diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake
index 9694a7bc59c12a96e1c0c33488895ae94dbf2a03..8220680cecf9af541d8a4f8bcc48082d1f1074d3 100644
--- a/cmake/inference_lib.cmake
+++ b/cmake/inference_lib.cmake
@@ -146,12 +146,12 @@ copy(inference_lib_dist
         SRCS ${THREADPOOL_INCLUDE_DIR}/ThreadPool.h
         DSTS ${dst_dir})
 
-# Only GPU need cudaErrorMessage.pb
+# GPU must copy externalErrorMsg.pb
 IF(WITH_GPU)
-        set(dst_dir "${PADDLE_INFERENCE_INSTALL_DIR}/third_party/cudaerror/data")
-        copy(inference_lib_dist
-                SRCS ${cudaerror_INCLUDE_DIR}
-                DSTS ${dst_dir})
+    set(dst_dir "${PADDLE_INFERENCE_INSTALL_DIR}/third_party/externalError/data")
+    copy(inference_lib_dist
+            SRCS ${externalError_INCLUDE_DIR}
+            DSTS ${dst_dir})
 ENDIF()
 
 # CMakeCache Info
@@ -259,7 +259,7 @@ copy(fluid_lib_dist
 set(module "platform")
 set(platform_lib_deps profiler_proto error_codes_proto)
 if(WITH_GPU)
-  set(platform_lib_deps ${platform_lib_deps} cuda_error_proto)
+  set(platform_lib_deps ${platform_lib_deps} external_error_proto)
 endif(WITH_GPU)
 
 add_dependencies(fluid_lib_dist ${platform_lib_deps})
diff --git a/cmake/third_party.cmake b/cmake/third_party.cmake
index 56edaff2a50dab0f7029ec1e85fc3d4ce8ac416e..8adc7a4e396e942700b30c87028e5592120dca69 100644
--- a/cmake/third_party.cmake
+++ b/cmake/third_party.cmake
@@ -111,10 +111,11 @@ FUNCTION(file_download_and_uncompress URL NAME)
   MESSAGE(STATUS "Download dependence[${NAME}] from ${URL}")
   SET(${NAME}_INCLUDE_DIR ${THIRD_PARTY_PATH}/${NAME}/data PARENT_SCOPE)
   ExternalProject_Add(
-      extern_download_${NAME}
+      download_${NAME}
       ${EXTERNAL_PROJECT_LOG_ARGS}
       PREFIX                ${THIRD_PARTY_PATH}/${NAME}
       URL                   ${URL}
+      TIMEOUT               120
       DOWNLOAD_DIR          ${THIRD_PARTY_PATH}/${NAME}/data/
       SOURCE_DIR            ${THIRD_PARTY_PATH}/${NAME}/data/
       DOWNLOAD_NO_PROGRESS  1
@@ -123,7 +124,7 @@ FUNCTION(file_download_and_uncompress URL NAME)
       UPDATE_COMMAND        ""
       INSTALL_COMMAND       ""
     )
-  set(third_party_deps ${third_party_deps} extern_download_${NAME} PARENT_SCOPE)
+  set(third_party_deps ${third_party_deps} download_${NAME} PARENT_SCOPE)
 ENDFUNCTION()
 
 
@@ -242,8 +243,20 @@ if(WITH_GPU)
         include(external/cub)       # download cub
         list(APPEND third_party_deps extern_cub)
     endif()
-    set(CUDAERROR_URL  "http://paddlepaddledeps.bj.bcebos.com/cudaErrorMessage.tar.gz" CACHE STRING "" FORCE)
-    file_download_and_uncompress(${CUDAERROR_URL} "cudaerror") # download file cudaErrorMessage
+    set(URL  "https://paddlepaddledeps.bj.bcebos.com/externalErrorMsg.tar.gz" CACHE STRING "" FORCE)
+    file_download_and_uncompress(${URL} "externalError")   # download file externalErrorMsg.tar.gz
+    if(WITH_TESTING)
+        # copy externalErrorMsg.pb for unittest 'enforce_test'
+        set(SRC_DIR ${THIRD_PARTY_PATH}/externalError/data)
+        if(WIN32 AND (NOT "${CMAKE_GENERATOR}" STREQUAL "Ninja"))
+            set(DST_DIR ${CMAKE_BINARY_DIR}/paddle/fluid/third_party/externalError/data)
+        else()
+            set(DST_DIR ${CMAKE_BINARY_DIR}/paddle/third_party/externalError/data)
+        endif()
+        add_custom_command(TARGET download_externalError POST_BUILD
+            COMMAND ${CMAKE_COMMAND} -E copy_directory ${SRC_DIR} ${DST_DIR}
+            COMMENT "copy_directory from ${SRC_DIR} to ${DST_DIR}")
+    endif()
 endif(WITH_GPU)
 
 if(WITH_XPU)
diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt
index 12a54fd7e87f447530888b85f15a66d1e16fdc9c..36a956762174e18ed7eef1d6e1158b82bf3ceeae 100644
--- a/paddle/fluid/platform/CMakeLists.txt
+++ b/paddle/fluid/platform/CMakeLists.txt
@@ -1,7 +1,7 @@
 proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto simple_threadpool)
 proto_library(error_codes_proto SRCS error_codes.proto)
 if(WITH_GPU)
-  proto_library(cuda_error_proto SRCS cuda_error.proto)
+  proto_library(external_error_proto SRCS external_error.proto)
 endif(WITH_GPU)
 
 if(WITH_XPU)
@@ -45,7 +45,7 @@ cc_test(errors_test SRCS errors_test.cc DEPS errors enforce)
 
 set(enforce_deps flags errors boost)
 if(WITH_GPU)
-  set(enforce_deps ${enforce_deps} cuda_error_proto)
+  set(enforce_deps ${enforce_deps} external_error_proto)
 endif()
 cc_library(enforce INTERFACE SRCS enforce.cc DEPS ${enforce_deps})
 cc_library(monitor SRCS monitor.cc)
diff --git a/paddle/fluid/platform/cudnn_helper.h b/paddle/fluid/platform/cudnn_helper.h
index 6c3c96b68c48a1314f4a90a97a2542ea3060446a..0d2a770ad8276b1f8251c19c553bee888ce06d28 100644
--- a/paddle/fluid/platform/cudnn_helper.h
+++ b/paddle/fluid/platform/cudnn_helper.h
@@ -34,35 +34,6 @@ DECLARE_bool(cudnn_deterministic);
 namespace paddle {
 namespace platform {
 
-inline const char* cudnnGetErrorString(cudnnStatus_t status) {
-  switch (status) {
-    case CUDNN_STATUS_SUCCESS:
-      return "CUDNN_STATUS_SUCCESS";
-    case CUDNN_STATUS_NOT_INITIALIZED:
-      return "CUDNN_STATUS_NOT_INITIALIZED";
-    case CUDNN_STATUS_ALLOC_FAILED:
-      return "CUDNN_STATUS_ALLOC_FAILED";
-    case CUDNN_STATUS_BAD_PARAM:
-      return "CUDNN_STATUS_BAD_PARAM";
-    case CUDNN_STATUS_INTERNAL_ERROR:
-      return "CUDNN_STATUS_INTERNAL_ERROR";
-    case CUDNN_STATUS_INVALID_VALUE:
-      return "CUDNN_STATUS_INVALID_VALUE";
-    case CUDNN_STATUS_ARCH_MISMATCH:
-      return "CUDNN_STATUS_ARCH_MISMATCH";
-    case CUDNN_STATUS_MAPPING_ERROR:
-      return "CUDNN_STATUS_MAPPING_ERROR";
-    case CUDNN_STATUS_EXECUTION_FAILED:
-      return "CUDNN_STATUS_EXECUTION_FAILED";
-    case CUDNN_STATUS_NOT_SUPPORTED:
-      return "CUDNN_STATUS_NOT_SUPPORTED";
-    case CUDNN_STATUS_LICENSE_ERROR:
-      return "CUDNN_STATUS_LICENSE_ERROR";
-    default:
-      return "Unknown cudnn error number";
-  }
-}
-
 #define CUDNN_VERSION_MIN(major, minor, patch) \
   (CUDNN_VERSION >= ((major)*1000 + (minor)*100 + (patch)))
 
diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h
index d42733823e669b03daa8f29dfa0c40be38de1069..d3890de89a5d140bfa09d04909f703f4ca771a05 100644
--- a/paddle/fluid/platform/enforce.h
+++ b/paddle/fluid/platform/enforce.h
@@ -34,7 +34,7 @@ limitations under the License. */
 #include <curand.h>
 #include <thrust/system/cuda/error.h>
 #include <thrust/system_error.h>
-#include "paddle/fluid/platform/cuda_error.pb.h"
+#include "paddle/fluid/platform/external_error.pb.h"
 #endif  // PADDLE_WITH_CUDA
 
 #ifdef PADDLE_WITH_HIP
@@ -682,41 +682,83 @@ struct EOFException : public std::exception {
     END_HANDLE_THE_ERROR                                                     \
   } while (0)
 
-/** CUDA PADDLE ENFORCE FUNCTIONS AND MACROS **/
+/**************************************************************************/
+/**************************** NVIDIA ERROR ********************************/
 #ifdef PADDLE_WITH_CUDA
 
-/***** CUDA ERROR *****/
-inline bool is_error(cudaError_t e) { return e != cudaSuccess; }
+namespace details {
 
-inline std::string GetCudaErrorWebsite(int32_t cuda_version) {
-  std::ostringstream webstr;
-  webstr << "https://docs.nvidia.com/cuda/";
-  if (cuda_version != -1) {
-    double version = cuda_version / 10;
-    webstr << "archive/" << std::fixed << std::setprecision(1) << version;
+template <typename T>
+struct ExternalApiType {};
+
+#define DEFINE_EXTERNAL_API_TYPE(type, success_value, proto_type) \
+  template <>                                                     \
+  struct ExternalApiType<type> {                                  \
+    using Type = type;                                            \
+    static constexpr Type kSuccess = success_value;               \
+    static constexpr const char* kTypeString = #proto_type;       \
+    static constexpr platform::proto::ApiType kProtoType =        \
+        platform::proto::ApiType::proto_type;                     \
   }
-  webstr << "/cuda-runtime-api/group__CUDART__TYPES.html"
-            "#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0038";
-  return webstr.str();
-}
 
-inline std::string build_nvidia_error_msg(cudaError_t e) {
-#if CUDA_VERSION >= 10000 && CUDA_VERSION < 11000
-  int32_t cuda_version = 100;
-#elif CUDA_VERSION >= 9000
-  int32_t cuda_version = 90;
-#else
-  int32_t cuda_version = -1;
+DEFINE_EXTERNAL_API_TYPE(cudaError_t, cudaSuccess, CUDA);
+DEFINE_EXTERNAL_API_TYPE(curandStatus_t, CURAND_STATUS_SUCCESS, CURAND);
+DEFINE_EXTERNAL_API_TYPE(cudnnStatus_t, CUDNN_STATUS_SUCCESS, CUDNN);
+DEFINE_EXTERNAL_API_TYPE(cublasStatus_t, CUBLAS_STATUS_SUCCESS, CUBLAS);
+DEFINE_EXTERNAL_API_TYPE(cusolverStatus_t, CUSOLVER_STATUS_SUCCESS, CUSOLVER);
+
+#if !defined(__APPLE__) && defined(PADDLE_WITH_NCCL)
+DEFINE_EXTERNAL_API_TYPE(ncclResult_t, ncclSuccess, NCCL);
 #endif
+
+}  // namespace details
+
+template <typename T>
+inline const char* GetErrorMsgUrl(T status) {
+  using __CUDA_STATUS_TYPE__ = decltype(status);
+  platform::proto::ApiType proto_type =
+      details::ExternalApiType<__CUDA_STATUS_TYPE__>::kProtoType;
+  switch (proto_type) {
+    case platform::proto::ApiType::CUDA:
+      return "https://docs.nvidia.com/cuda/cuda-runtime-api/"
+             "group__CUDART__TYPES.html#group__CUDART__TYPES_"
+             "1g3f51e3575c2178246db0a94a430e0038";
+      break;
+    case platform::proto::ApiType::CURAND:
+      return "https://docs.nvidia.com/cuda/curand/"
+             "group__HOST.html#group__HOST_1gb94a31d5c165858c96b6c18b70644437";
+      break;
+    case platform::proto::ApiType::CUDNN:
+      return "https://docs.nvidia.com/deeplearning/cudnn/api/"
+             "index.html#cudnnStatus_t";
+      break;
+    case platform::proto::ApiType::CUBLAS:
+      return "https://docs.nvidia.com/cuda/cublas/index.html#cublasstatus_t";
+      break;
+    case platform::proto::ApiType::CUSOLVER:
+      return "https://docs.nvidia.com/cuda/cusolver/"
+             "index.html#cuSolverSPstatus";
+      break;
+    case platform::proto::ApiType::NCCL:
+      return "https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/api/"
+             "types.html#ncclresult-t";
+      break;
+    default:
+      return "Unknown type of External API, can't get error message URL!";
+      break;
+  }
+}
+
+template <typename T>
+inline std::string GetExternalErrorMsg(T status) {
   std::ostringstream sout;
-  sout << " Cuda error(" << e << "), " << cudaGetErrorString(e) << ".";
-  static platform::proto::cudaerrorDesc cudaerror;
-  static bool _initSucceed = false;
-  if (cudaerror.ByteSizeLong() == 0) {
+  bool _initSucceed = false;
+  platform::proto::ExternalErrorDesc externalError;
+  if (externalError.ByteSizeLong() == 0) {
     std::string filePath;
 #if !defined(_WIN32)
     Dl_info info;
-    if (dladdr(reinterpret_cast<void*>(GetCudaErrorWebsite), &info)) {
+    if (dladdr(reinterpret_cast<void*>(GetCurrentTraceBackString), &info)) {
       std::string strModule(info.dli_fname);
       const size_t last_slash_idx = strModule.find_last_of("/");
       std::string compare_path = strModule.substr(strModule.length() - 6);
@@ -724,18 +766,19 @@ inline std::string build_nvidia_error_msg(cudaError_t e) {
         strModule.erase(last_slash_idx, std::string::npos);
       }
       if (compare_path.compare("avx.so") == 0) {
-        filePath = strModule +
-                   "/../include/third_party/cudaerror/data/cudaErrorMessage.pb";
-      } else {
         filePath =
-            strModule + "/../../thirl_party/cudaerror/data/cudaErrorMessage.pb";
+            strModule +
+            "/../include/third_party/externalError/data/externalErrorMsg.pb";
+      } else {
+        filePath = strModule +
+                   "/../../third_party/externalError/data/externalErrorMsg.pb";
       }
     }
 #else
     char buf[100];
     MEMORY_BASIC_INFORMATION mbi;
     HMODULE h_module =
-        (::VirtualQuery(GetCudaErrorWebsite, &mbi, sizeof(mbi)) != 0)
+        (::VirtualQuery(GetCurrentTraceBackString, &mbi, sizeof(mbi)) != 0)
             ? (HMODULE)mbi.AllocationBase
             : NULL;
     GetModuleFileName(h_module, buf, 100);
@@ -746,198 +789,118 @@ inline std::string build_nvidia_error_msg(cudaError_t e) {
       strModule.erase(last_slash_idx, std::string::npos);
     }
     if (compare_path.compare("avx.pyd") == 0) {
-      filePath =
-          strModule +
-          "\\..\\include\\third_party\\cudaerror\\data\\cudaErrorMessage.pb";
+      filePath = strModule +
+                 "\\..\\include\\third_"
+                 "party\\externalerror\\data\\externalErrorMsg.pb";
     } else {
       filePath =
-          strModule + "\\..\\third_party\\cudaerror\\data\\cudaErrorMessage.pb";
+          strModule +
+          "\\..\\..\\third_party\\externalerror\\data\\externalErrorMsg.pb";
     }
 #endif
     std::ifstream fin(filePath, std::ios::in | std::ios::binary);
-    _initSucceed = cudaerror.ParseFromIstream(&fin);
+    _initSucceed = externalError.ParseFromIstream(&fin);
   }
+  using __CUDA_STATUS_TYPE__ = decltype(status);
+  platform::proto::ApiType proto_type =
+      details::ExternalApiType<__CUDA_STATUS_TYPE__>::kProtoType;
   if (_initSucceed) {
-    for (int i = 0; i < cudaerror.allmessages_size(); ++i) {
-      if (cuda_version == cudaerror.allmessages(i).version()) {
-        for (int j = 0; j < cudaerror.allmessages(i).messages_size(); ++j) {
-          if (e == cudaerror.allmessages(i).messages(j).errorcode()) {
-            sout << "\n  [Advise: "
-                 << cudaerror.allmessages(i).messages(j).errormessage() << "]";
+    for (int i = 0; i < externalError.errors_size(); ++i) {
+      if (proto_type == externalError.errors(i).type()) {
+        for (int j = 0; j < externalError.errors(i).messages_size(); ++j) {
+          if (status == externalError.errors(i).messages(j).code()) {
+            sout << "\n  [Hint: "
+                 << externalError.errors(i).messages(j).message() << "]";
             return sout.str();
           }
         }
       }
     }
   }
-  sout << "\n  [Advise: Please search for the error code(" << e
-       << ") on website( " << GetCudaErrorWebsite(cuda_version)
-       << " ) to get Nvidia's official solution about CUDA Error.]";
+
+  sout << "\n  [Hint: Please search for the error code(" << status
+       << ") on website (" << GetErrorMsgUrl(status)
+       << ") to get Nvidia's official solution and advice about "
+       << details::ExternalApiType<__CUDA_STATUS_TYPE__>::kTypeString
+       << " Error.]";
   return sout.str();
 }
 
-/** curand ERROR **/
-inline bool is_error(curandStatus_t stat) {
-  return stat != CURAND_STATUS_SUCCESS;
+template std::string GetExternalErrorMsg<cudaError_t>(cudaError_t);
+template std::string GetExternalErrorMsg<curandStatus_t>(curandStatus_t);
+template std::string GetExternalErrorMsg<cudnnStatus_t>(cudnnStatus_t);
+template std::string GetExternalErrorMsg<cublasStatus_t>(cublasStatus_t);
+template std::string GetExternalErrorMsg<cusolverStatus_t>(cusolverStatus_t);
+#if !defined(__APPLE__) && defined(PADDLE_WITH_NCCL)
+template std::string GetExternalErrorMsg<ncclResult_t>(ncclResult_t);
+#endif
+
+/*************** CUDA ERROR ***************/
+inline bool is_error(cudaError_t e) { return e != cudaSuccess; }
+
+inline std::string build_nvidia_error_msg(cudaError_t e) {
+  std::ostringstream sout;
+  sout << "CUDA error(" << e << "), " << cudaGetErrorString(e) << ". "
+       << GetExternalErrorMsg(e);
+  return sout.str();
 }
 
-inline const char* curandGetErrorString(curandStatus_t stat) {
-  switch (stat) {
-    case CURAND_STATUS_SUCCESS:
-      return "`CURAND_STATUS_SUCCESS`. No errors.";
-    case CURAND_STATUS_VERSION_MISMATCH:
-      return "`CURAND_STATUS_VERSION_MISMATCH`. Header file and linked library "
-             "version do not match.";
-    case CURAND_STATUS_NOT_INITIALIZED:
-      return "`CURAND_STATUS_NOT_INITIALIZED`. Generator not initialized.";
-    case CURAND_STATUS_ALLOCATION_FAILED:
-      return "`CURAND_STATUS_ALLOCATION_FAILED`. Memory allocation failed.";
-    case CURAND_STATUS_TYPE_ERROR:
-      return "`CURAND_STATUS_TYPE_ERROR`. Generator is wrong type.";
-    case CURAND_STATUS_OUT_OF_RANGE:
-      return "`CURAND_STATUS_OUT_OF_RANGE`. Argument out of range.";
-    case CURAND_STATUS_LENGTH_NOT_MULTIPLE:
-      return "`CURAND_STATUS_LENGTH_NOT_MULTIPLE`. Length requested is not a "
-             "multple of dimension.";
-    case CURAND_STATUS_DOUBLE_PRECISION_REQUIRED:
-      return "`CURAND_STATUS_DOUBLE_PRECISION_REQUIRED`. GPU does not have "
-             "double precision required by MRG32k3a.";
-    case CURAND_STATUS_LAUNCH_FAILURE:
-      return "`CURAND_STATUS_LAUNCH_FAILURE`. Kernel launch failure.";
-    case CURAND_STATUS_PREEXISTING_FAILURE:
-      return "`CURAND_STATUS_PREEXISTING_FAILURE`. Preexisting failure on "
-             "library entry.";
-    case CURAND_STATUS_INITIALIZATION_FAILED:
-      return "`CURAND_STATUS_INITIALIZATION_FAILED`. Initialization of CUDA "
-             "failed.";
-    case CURAND_STATUS_ARCH_MISMATCH:
-      return "`CURAND_STATUS_ARCH_MISMATCH`. Architecture mismatch, GPU does "
-             "not support requested feature.";
-    case CURAND_STATUS_INTERNAL_ERROR:
-      return "`CURAND_STATUS_INTERNAL_ERROR`. Internal library error.";
-    default:
-      return "Unknown curand status";
-  }
+/*************** CURAND ERROR ***************/
+inline bool is_error(curandStatus_t stat) {
+  return stat != CURAND_STATUS_SUCCESS;
 }
 
 inline std::string build_nvidia_error_msg(curandStatus_t stat) {
-  std::string msg(" Curand error, ");
-  return msg + curandGetErrorString(stat) + " ";
+  std::ostringstream sout;
+  sout << "CURAND error(" << stat << "). " << GetExternalErrorMsg(stat);
+  return sout.str();
 }
 
-/***** CUDNN ERROR *****/
+/*************** CUDNN ERROR ***************/
 inline bool is_error(cudnnStatus_t stat) {
   return stat != CUDNN_STATUS_SUCCESS;
 }
 
 inline std::string build_nvidia_error_msg(cudnnStatus_t stat) {
-  std::string msg(" Cudnn error, ");
-  return msg + platform::dynload::cudnnGetErrorString(stat) + " ";
+  std::ostringstream sout;
+  sout << "CUDNN error(" << stat << "), "
+       << platform::dynload::cudnnGetErrorString(stat) << ". "
+       << GetExternalErrorMsg(stat);
+  return sout.str();
 }
 
-/***** CUBLAS ERROR *****/
+/*************** CUBLAS ERROR ***************/
 inline bool is_error(cublasStatus_t stat) {
   return stat != CUBLAS_STATUS_SUCCESS;
 }
 
-inline const char* cublasGetErrorString(cublasStatus_t stat) {
-  switch (stat) {
-    case CUBLAS_STATUS_NOT_INITIALIZED:
-      return "`CUBLAS_STATUS_NOT_INITIALIZED`. The cuBLAS library was not "
-             "initialized.";
-    case CUBLAS_STATUS_ALLOC_FAILED:
-      return "`CUBLAS_STATUS_ALLOC_FAILED`. Resource allocation failed inside "
-             "the cuBLAS library.";
-    case CUBLAS_STATUS_INVALID_VALUE:
-      return "`CUBLAS_STATUS_INVALID_VALUE`. An unsupported value or parameter "
-             "was passed to the function (a negative vector size, for "
-             "example).";
-    case CUBLAS_STATUS_ARCH_MISMATCH:
-      return "`CUBLAS_STATUS_ARCH_MISMATCH`. The function requires a feature "
-             "absent from the device architecture; usually caused by the lack "
-             "of support for double precision.";
-    case CUBLAS_STATUS_MAPPING_ERROR:
-      return "`CUBLAS_STATUS_MAPPING_ERROR`. An access to GPU memory space "
-             "failed, which is usually caused by a failure to bind a texture.";
-    case CUBLAS_STATUS_EXECUTION_FAILED:
-      return "`CUBLAS_STATUS_EXECUTION_FAILED`. The GPU program failed to "
-             "execute. This is often caused by a launch failure of the kernel "
-             "on the GPU, which can be caused by multiple reasons.";
-    case CUBLAS_STATUS_INTERNAL_ERROR:
-      return "`CUBLAS_STATUS_INTERNAL_ERROR`. An internal cuBLAS operation "
-             "failed. This error is usually caused by a cudaMemcpyAsync() "
-             "failure.";
-    case CUBLAS_STATUS_NOT_SUPPORTED:
-      return "`CUBLAS_STATUS_NOT_SUPPORTED`. The functionality requested is "
-             "not supported.";
-    case CUBLAS_STATUS_LICENSE_ERROR:
-      return "`CUBLAS_STATUS_LICENSE_ERROR`. The functionality requested "
-             "requires some license and an error was detected when trying to "
-             "check the current licensing.";
-    default:
-      return "Unknown cublas status";
-  }
-}
-
 inline std::string build_nvidia_error_msg(cublasStatus_t stat) {
-  std::string msg(" Cublas error, ");
-  return msg + cublasGetErrorString(stat) + " ";
+  std::ostringstream sout;
+  sout << "CUBLAS error(" << stat << "). " << GetExternalErrorMsg(stat);
+  return sout.str();
 }
 
-/***** CUSOLVER ERROR *****/
+/*************** CUSOLVER ERROR ***************/
 inline bool is_error(cusolverStatus_t stat) {
   return stat != CUSOLVER_STATUS_SUCCESS;
 }
 
-inline const char* cusolverGetErrorString(cusolverStatus_t stat) {
-  switch (stat) {
-    case CUSOLVER_STATUS_NOT_INITIALIZED:
-      return "`CUSOLVER_STATUS_NOT_INITIALIZED`. The cuSolver library was not "
-             "initialized. This is usually caused by the lack of a prior call, "
-             "an error in the CUDA Runtime API called by the cuSolver routine, "
-             "or an error in the hardware setup.";
-    case CUSOLVER_STATUS_ALLOC_FAILED:
-      return "`CUSOLVER_STATUS_ALLOC_FAILED`. Resource allocation failed "
-             "inside the cuSolver library. This is usually caused by a "
-             "cudaMalloc() failure.";
-    case CUSOLVER_STATUS_INVALID_VALUE:
-      return "`CUSOLVER_STATUS_INVALID_VALUE`. An unsupported value or "
-             "parameter was passed to the function (a negative vector size, "
-             "for example).";
-    case CUSOLVER_STATUS_ARCH_MISMATCH:
-      return "`CUSOLVER_STATUS_ARCH_MISMATCH`. The function requires a feature "
-             "absent from the device architecture; usually caused by the lack "
-             "of support for atomic operations or double precision.";
-    case CUSOLVER_STATUS_EXECUTION_FAILED:
-      return "`CUSOLVER_STATUS_EXECUTION_FAILED`. The GPU program failed to "
-             "execute. This is often caused by a launch failure of the kernel "
-             "on the GPU, which can be caused by multiple reasons.";
-    case CUSOLVER_STATUS_INTERNAL_ERROR:
-      return "`CUSOLVER_STATUS_INTERNAL_ERROR`. An internal cuSolver operation "
-             "failed. This error is usually caused by a cudaMemcpyAsync() "
-             "failure.";
-    case CUSOLVER_STATUS_MATRIX_TYPE_NOT_SUPPORTED:
-      return "`CUSOLVER_STATUS_MATRIX_TYPE_NOT_SUPPORTED`. The matrix type is "
-             "not supported by this function. This is usually caused by "
-             "passing an invalid matrix descriptor to the function.";
-    default:
-      return "Unknown cusolver status";
-  }
-}
-
 inline std::string build_nvidia_error_msg(cusolverStatus_t stat) {
-  std::string msg(" Cublas error, ");
-  return msg + cusolverGetErrorString(stat) + " ";
+  std::ostringstream sout;
+  sout << "CUSOLVER error(" << stat << "). " << GetExternalErrorMsg(stat);
+  return sout.str();
 }
 
-/****** NCCL ERROR ******/
+/**************** NCCL ERROR ****************/
 #if !defined(__APPLE__) && defined(PADDLE_WITH_NCCL)
 inline bool is_error(ncclResult_t nccl_result) {
   return nccl_result != ncclSuccess;
 }
 
 inline std::string build_nvidia_error_msg(ncclResult_t nccl_result) {
-  std::string msg(" Nccl error, ");
+  std::ostringstream sout;
+  sout << "NCCL error(" << nccl_result << "), "
+       << platform::dynload::ncclGetErrorString(nccl_result) << ". ";
   if (errno == ENOSPC || errno == EAGAIN) {
     std::string detail(strerror(errno));
     detail += "\nPlease try one of the following solutions:";
@@ -947,42 +910,19 @@ inline std::string build_nvidia_error_msg(ncclResult_t nccl_result) {
         "\n3. Increase shared memory by setting the -shm-size "
         "option when starting docker container, e.g., setting "
         " -shm-size=2g.\n";
-    return msg + platform::dynload::ncclGetErrorString(nccl_result) +
-           ", detail: " + detail + " ";
+    sout << " Detail: " + detail;
   }
-  return msg + platform::dynload::ncclGetErrorString(nccl_result) + " ";
+  sout << GetExternalErrorMsg(nccl_result);
+  return sout.str();
 }
 #endif  // not(__APPLE__) and PADDLE_WITH_NCCL
 
-namespace details {
-
-template <typename T>
-struct CudaStatusType {};
-
-#define DEFINE_CUDA_STATUS_TYPE(type, success_value) \
-  template <>                                        \
-  struct CudaStatusType<type> {                      \
-    using Type = type;                               \
-    static constexpr Type kSuccess = success_value;  \
-  }
-
-DEFINE_CUDA_STATUS_TYPE(cudaError_t, cudaSuccess);
-DEFINE_CUDA_STATUS_TYPE(curandStatus_t, CURAND_STATUS_SUCCESS);
-DEFINE_CUDA_STATUS_TYPE(cudnnStatus_t, CUDNN_STATUS_SUCCESS);
-DEFINE_CUDA_STATUS_TYPE(cublasStatus_t, CUBLAS_STATUS_SUCCESS);
-DEFINE_CUDA_STATUS_TYPE(cusolverStatus_t, CUSOLVER_STATUS_SUCCESS);
-
-#if !defined(__APPLE__) && defined(PADDLE_WITH_NCCL)
-DEFINE_CUDA_STATUS_TYPE(ncclResult_t, ncclSuccess);
-#endif
-}  // namespace details
-
 #define PADDLE_ENFORCE_CUDA_SUCCESS(COND)                        \
   do {                                                           \
     auto __cond__ = (COND);                                      \
     using __CUDA_STATUS_TYPE__ = decltype(__cond__);             \
     constexpr auto __success_type__ =                            \
-        ::paddle::platform::details::CudaStatusType<             \
+        ::paddle::platform::details::ExternalApiType<            \
             __CUDA_STATUS_TYPE__>::kSuccess;                     \
     if (UNLIKELY(__cond__ != __success_type__)) {                \
       auto __summary__ = ::paddle::platform::errors::External(   \
@@ -1023,7 +963,7 @@ inline void retry_sleep(unsigned milliseconds) {
     int retry_count = 1;                                                \
     using __CUDA_STATUS_TYPE__ = decltype(__cond__);                    \
     constexpr auto __success_type__ =                                   \
-        ::paddle::platform::details::CudaStatusType<                    \
+        ::paddle::platform::details::ExternalApiType<                   \
             __CUDA_STATUS_TYPE__>::kSuccess;                            \
     while (UNLIKELY(__cond__ != __success_type__) && retry_count < 5) { \
       retry_sleep(FLAGS_gpu_allocator_retry_time);                      \
@@ -1037,10 +977,11 @@ inline void retry_sleep(unsigned milliseconds) {
     }                                                                   \
   } while (0)
 
-#undef DEFINE_CUDA_STATUS_TYPE
+#undef DEFINE_EXTERNAL_API_TYPE
 #endif  // PADDLE_WITH_CUDA
 
-/** HIP PADDLE ENFORCE FUNCTIONS AND MACROS **/
+/**************************************************************************/
+/***************************** HIP ERROR **********************************/
 #ifdef PADDLE_WITH_HIP
 
 /***** HIP ERROR *****/
@@ -1052,7 +993,7 @@ inline std::string build_rocm_error_msg(hipError_t e) {
   return sout.str();
 }
 
-/** HIPRAND ERROR **/
+/***** HIPRAND ERROR *****/
 inline bool is_error(hiprandStatus_t stat) {
   return stat != HIPRAND_STATUS_SUCCESS;
 }
@@ -1153,22 +1094,22 @@ inline std::string build_rocm_error_msg(ncclResult_t nccl_result) {
 namespace details {
 
 template <typename T>
-struct CudaStatusType {};
+struct ExternalApiType {};
 
-#define DEFINE_CUDA_STATUS_TYPE(type, success_value) \
-  template <>                                        \
-  struct CudaStatusType<type> {                      \
-    using Type = type;                               \
-    static constexpr Type kSuccess = success_value;  \
+#define DEFINE_EXTERNAL_API_TYPE(type, success_value) \
+  template <>                                         \
+  struct ExternalApiType<type> {                      \
+    using Type = type;                                \
+    static constexpr Type kSuccess = success_value;   \
   }
 
-DEFINE_CUDA_STATUS_TYPE(hipError_t, hipSuccess);
-DEFINE_CUDA_STATUS_TYPE(hiprandStatus_t, HIPRAND_STATUS_SUCCESS);
-DEFINE_CUDA_STATUS_TYPE(miopenStatus_t, miopenStatusSuccess);
-DEFINE_CUDA_STATUS_TYPE(rocblas_status, rocblas_status_success);
+DEFINE_EXTERNAL_API_TYPE(hipError_t, hipSuccess);
+DEFINE_EXTERNAL_API_TYPE(hiprandStatus_t, HIPRAND_STATUS_SUCCESS);
+DEFINE_EXTERNAL_API_TYPE(miopenStatus_t, miopenStatusSuccess);
+DEFINE_EXTERNAL_API_TYPE(rocblas_status, rocblas_status_success);
 
 #if !defined(__APPLE__) && defined(PADDLE_WITH_RCCL)
-DEFINE_CUDA_STATUS_TYPE(ncclResult_t, ncclSuccess);
+DEFINE_EXTERNAL_API_TYPE(ncclResult_t, ncclSuccess);
 #endif
 
 }  // namespace details
@@ -1178,7 +1119,7 @@ DEFINE_CUDA_STATUS_TYPE(ncclResult_t, ncclSuccess);
     auto __cond__ = (COND);                                    \
     using __CUDA_STATUS_TYPE__ = decltype(__cond__);           \
     constexpr auto __success_type__ =                          \
-        ::paddle::platform::details::CudaStatusType<           \
+        ::paddle::platform::details::ExternalApiType<          \
             __CUDA_STATUS_TYPE__>::kSuccess;                   \
     if (UNLIKELY(__cond__ != __success_type__)) {              \
       auto __summary__ = ::paddle::platform::errors::External( \
@@ -1201,7 +1142,7 @@ inline void retry_sleep(unsigned millisecond) {
     int retry_count = 1;                                                \
     using __CUDA_STATUS_TYPE__ = decltype(__cond__);                    \
     constexpr auto __success_type__ =                                   \
-        ::paddle::platform::details::CudaStatusType<                    \
+        ::paddle::platform::details::ExternalApiType<                   \
             __CUDA_STATUS_TYPE__>::kSuccess;                            \
     while (UNLIKELY(__cond__ != __success_type__) && retry_count < 5) { \
       retry_sleep(FLAGS_gpu_allocator_retry_time);                      \
@@ -1215,7 +1156,7 @@ inline void retry_sleep(unsigned millisecond) {
     }                                                                   \
   } while (0)
 
-#undef DEFINE_CUDA_STATUS_TYPE
+#undef DEFINE_EXTERNAL_API_TYPE
 #endif  // PADDLE_WITH_HIP
 
 #ifdef PADDLE_WITH_ASCEND_CL
diff --git a/paddle/fluid/platform/enforce_test.cc b/paddle/fluid/platform/enforce_test.cc
index 39f3d3f00c9997eea3f4ab1e5652fcc78f1be0a6..842d4cc139281aab48131759f63003b3fe3890c2 100644
--- a/paddle/fluid/platform/enforce_test.cc
+++ b/paddle/fluid/platform/enforce_test.cc
@@ -304,6 +304,7 @@ bool CheckCudaStatusFailure(T value, const std::string& msg) {
     return false;
   } catch (paddle::platform::EnforceNotMet& error) {
     std::string ex_msg = error.what();
+    std::cout << ex_msg << std::endl;
     return ex_msg.find(msg) != std::string::npos;
   }
 }
@@ -338,30 +339,98 @@ TEST(enforce, hip_success) {
 #else
 TEST(enforce, cuda_success) {
   EXPECT_TRUE(CheckCudaStatusSuccess(cudaSuccess));
-  EXPECT_TRUE(CheckCudaStatusFailure(cudaErrorInvalidValue, "Cuda error"));
-  EXPECT_TRUE(CheckCudaStatusFailure(cudaErrorMemoryAllocation, "Cuda error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(cudaErrorInvalidValue, "CUDA error"));
+
+  EXPECT_TRUE(CheckCudaStatusFailure(cudaErrorMemoryAllocation, "CUDA error"));
+
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      cudaErrorInsufficientDriver,
+      "This indicates that the installed NVIDIA CUDA driver is older than the "
+      "CUDA runtime library. This is not a supported configuration.Users "
+      "should install an updated NVIDIA display driver to allow the "
+      "application to run"));
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      cudaErrorContextIsDestroyed,
+      "This error indicates that the context current to the calling thread has "
+      "been destroyed using cuCtxDestroy, or is a primary context which has "
+      "not yet been initialized"));
 
   EXPECT_TRUE(CheckCudaStatusSuccess(CURAND_STATUS_SUCCESS));
   EXPECT_TRUE(
-      CheckCudaStatusFailure(CURAND_STATUS_VERSION_MISMATCH, "Curand error"));
+      CheckCudaStatusFailure(CURAND_STATUS_VERSION_MISMATCH, "CURAND error"));
   EXPECT_TRUE(
-      CheckCudaStatusFailure(CURAND_STATUS_NOT_INITIALIZED, "Curand error"));
+      CheckCudaStatusFailure(CURAND_STATUS_NOT_INITIALIZED, "CURAND error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      CURAND_STATUS_ARCH_MISMATCH,
+      "Architecture mismatch, GPU does not support requested feature"));
+  EXPECT_TRUE(
+      CheckCudaStatusFailure(CURAND_STATUS_LENGTH_NOT_MULTIPLE,
+                             "Length requested is not a multple of dimension"));
 
   EXPECT_TRUE(CheckCudaStatusSuccess(CUDNN_STATUS_SUCCESS));
   EXPECT_TRUE(
-      CheckCudaStatusFailure(CUDNN_STATUS_NOT_INITIALIZED, "Cudnn error"));
-  EXPECT_TRUE(CheckCudaStatusFailure(CUDNN_STATUS_ALLOC_FAILED, "Cudnn error"));
+      CheckCudaStatusFailure(CUDNN_STATUS_NOT_INITIALIZED, "CUDNN error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(CUDNN_STATUS_ALLOC_FAILED, "CUDNN error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      CUDNN_STATUS_BAD_PARAM,
+      "An incorrect value or parameter was passed to the function. To correct, "
+      "ensure that all the parameters being passed have valid values"));
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      CUDNN_STATUS_LICENSE_ERROR,
+      "The functionality requested requires some license and an error was "
+      "detected when trying to check the current licensing. This error can "
+      "happen if the license is not present or is expired or if the "
+      "environment variable NVIDIA_LICENSE_FILE is not set properly"));
 
   EXPECT_TRUE(CheckCudaStatusSuccess(CUBLAS_STATUS_SUCCESS));
   EXPECT_TRUE(
-      CheckCudaStatusFailure(CUBLAS_STATUS_NOT_INITIALIZED, "Cublas error"));
+      CheckCudaStatusFailure(CUBLAS_STATUS_NOT_INITIALIZED, "CUBLAS error"));
+  EXPECT_TRUE(
+      CheckCudaStatusFailure(CUBLAS_STATUS_INVALID_VALUE, "CUBLAS error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      CUBLAS_STATUS_EXECUTION_FAILED,
+      "The GPU program failed to execute. This is often caused by a launch "
+      "failure of the kernel on the GPU, which can be caused by multiple "
+      "reasons.  To correct: check that the hardware, an appropriate version "
+      "of the driver, and the cuBLAS library are correctly installed"));
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      CUBLAS_STATUS_MAPPING_ERROR,
+      "An access to GPU memory space failed, which is usually caused by a "
+      "failure to bind a texture. To correct: prior to the function call, "
+      "unbind any previously bound textures"));
+
+  EXPECT_TRUE(CheckCudaStatusSuccess(CUSOLVER_STATUS_SUCCESS));
+  EXPECT_TRUE(CheckCudaStatusFailure(CUSOLVER_STATUS_NOT_INITIALIZED,
+                                     "CUSOLVER error"));
   EXPECT_TRUE(
-      CheckCudaStatusFailure(CUBLAS_STATUS_INVALID_VALUE, "Cublas error"));
+      CheckCudaStatusFailure(CUSOLVER_STATUS_ALLOC_FAILED, "CUSOLVER error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      CUSOLVER_STATUS_INTERNAL_ERROR,
+      "An internal cuSolver operation failed. This error is usually caused by "
+      "a cudaMemcpyAsync() failure.To correct: check that the hardware, an "
+      "appropriate version of the driver, and the cuSolver library are "
+      "correctly installed. Also, check that the memory passed as a parameter "
+      "to the routine is not being deallocated prior to the routine’s "
+      "completion"));
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      CUSOLVER_STATUS_INVALID_VALUE,
+      "An unsupported value or parameter was passed to the function (a "
+      "negative vector size, for example).To correct: ensure that all the "
+      "parameters being passed have valid values"));
+  /*
 #if !defined(__APPLE__) && defined(PADDLE_WITH_NCCL)
   EXPECT_TRUE(CheckCudaStatusSuccess(ncclSuccess));
-  EXPECT_TRUE(CheckCudaStatusFailure(ncclUnhandledCudaError, "Nccl error"));
-  EXPECT_TRUE(CheckCudaStatusFailure(ncclSystemError, "Nccl error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(ncclUnhandledCudaError, "NCCL error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(ncclSystemError, "NCCL error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(ncclInternalError,
+                                     "An internal check failed. This is either "
+                                     "a bug in NCCL or due to memory "
+                                     "corruption"));
+  EXPECT_TRUE(CheckCudaStatusFailure(ncclInvalidUsage,
+                                     "The call to NCCL is incorrect. This is "
+                                     "usually reflecting a programming error"));
 #endif
+*/
 }
 #endif
 #endif
diff --git a/paddle/fluid/platform/cuda_error.proto b/paddle/fluid/platform/external_error.proto
similarity index 58%
rename from paddle/fluid/platform/cuda_error.proto
rename to paddle/fluid/platform/external_error.proto
index b55e0af81ee6f8fb47d558287c7f902ef0fde81b..2094de7e10f69e98cc450d4221a85c6f904770ed 100644
--- a/paddle/fluid/platform/cuda_error.proto
+++ b/paddle/fluid/platform/external_error.proto
@@ -15,21 +15,32 @@ limitations under the License. */
 syntax = "proto2";
 package paddle.platform.proto;
 
+// (NOTE:zhouwei): ApiType describes which kind of external third party API
+// More external third party API can be added.
+enum ApiType {
+  CUDA = 0;
+  CURAND = 1;
+  CUDNN = 2;
+  CUBLAS = 3;
+  CUSOLVER = 4;
+  NCCL = 5;
+}
+
 message MessageDesc {
-  // Indicates the type of error
-  required int32 errorCode = 1;
+  // Indicates the code of error
+  required int32 code = 1;
   // Indicates the message of error
-  required string errorMessage = 2;
+  required string message = 2;
 }
 
 message AllMessageDesc {
-  // Version of cuda API
-  required int32 version = 1;
+  // Indicates which kind of third-party API
+  required ApiType type = 1;
   // Error messages of different errortype
-  repeated MessageDesc Messages = 2;
+  repeated MessageDesc messages = 2;
 }
 
-message cudaerrorDesc {
-  // Error messages of different cuda versions(9.0/10.0/10.2)
-  repeated AllMessageDesc AllMessages = 2;
+message ExternalErrorDesc {
+  // Error messages of different kind of external third party API
+  repeated AllMessageDesc errors = 1;
 }
\ No newline at end of file
diff --git a/paddle/scripts/paddle_build.bat b/paddle/scripts/paddle_build.bat
index dd8146aa3a1147b2d4e77185647be720468043f7..8c323490cc964c5e3b69d6b512fdee22041d9803 100644
--- a/paddle/scripts/paddle_build.bat
+++ b/paddle/scripts/paddle_build.bat
@@ -76,6 +76,7 @@ if not defined PYTHON_ROOT set PYTHON_ROOT=C:\Python37
 
 rem -------set cache build directory-----------
 rmdir build\python /s/q
+rmdir build\paddle\third_party\externalError /s/q
 rmdir build\paddle\fluid\pybind /s/q
 rmdir build\paddle_install_dir /s/q
 rmdir build\paddle_inference_install_dir /s/q
@@ -506,7 +507,6 @@ echo    ========================================
 echo    Step 4. Running unit tests ...
 echo    ========================================
 
-
 : set CI_SKIP_CPP_TEST if only *.py changed
 git diff --name-only %BRANCH% | findstr /V "\.py" || set CI_SKIP_CPP_TEST=ON
 
diff --git a/python/setup.py.in b/python/setup.py.in
index 79c67182f9c7911aaae32dfaf660c49dbe683e1c..3fbe796a8137d8da774e9fb4b464f104f6cf154a 100644
--- a/python/setup.py.in
+++ b/python/setup.py.in
@@ -412,7 +412,8 @@ if '${WITH_MKLDNN}' == 'ON':
     headers += list(find_files('*', '${MKLDNN_INSTALL_DIR}/include')) # mkldnn
 
 if '${WITH_GPU}' == 'ON' or '${WITH_ROCM}' == 'ON':
-    headers += list(find_files('*.pb', '${cudaerror_INCLUDE_DIR}'))   # errorMessage.pb for errormessage
+    # externalErrorMsg.pb for External Error message
+    headers += list(find_files('*.pb', '${externalError_INCLUDE_DIR}'))
 
 class InstallCommand(InstallCommandBase):
     def finalize_options(self):
diff --git a/tools/cudaError/README.md b/tools/cudaError/README.md
deleted file mode 100644
index df7434c33a9fd7c6dfcf8c3cd7479169d748ca48..0000000000000000000000000000000000000000
--- a/tools/cudaError/README.md
+++ /dev/null
@@ -1,22 +0,0 @@
-Usage:
-
-Please run:
-```
-bash start.sh
-```
-
-The error message of CUDA9.0 / CUDA10.0 / CUDA-latest-version will be crawled by default.
-
-If you want to crawl a specified version of CUDA, Please run:
-```
-bash start.sh <version> <URL(optional)>
-```
-URL can be derived by default, so you don't have to enter a URL.
-
-for example:
-```
-bash start.sh 11.0
-```
-will capture error message of CUDA11.0(in future).
-
-Every time when Nvidia upgrade the CUDA major version, you need to run `bash start.sh` in current directory, and upload cudaErrorMessage.tar.gz to https://paddlepaddledeps.bj.bcebos.com/cudaErrorMessage.tar.gz
diff --git a/tools/cudaError/spider.py b/tools/cudaError/spider.py
deleted file mode 100644
index c2c3dc97f422202e96d9c6ab58ce462e7dbd980e..0000000000000000000000000000000000000000
--- a/tools/cudaError/spider.py
+++ /dev/null
@@ -1,124 +0,0 @@
-# Copyright (c) 2020 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 ssl
-import re
-import urllib2
-import json
-import collections
-import sys, getopt
-import cuda_error_pb2
-
-
-def parsing(cuda_errorDesc, version, url):
-    All_Messages = cuda_errorDesc.AllMessages.add()
-    All_Messages.version = int(version)
-
-    ssl._create_default_https_context = ssl._create_unverified_context
-    html = urllib2.urlopen(url).read()
-    res_div = r'<div class="section">.*?<p>CUDA error types </p>.*?</div>.*?<div class="enum-members">(.*?)</div>'
-    m_div = re.findall(res_div, html, re.S | re.M)
-
-    url_list = url.split('/')
-    url_prefix = '/'.join(url_list[0:url_list.index('cuda-runtime-api') + 1])
-
-    dic = collections.OrderedDict()
-    dic_message = collections.OrderedDict()
-    for line in m_div:
-        res_dt = r'<dt>(.*?)</dt>.*?<dd>(.*?)</dd>'
-        m_dt = re.findall(res_dt, line, re.S | re.M)
-        for error in m_dt:
-            res_type = r'<span class="ph ph apiData">(.*?)</span>'
-            m_type = re.findall(res_type, error[0], re.S | re.M)[0]
-            m_message = error[1]
-            m_message = m_message.replace('\n', '')
-            res_a = r'(<a class=.*?</a>)'
-            res_shape = r'<a class=.*?>(.*?)</a>'
-            list_a = re.findall(res_a, m_message, re.S | re.M)
-            list_shape = re.findall(res_shape, m_message, re.S | re.M)
-            assert len(list_a) == len(list_shape)
-            for idx in range(len(list_a)):
-                m_message = m_message.replace(list_a[idx], list_shape[idx])
-
-            m_message = m_message.replace(
-                '<h6 class=\"deprecated_header\">Deprecated</h6>', '')
-
-            res_span = r'(<span class=.*?</span>)'
-            res_span_detail = r'<span class=.*?>(.*?)</span>'
-            list_span = re.findall(res_span, m_message, re.S | re.M)
-            list_span_detail = re.findall(res_span_detail, m_message, re.S |
-                                          re.M)
-            assert len(list_span) == len(list_span_detail)
-            for idx in range(len(list_span)):
-                m_message = m_message.replace(list_span[idx],
-                                              list_span_detail[idx])
-
-            res_p = r'(<p>.*?</p>)'
-            res_p_detail = r'<p>(.*?)</p>'
-            list_p = re.findall(res_p, m_message, re.S | re.M)
-            list_p_detail = re.findall(res_p_detail, m_message, re.S | re.M)
-            assert len(list_p) == len(list_p_detail)
-            for idx in range(len(list_p)):
-                m_message = m_message.replace(list_p[idx], list_p_detail[idx])
-
-            m_message = m_message.replace('  ', '')
-            _Messages = All_Messages.Messages.add()
-            try:
-                _Messages.errorCode = int(m_type)
-            except ValueError:
-                if re.match('0x', m_type):
-                    _Messages.errorCode = int(m_type, 16)
-                else:
-                    raise ValueError
-            _Messages.errorMessage = m_message  # save for cudaErrorMessage.pb from python-protobuf interface
-
-
-def main(argv):
-    version = []
-    url = []
-    try:
-        opts, args = getopt.getopt(argv, "hv:u:", ["help", "version=", "url="])
-    except getopt.GetoptError:
-        print 'python spider.py -v <version1,version2,...,> -u <url1,url2,...,>'
-        sys.exit(2)
-    for opt, arg in opts:
-        if opt in ("-h", "--help"):
-            print 'python spider.py -v <version1,version2,...,> -u <url1,url2,...,>'
-            sys.exit()
-        elif opt in ("-v", "--version"):
-            version = arg
-        elif opt in ("-u", "--url"):
-            url = arg
-    version = version.split(',')
-    url = url.split(',')
-    assert len(version) == len(url)
-    cuda_errorDesc = cuda_error_pb2.cudaerrorDesc()
-    for idx in range(len(version)):
-        if version[idx] == "-1":
-            print("crawling errorMessage for CUDA%s from %s" %
-                  ("-latest-version", url[idx]))
-        else:
-            print("crawling errorMessage for CUDA%s from %s" %
-                  (version[idx], url[idx]))
-        parsing(cuda_errorDesc, version[idx], url[idx])
-
-    serializeToString = cuda_errorDesc.SerializeToString()
-    with open("cudaErrorMessage.pb", "wb") as f:
-        f.write(serializeToString
-                )  # save for cudaErrorMessage.pb from python-protobuf interface
-    print("crawling errorMessage for CUDA has been done!!!")
-
-
-if __name__ == "__main__":
-    main(sys.argv[1:])
diff --git a/tools/externalError/README.md b/tools/externalError/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..029efd8cb949190135aeb12635e6951f5d979d7d
--- /dev/null
+++ b/tools/externalError/README.md
@@ -0,0 +1,9 @@
+Usage:
+
+Please run:
+```
+bash start.sh
+```
+
+If you want to update all external error message, you need to run command `bash start.sh` in current directory, 
+and upload the generated file `externalErrorMsg.tar.gz` to https://paddlepaddledeps.bj.bcebos.com/externalErrorMsg.tar.gz
diff --git a/tools/externalError/spider.py b/tools/externalError/spider.py
new file mode 100644
index 0000000000000000000000000000000000000000..a74d82f40ebebd3e59cd2b94715905e7157cbbef
--- /dev/null
+++ b/tools/externalError/spider.py
@@ -0,0 +1,363 @@
+# Copyright (c) 2021 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 ssl
+import re
+import urllib.request
+import json
+import collections
+import sys, getopt
+import external_error_pb2
+
+
+def parsing(externalErrorDesc):
+    #*********************************************************************************************#
+    #*********************************** CUDA Error Message **************************************#
+    print("start crawling errorMessage for nvidia CUDA API--->")
+    url = 'https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0038'
+
+    allMessageDesc = externalErrorDesc.errors.add()
+    allMessageDesc.type = external_error_pb2.CUDA
+
+    ssl._create_default_https_context = ssl._create_unverified_context
+    html = urllib.request.urlopen(url).read().decode('utf-8')
+    res_div = r'<div class="section">.*?<p>CUDA error types </p>.*?</div>.*?<div class="enum-members">(.*?)</div>'
+    m_div = re.findall(res_div, html, re.S | re.M)[0]
+
+    res_dt = r'<dt>(.*?)</dt>.*?<dd>(.*?)</dd>'
+    m_dt = re.findall(res_dt, m_div, re.S | re.M)
+    for error in m_dt:
+        res_type = r'<span class="enum-member-name-def">(.*?) = <span class="ph ph apiData">(.*?)</span></span>'
+        m_type = re.findall(res_type, error[0], re.S | re.M)[0]
+        m_message = error[1]
+        m_message = m_message.replace('\n', '')
+        res_a = r'(<a class=.*?</a>)'
+        res_shape = r'<a class=.*?>(.*?)</a>'
+        list_a = re.findall(res_a, m_message, re.S | re.M)
+        list_shape = re.findall(res_shape, m_message, re.S | re.M)
+        assert len(list_a) == len(list_shape)
+        for idx in range(len(list_a)):
+            m_message = m_message.replace(list_a[idx], list_shape[idx])
+
+        m_message = m_message.replace(
+            '<h6 class=\"deprecated_header\">Deprecated</h6>', '')
+
+        res_span = r'(<span class=.*?</span>)'
+        res_span_detail = r'<span class=.*?>(.*?)</span>'
+        list_span = re.findall(res_span, m_message, re.S | re.M)
+        list_span_detail = re.findall(res_span_detail, m_message, re.S | re.M)
+        assert len(list_span) == len(list_span_detail)
+        for idx in range(len(list_span)):
+            m_message = m_message.replace(list_span[idx], list_span_detail[idx])
+
+        res_p = r'(<p>.*?</p>)'
+        res_p_detail = r'<p>(.*?)</p>'
+        list_p = re.findall(res_p, m_message, re.S | re.M)
+        list_p_detail = re.findall(res_p_detail, m_message, re.S | re.M)
+        assert len(list_p) == len(list_p_detail)
+        for idx in range(len(list_p)):
+            m_message = m_message.replace(list_p[idx], list_p_detail[idx])
+
+        m_message = m_message.replace('  ', '')
+        _Messages = allMessageDesc.messages.add()
+        try:
+            _Messages.code = int(m_type[1])
+        except ValueError:
+            if re.match('0x', m_type[1]):
+                _Messages.code = int(m_type[1], 16)
+            else:
+                raise ValueError
+        _Messages.message = "'%s'. %s" % (m_type[0], m_message)
+    print("End crawling errorMessage for nvidia CUDA API!\n")
+
+    #***********************************************************************************************#
+    #*********************************** CURAND Error Message **************************************#
+    print("start crawling errorMessage for nvidia CURAND API--->")
+    url = 'https://docs.nvidia.com/cuda/curand/group__HOST.html#group__HOST_1gb94a31d5c165858c96b6c18b70644437'
+
+    allMessageDesc = externalErrorDesc.errors.add()
+    allMessageDesc.type = external_error_pb2.CURAND
+
+    html = urllib.request.urlopen(url).read().decode('utf-8')
+
+    res_div = r'<div class="section">.*?<p>CURAND function call status types </p>.*?</div>.*?<div class="enum-members">(.*?)</div>'
+    m_div = re.findall(res_div, html, re.S | re.M)[0]
+
+    res_dt = r'<dt>(.*?)</dt>.*?<dd>(.*?)</dd>'
+    m_dt = re.findall(res_dt, m_div, re.S | re.M)
+    for error in m_dt:
+        res_type = r'<span class="enum-member-name-def">(.*?) = <span class="ph ph apiData">(.*?)</span></span>'
+        m_type = re.findall(res_type, error[0], re.S | re.M)[0]
+        m_message = error[1]
+
+        _Messages = allMessageDesc.messages.add()
+        try:
+            _Messages.code = int(m_type[1])
+        except ValueError:
+            if re.match('0x', m_type[1]):
+                _Messages.code = int(m_type[1], 16)
+            else:
+                raise ValueError
+        _Messages.message = "'%s'. %s" % (m_type[0], m_message)
+    print("End crawling errorMessage for nvidia CURAND API!\n")
+
+    #**************************************************************************************************#
+    #*********************************** CUDNN Error Message ******************************************#
+    cudnnStatus_t = {
+        "CUDNN_STATUS_SUCCESS": 0,
+        "CUDNN_STATUS_NOT_INITIALIZED": 1,
+        "CUDNN_STATUS_ALLOC_FAILED": 2,
+        "CUDNN_STATUS_BAD_PARAM": 3,
+        "CUDNN_STATUS_INTERNAL_ERROR": 4,
+        "CUDNN_STATUS_INVALID_VALUE": 5,
+        "CUDNN_STATUS_ARCH_MISMATCH": 6,
+        "CUDNN_STATUS_MAPPING_ERROR": 7,
+        "CUDNN_STATUS_EXECUTION_FAILED": 8,
+        "CUDNN_STATUS_NOT_SUPPORTED": 9,
+        "CUDNN_STATUS_LICENSE_ERROR": 10,
+        "CUDNN_STATUS_RUNTIME_PREREQUISITE_MISSING": 11,
+        "CUDNN_STATUS_RUNTIME_IN_PROGRESS": 12,
+        "CUDNN_STATUS_RUNTIME_FP_OVERFLOW": 13,
+    }
+
+    print("start crawling errorMessage for nvidia CUDNN API--->")
+    url = 'https://docs.nvidia.com/deeplearning/cudnn/api/index.html#cudnnStatus_t'
+
+    allMessageDesc = externalErrorDesc.errors.add()
+    allMessageDesc.type = external_error_pb2.CUDNN
+
+    html = urllib.request.urlopen(url).read().decode('utf-8')
+    f = open('1.txt', 'w')
+    f.write(html)
+
+    res_div = r'<div class="section" id="cudnnStatus_t__section_lmp_dgr_2jb"><a name="cudnnStatus_t__section_lmp_dgr_2jb" shape="rect">(.*?)</div>'
+    m_div = re.findall(res_div, html, re.S | re.M)[0]
+
+    res_dt = r'<dt class="dt dlterm"><samp class="ph codeph">(.*?)</samp></dt>.*?<dd class="dd">(.*?)</dd>'
+    m_dt = re.findall(res_dt, m_div, re.S | re.M)
+    for error in m_dt:
+        m_message = error[1]
+
+        res_class = r'<p class="p">.*?</p>'
+        res_class_detail = r'<p class="p">(.*?)</p>'
+        list_class = re.findall(res_class, m_message, re.S | re.M)
+        list_class_detail = re.findall(res_class_detail, m_message, re.S | re.M)
+        assert len(list_class) == len(list_class_detail)
+        for idx in range(len(list_class)):
+            m_message = m_message.replace(list_class[idx],
+                                          list_class_detail[idx])
+
+        res_a = r'(<a class="xref".*?</a>)'
+        res_shape = r'<a class="xref".*?>(.*?)</a>'
+        list_a = re.findall(res_a, m_message, re.S | re.M)
+        list_shape = re.findall(res_shape, m_message, re.S | re.M)
+        assert len(list_a) == len(list_shape)
+        for idx in range(len(list_a)):
+            m_message = m_message.replace(list_a[idx], list_shape[idx])
+
+        res_span = r'(<span class="ph">.*?</span>)'
+        res_span_detail = r'<span class="ph">(.*?)</span>'
+        list_span = re.findall(res_span, m_message, re.S | re.M)
+        list_span_detail = re.findall(res_span_detail, m_message, re.S | re.M)
+        assert len(list_span) == len(list_span_detail)
+        for idx in range(len(list_span)):
+            m_message = m_message.replace(list_span[idx], list_span_detail[idx])
+
+        res_samp = r'(<samp class="ph codeph">.*?</samp>)'
+        res_samp_detail = r'<samp class="ph codeph">(.*?)</samp>'
+        list_samp = re.findall(res_samp, m_message, re.S | re.M)
+        list_samp_detail = re.findall(res_samp_detail, m_message, re.S | re.M)
+        assert len(list_samp) == len(list_samp_detail)
+        for idx in range(len(list_samp)):
+            m_message = m_message.replace(list_samp[idx], list_samp_detail[idx])
+
+        m_message = re.sub(r'\n +', ' ', m_message)
+
+        _Messages = allMessageDesc.messages.add()
+        _Messages.code = int(cudnnStatus_t[error[0]])
+        _Messages.message = "'%s'. %s" % (error[0], m_message)
+    print("End crawling errorMessage for nvidia CUDNN API!\n")
+
+    #*************************************************************************************************#
+    #*********************************** CUBLAS Error Message ****************************************#
+    cublasStatus_t = {
+        "CUBLAS_STATUS_SUCCESS": 0,
+        "CUBLAS_STATUS_NOT_INITIALIZED": 1,
+        "CUBLAS_STATUS_ALLOC_FAILED": 3,
+        "CUBLAS_STATUS_INVALID_VALUE": 7,
+        "CUBLAS_STATUS_ARCH_MISMATCH": 8,
+        "CUBLAS_STATUS_MAPPING_ERROR": 11,
+        "CUBLAS_STATUS_EXECUTION_FAILED": 13,
+        "CUBLAS_STATUS_INTERNAL_ERROR": 14,
+        "CUBLAS_STATUS_NOT_SUPPORTED": 15,
+        "CUBLAS_STATUS_LICENSE_ERROR": 16
+    }
+
+    print("start crawling errorMessage for nvidia CUBLAS API--->")
+    url = 'https://docs.nvidia.com/cuda/cublas/index.html#cublasstatus_t'
+
+    allMessageDesc = externalErrorDesc.errors.add()
+    allMessageDesc.type = external_error_pb2.CUBLAS
+
+    html = urllib.request.urlopen(url).read().decode('utf-8')
+
+    res_div = r'<p class="p">The type is used for function status returns. All cuBLAS library.*?<div class="tablenoborder">(.*?)</div>'
+    m_div = re.findall(res_div, html, re.S | re.M)[0]
+
+    res_dt = r'<p class="p"><samp class="ph codeph">(.*?)</samp></p>.*?colspan="1">(.*?)</td>'
+    m_dt = re.findall(res_dt, m_div, re.S | re.M)
+
+    for error in m_dt:
+        m_message = error[1]
+        m_message = re.sub(r'\n +', ' ', m_message)
+
+        res_p = r'<p class="p">.*?</p>'
+        res_p_detail = r'<p class="p">(.*?)</p>'
+        list_p = re.findall(res_p, m_message, re.S | re.M)
+        list_p_detail = re.findall(res_p_detail, m_message, re.S | re.M)
+        assert len(list_p) == len(list_p_detail)
+        for idx in range(len(list_p)):
+            m_message = m_message.replace(list_p[idx], list_p_detail[idx])
+
+        res_samp = r'<samp class="ph codeph">.*?</samp>'
+        res_samp_detail = r'<samp class="ph codeph">(.*?)</samp>'
+        list_samp = re.findall(res_samp, m_message, re.S | re.M)
+        list_samp_detail = re.findall(res_samp_detail, m_message, re.S | re.M)
+        assert len(list_samp) == len(list_samp_detail)
+        for idx in range(len(list_samp)):
+            m_message = m_message.replace(list_samp[idx], list_samp_detail[idx])
+
+        _Messages = allMessageDesc.messages.add()
+        _Messages.code = int(cublasStatus_t[error[0]])
+        _Messages.message = "'%s'. %s" % (error[0], m_message)
+    print("End crawling errorMessage for nvidia CUBLAS API!\n")
+
+    #*************************************************************************************************#
+    #*********************************** CUSOLVER Error Message **************************************#
+    cusolverStatus_t = {
+        "CUSOLVER_STATUS_SUCCESS": 0,
+        "CUSOLVER_STATUS_NOT_INITIALIZED": 1,
+        "CUSOLVER_STATUS_ALLOC_FAILED": 2,
+        "CUSOLVER_STATUS_INVALID_VALUE": 3,
+        "CUSOLVER_STATUS_ARCH_MISMATCH": 4,
+        "CUSOLVER_STATUS_MAPPING_ERROR": 5,
+        "CUSOLVER_STATUS_EXECUTION_FAILED": 6,
+        "CUSOLVER_STATUS_INTERNAL_ERROR": 7,
+        "CUSOLVER_STATUS_MATRIX_TYPE_NOT_SUPPORTED": 8,
+        "CUSOLVER_STATUS_NOT_SUPPORTED": 9,
+        "CUSOLVER_STATUS_ZERO_PIVOT": 10,
+        "CUSOLVER_STATUS_INVALID_LICENSE": 11,
+        "CUSOLVER_STATUS_IRS_PARAMS_NOT_INITIALIZED": 12,
+        "CUSOLVER_STATUS_IRS_PARAMS_INVALID": 13,
+        "CUSOLVER_STATUS_IRS_INTERNAL_ERROR": 14,
+        "CUSOLVER_STATUS_IRS_NOT_SUPPORTED": 15,
+        "CUSOLVER_STATUS_IRS_OUT_OF_RANGE": 16,
+        "CUSOLVER_STATUS_IRS_NRHS_NOT_SUPPORTED_FOR_REFINE_GMRES": 17,
+        "CUSOLVER_STATUS_IRS_INFOS_NOT_INITIALIZED": 18
+    }
+    print("start crawling errorMessage for nvidia CUSOLVER API--->")
+    url = 'https://docs.nvidia.com/cuda/cusolver/index.html#cuSolverSPstatus'
+
+    allMessageDesc = externalErrorDesc.errors.add()
+    allMessageDesc.type = external_error_pb2.CUSOLVER
+
+    html = urllib.request.urlopen(url).read().decode('utf-8')
+
+    res_div = r'This is a status type returned by the library functions and.*?<div class="tablenoborder">(.*?)</div>'
+    m_div = re.findall(res_div, html, re.S | re.M)[0]
+
+    res_dt = r'<samp class="ph codeph">(.*?)</samp></td>.*?colspan="1">(.*?)</td>'
+    m_dt = re.findall(res_dt, m_div, re.S | re.M)
+
+    for error in m_dt:
+        m_message = error[1]
+        m_message = re.sub(r'\n +', '', m_message)
+        m_message = re.sub(r'<p class="p"></p>', '', m_message)
+
+        res_p = r'<p class="p">.*?</p>'
+        res_p_detail = r'<p class="p">(.*?)</p>'
+        list_p = re.findall(res_p, m_message, re.S | re.M)
+        list_p_detail = re.findall(res_p_detail, m_message, re.S | re.M)
+        assert len(list_p) == len(list_p_detail)
+        for idx in range(len(list_p)):
+            m_message = m_message.replace(list_p[idx], list_p_detail[idx])
+
+        res_samp = r'<samp class="ph codeph">.*?</samp>'
+        res_samp_detail = r'<samp class="ph codeph">(.*?)</samp>'
+        list_samp = re.findall(res_samp, m_message, re.S | re.M)
+        list_samp_detail = re.findall(res_samp_detail, m_message, re.S | re.M)
+        assert len(list_samp) == len(list_samp_detail)
+        for idx in range(len(list_samp)):
+            m_message = m_message.replace(list_samp[idx], list_samp_detail[idx])
+
+        res_strong = r'<strong class="ph b">.*?</strong>'
+        res_strong_detail = r'<strong class="ph b">(.*?)</strong>'
+        list_strong = re.findall(res_strong, m_message, re.S | re.M)
+        list_strong_detail = re.findall(res_strong_detail, m_message, re.S |
+                                        re.M)
+        assert len(list_strong) == len(list_strong_detail)
+        for idx in range(len(list_strong)):
+            m_message = m_message.replace(list_strong[idx],
+                                          list_strong_detail[idx])
+
+        _Messages = allMessageDesc.messages.add()
+        _Messages.code = int(cusolverStatus_t[error[0]])
+        _Messages.message = "'%s'. %s" % (error[0], m_message)
+    print("End crawling errorMessage for nvidia CUSOLVER API!\n")
+
+    #**********************************************************************************************#
+    #*************************************** NCCL error *******************************************#
+    print("start crawling errorMessage for nvidia NCCL API--->")
+    url = 'https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/api/types.html#ncclresult-t'
+    allMessageDesc = externalErrorDesc.errors.add()
+    allMessageDesc.type = external_error_pb2.NCCL
+    html = urllib.request.urlopen(url).read().decode('utf-8')
+    res_div = r'<code class="descname">ncclResult_t</code>(.*?)</div>'
+    m_div = re.findall(res_div, html, re.S | re.M)[0]
+
+    res_dt = r'<code class="descname">(.*?)</code>.*?<span class="pre">(.*?)</span></code>\)(.*?)</p>\n</dd></dl>'
+    m_dt = re.findall(res_dt, m_div, re.S | re.M)
+    for error in m_dt:
+        m_message = re.sub(r'\n', '', error[2])
+        _Messages = allMessageDesc.messages.add()
+        _Messages.code = int(error[1])
+        _Messages.message = "'%s'. %s" % (error[0], m_message)
+    print("End crawling errorMessage for nvidia NCCL API!\n")
+
+
+def main(argv):
+    try:
+        opts, _ = getopt.getopt(argv, "h", ["help"])
+    except getopt.GetoptError:
+        print('python spider.py')
+        sys.exit(2)
+    for opt, _ in opts:
+        if opt in ("-h", "--help"):
+            print('python spider.py')
+            sys.exit(2)
+    externalErrorDesc = external_error_pb2.ExternalErrorDesc()
+    parsing(externalErrorDesc)
+
+    serializedString = externalErrorDesc.SerializeToString()
+    with open("externalErrorMsg.pb", "wb") as f:
+        # save for externalErrorMsg.pb from Python-protobuf interface
+        # load from C++-protobuf interface and get error message
+        f.write(serializedString)
+    print(
+        "Generating data file [externalErrorMsg.pb] for external third_party API error has been done!"
+    )
+
+
+if __name__ == "__main__":
+    main(sys.argv[1:])
diff --git a/tools/cudaError/start.sh b/tools/externalError/start.sh
similarity index 59%
rename from tools/cudaError/start.sh
rename to tools/externalError/start.sh
index 66e56b8485d8c6d40937bf821c1889424da33527..32ef63c261268191646b03648ca46fbc15485087 100644
--- a/tools/cudaError/start.sh
+++ b/tools/externalError/start.sh
@@ -29,19 +29,7 @@ else
     echo "please run on Mac/Linux"
     exit 1
 fi
-protobuf/bin/protoc -I../../paddle/fluid/platform/ --python_out . ../../paddle/fluid/platform/cuda_error.proto
+protobuf/bin/protoc -I../../paddle/fluid/platform/ --python_out . ../../paddle/fluid/platform/external_error.proto
 
-version=90,100,-1    # -1 represent the latest cuda-version 
-url=https://docs.nvidia.com/cuda/archive/9.0/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0038,https://docs.nvidia.com/cuda/archive/10.0/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0038,https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0038
-
-if [ "$1" != "" ]; then
-    version=$version,$(($1*10))
-    if [ "$2" != "" ]; then
-        url=$url,$2
-    else
-        url=$url,https://docs.nvidia.com/cuda/archive/$1/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0038
-    fi
-fi
-
-python spider.py --version=$version --url=$url
-tar czf cudaErrorMessage.tar.gz cudaErrorMessage.pb
+python3.7 spider.py
+tar czvf externalErrorMsg.tar.gz externalErrorMsg.pb