "paddle/fluid/platform/init.h"
+
 namespace paddle {
 namespace experimental {

@@ -28,6 +30,9 @@ DeviceContextPool& DeviceContextPool::Instance() {
 const phi::DeviceContext* DeviceContextPool::Get(const Place& place) {
  auto it = context_map_.find(place);
  if (it == context_map_.end()) {
+    if (!paddle::platform::DeviceContextPool::IsInitialized()) {
+      paddle::framework::InitDevices();
+    }
    // only when we need the specific DeviceContext, get and cache it
    auto* dev_ctx = paddle::platform::DeviceContextPool::Instance().Get(place);
    {

--- a/paddle/phi/api/lib/tensor.cc
+++ b/paddle/phi/api/lib/tensor.cc
@@ -20,7 +20,9 @@ limitations under the License. */

 #include "glog/logging.h"

+#include "paddle/phi/api/include/context_pool.h"
 #include "paddle/phi/api/lib/utils/allocator.h"
+#include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/backends/gpu/gpu_info.h"
 #include "paddle/phi/core/ddim.h"
 #include "paddle/phi/core/dense_tensor.h"
@@ -32,8 +34,7 @@ limitations under the License. */
 #include "paddle/phi/core/tensor_base.h"
 #include "paddle/phi/core/tensor_meta.h"
 #include "paddle/phi/core/tensor_utils.h"
-
-#include "paddle/fluid/platform/stream/cuda_stream.h"
+// clang-format off

 namespace paddle {
 namespace experimental {
@@ -305,7 +306,10 @@ void Tensor::set_impl(std::shared_ptr<phi::TensorBase> &&impl) {

 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 gpuStream_t Tensor::stream() const {
-  return platform::stream::get_current_stream(-1)->raw_stream();
+  int device_id = phi::backends::gpu::GetCurrentDeviceId();
+  auto* gpu_context = DeviceContextPool::Instance()
+    .Get<AllocationType::GPU>(GPUPlace(device_id));
+  return gpu_context->stream();
 }
 #endif


--- a/python/paddle/fluid/dygraph/jit.py
+++ b/python/paddle/fluid/dygraph/jit.py
@@ -64,8 +64,8 @@ def _extract_vars(inputs, result_list, err_tag='inputs'):
            _extract_vars(var, result_list, err_tag)
    else:
        raise TypeError(
-            "The type of 'each element of {}' in fluid.dygraph.jit.TracedLayer.trace must be fluid.Variable, but received {}.".
-            format(err_tag, type(inputs)))
+            "The type of 'each element of {}' in fluid.dygraph.jit.TracedLayer.trace must be fluid.Variable, but received {}."
+            .format(err_tag, type(inputs)))


 def extract_vars(inputs, err_tag='inputs'):
@@ -211,20 +211,19 @@ def declarative(function=None, input_spec=None, build_strategy=None):
        _, python_func = unwrap_decorators(python_func)

        # Step 2. copy some attributes from original python function.
-        static_layer = copy_decorator_attrs(
-            original_func=python_func,
-            decorated_obj=StaticFunction(
-                function=python_func,
-                input_spec=input_spec,
-                build_strategy=build_strategy))
+        static_layer = copy_decorator_attrs(original_func=python_func,
+                                            decorated_obj=StaticFunction(
+                                                function=python_func,
+                                                input_spec=input_spec,
+                                                build_strategy=build_strategy))

        return static_layer

    build_strategy = build_strategy or BuildStrategy()
    if not isinstance(build_strategy, BuildStrategy):
        raise TypeError(
-            "Required type(build_strategy) shall be `paddle.static.BuildStrategy`, but received {}".
-            format(type(build_strategy).__name__))
+            "Required type(build_strategy) shall be `paddle.static.BuildStrategy`, but received {}"
+            .format(type(build_strategy).__name__))

    # for usage: `declarative(foo, ...)`
    if function is not None:
@@ -232,8 +231,8 @@ def declarative(function=None, input_spec=None, build_strategy=None):
            if isinstance(function.forward, StaticFunction):
                class_name = function.__class__.__name__
                logging_utils.warn(
-                    "`{}.forward` has already been decorated somewhere. It will be redecorated to replace previous one.".
-                    format(class_name))
+                    "`{}.forward` has already been decorated somewhere. It will be redecorated to replace previous one."
+                    .format(class_name))
            function.forward = decorated(function.forward)
            return function
        else:
@@ -284,6 +283,7 @@ def not_to_static(func=None):


 class _SaveLoadConfig(object):
+
    def __init__(self):
        self._output_spec = None
        self._model_filename = None
@@ -371,7 +371,7 @@ class _SaveLoadConfig(object):


 def _parse_save_configs(configs):
-    supported_configs = ['output_spec', "with_hook"]
+    supported_configs = ['output_spec', "with_hook", "clip_extra"]

    # input check
    for key in configs:
@@ -384,6 +384,7 @@ def _parse_save_configs(configs):
    inner_config = _SaveLoadConfig()
    inner_config.output_spec = configs.get('output_spec', None)
    inner_config.with_hook = configs.get('with_hook', False)
+    inner_config.clip_extra = configs.get("clip_extra", False)

    return inner_config

@@ -622,6 +623,7 @@ def _remove_save_pre_hook(hook):


 def _run_save_pre_hooks(func):
+
    def wrapper(layer, path, input_spec=None, **configs):
        global _save_pre_hooks
        for hook in _save_pre_hooks:
@@ -775,8 +777,8 @@ def save(layer, path, input_spec=None, **configs):
            "The paddle.jit.save doesn't work when setting ProgramTranslator.enable to False."
        )

-    if not (isinstance(layer, Layer) or inspect.isfunction(layer) or isinstance(
-            layer, StaticFunction)):
+    if not (isinstance(layer, Layer) or inspect.isfunction(layer)
+            or isinstance(layer, StaticFunction)):
        raise TypeError(
            "The input of paddle.jit.save should be 'Layer' or 'Function', but received input type is %s."
            % type(layer))
@@ -837,7 +839,7 @@ def save(layer, path, input_spec=None, **configs):
    # parse configs
    configs = _parse_save_configs(configs)
    # whether outermost layer has pre/post hook, if does, we need also save
-    # these operators in program. 
+    # these operators in program.
    with_hook = configs.with_hook

    scope = core.Scope()
@@ -848,7 +850,9 @@ def save(layer, path, input_spec=None, **configs):
            with_hook = True
    else:
        # layer is function
-        functions = [layer, ]
+        functions = [
+            layer,
+        ]
    for attr_func in functions:
        if isinstance(layer, Layer):
            static_func = getattr(inner_layer, attr_func, None)
@@ -862,8 +866,8 @@ def save(layer, path, input_spec=None, **configs):
                if inner_input_spec:
                    inner_input_spec = pack_sequence_as(input_spec,
                                                        inner_input_spec)
-                static_forward = declarative(
-                    inner_layer.forward, input_spec=inner_input_spec)
+                static_forward = declarative(inner_layer.forward,
+                                             input_spec=inner_input_spec)
                concrete_program = static_forward.concrete_program_specify_input_spec(
                    with_hook=with_hook)
                # the input_spec has been used in declarative, which is equal to
@@ -882,14 +886,14 @@ def save(layer, path, input_spec=None, **configs):
                if inner_input_spec:
                    inner_input_spec = pack_sequence_as(input_spec,
                                                        inner_input_spec)
-                static_function = declarative(
-                    attr_func, input_spec=inner_input_spec)
+                static_function = declarative(attr_func,
+                                              input_spec=inner_input_spec)
                concrete_program = static_function.concrete_program

                if static_function._class_instance is None:
                    warnings.warn(
-                        '`jit.save` will only save the `Program`, not the parameters. If you have to save the parameters, please make sure that {} is a member function of `paddle.nn.Layer` and the saved parameters are in `state_dict`'.
-                        format(layer))
+                        '`jit.save` will only save the `Program`, not the parameters. If you have to save the parameters, please make sure that {} is a member function of `paddle.nn.Layer` and the saved parameters are in `state_dict`'
+                        .format(layer))

        dygraph_state_dict = None
        if isinstance(inner_layer, Layer):
@@ -922,8 +926,8 @@ def save(layer, path, input_spec=None, **configs):
                        param_or_buffer_tensor = scope.var(
                            param_or_buffer.name).get_tensor()
                        #src_tensor = param_or_buffer.value().get_tensor()
-                        src_tensor = state_var_dict[param_or_buffer.name].value(
-                        ).get_tensor()
+                        src_tensor = state_var_dict[
+                            param_or_buffer.name].value().get_tensor()
                        param_or_buffer_tensor._share_data_with(src_tensor)
                    # record var info
                    if param_or_buffer.name not in extra_var_info:
@@ -986,7 +990,7 @@ def save(layer, path, input_spec=None, **configs):
                params_filename=params_filename,
                export_for_deployment=configs._export_for_deployment,
                program_only=configs._program_only,
-                clip_extra=False)
+                clip_extra=configs.clip_extra)

    # NOTE(chenweihang): [ Save extra variable info ]
    # save_inference_model will lose some important variable information, including:
@@ -1534,14 +1538,16 @@ class TracedLayer(object):
                   "fluid.dygraph.jit.TracedLayer.save_inference_model")
        if isinstance(feed, list):
            for f in feed:
-                check_type(f, "each element of feed", int,
-                           "fluid.dygraph.jit.TracedLayer.save_inference_model")
+                check_type(
+                    f, "each element of feed", int,
+                    "fluid.dygraph.jit.TracedLayer.save_inference_model")
        check_type(fetch, "fetch", (type(None), list),
                   "fluid.dygraph.jit.TracedLayer.save_inference_model")
        if isinstance(fetch, list):
            for f in fetch:
-                check_type(f, "each element of fetch", int,
-                           "fluid.dygraph.jit.TracedLayer.save_inference_model")
+                check_type(
+                    f, "each element of fetch", int,
+                    "fluid.dygraph.jit.TracedLayer.save_inference_model")
        clip_extra = kwargs.get('clip_extra', False)
        # path check
        file_prefix = os.path.basename(path)
@@ -1575,12 +1581,11 @@ class TracedLayer(object):
            model_filename = file_prefix + INFER_MODEL_SUFFIX
            params_filename = file_prefix + INFER_PARAMS_SUFFIX

-            save_inference_model(
-                dirname=dirname,
-                feeded_var_names=feeded_var_names,
-                target_vars=target_vars,
-                executor=self._exe,
-                main_program=self._program.clone(),
-                model_filename=model_filename,
-                params_filename=params_filename,
-                clip_extra=clip_extra)
+            save_inference_model(dirname=dirname,
+                                 feeded_var_names=feeded_var_names,
+                                 target_vars=target_vars,
+                                 executor=self._exe,
+                                 main_program=self._program.clone(),
+                                 model_filename=model_filename,
+                                 params_filename=params_filename,
+                                 clip_extra=clip_extra)
--- a/python/setup.py.in
+++ b/python/setup.py.in
@@ -570,7 +570,8 @@ if '${CMAKE_BUILD_TYPE}' == 'Release':
            commands = ["install_name_tool -id '@loader_path/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/${FLUID_CORE_NAME}" + '.so']
            commands.append("install_name_tool -add_rpath '@loader_path/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/${FLUID_CORE_NAME}" + '.so')
        else:
-            commands = ["patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/${FLUID_CORE_NAME}" + '.so']
+            commands = ["patchelf --set-soname '${FLUID_CORE_NAME}.so' ${PADDLE_BINARY_DIR}/python/paddle/fluid/${FLUID_CORE_NAME}" + '.so']
+            commands.append("patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/${FLUID_CORE_NAME}" + '.so')
        # The sw_64 not suppot patchelf, so we just disable that.
        if platform.machine() != 'sw_64' and platform.machine() != 'mips64':
            for command in commands: