Merge branch 'develop' of https://github.com/PaddlePaddle/Serving into fix_dead_link

9cf29f60 · HexToString · 2422a8d6 · 1b48cf5a · 9cf29f60 · 9cf29f60
17 changed file
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -59,6 +59,8 @@ option(PACK                 "Compile for whl"
 option(WITH_TRT             "Compile Paddle Serving with TRT"                   OFF)
 option(PADDLE_ON_INFERENCE  "Compile for encryption"                             ON)
 option(WITH_OPENCV	    "Compile Paddle Serving with OPENCV"                    OFF)
+option(WITH_ROCM	    "Compile Paddle Serving with ROCM"                    OFF)
+option(WITH_ASCEND_CL       "Compile PaddlePaddle with ASCEND CL"               OFF)

 if(NOT DEFINED VERSION_TAG)
  set(VERSION_TAG "0.0.0")
@@ -163,6 +165,14 @@ if(SERVER)
    endif()

    list(APPEND EXTERNAL_LIBS paddlepaddle)
+    if(WITH_ROCM)
+        include(hip)
+    endif()
+
+    if(WITH_ASCEND_CL)
+	    include(external/ascend)
+	    list(APPEND EXTERNAL_LIBS ascend ascend_cl)
+    endif()
 endif()



--- a/cmake/external/ascend.cmake
+++ b/cmake/external/ascend.cmake
+# 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.
+
+
+#NOTE: Logic is from
+# https://github.com/mindspore-ai/graphengine/blob/master/CMakeLists.txt
+if(DEFINED ENV{ASCEND_CUSTOM_PATH})
+    set(ASCEND_DIR $ENV{ASCEND_CUSTOM_PATH})
+else()
+    set(ASCEND_DIR /usr/local/Ascend)
+endif()
+
+if(EXISTS ${ASCEND_DIR}/ascend-toolkit/latest/fwkacllib/include/graph/ascend_string.h)
+  # It means CANN 20.2 +
+  add_definitions(-DPADDLE_WITH_ASCEND_STRING)
+endif()
+
+
+if(WITH_ASCEND OR WITH_ASCEND_CL)
+  set(ASCEND_DRIVER_DIR ${ASCEND_DIR}/driver/lib64)
+  set(ASCEND_DRIVER_COMMON_DIR ${ASCEND_DIR}/driver/lib64/common)
+  set(ASCEND_DRIVER_SHARE_DIR ${ASCEND_DIR}/driver/lib64/share)
+  set(ASCEND_RUNTIME_DIR ${ASCEND_DIR}/fwkacllib/lib64)
+  set(ASCEND_ATC_DIR ${ASCEND_DIR}/atc/lib64)
+  set(ASCEND_ACL_DIR ${ASCEND_DIR}/acllib/lib64)
+  set(STATIC_ACL_LIB ${ASCEND_ACL_DIR})
+
+  set(ASCEND_MS_RUNTIME_PATH ${ASCEND_RUNTIME_DIR} ${ASCEND_ACL_DIR} ${ASCEND_ATC_DIR})
+  set(ASCEND_MS_DRIVER_PATH ${ASCEND_DRIVER_DIR} ${ASCEND_DRIVER_COMMON_DIR})
+  set(ATLAS_RUNTIME_DIR ${ASCEND_DIR}/ascend-toolkit/latest/fwkacllib/lib64)
+  set(ATLAS_RUNTIME_INC_DIR ${ASCEND_DIR}/ascend-toolkit/latest/fwkacllib/include)
+  set(ATLAS_ACL_DIR ${ASCEND_DIR}/ascend-toolkit/latest/acllib/lib64)
+  set(ATLAS_ATC_DIR ${ASCEND_DIR}/ascend-toolkit/latest/atc/lib64)
+  set(ATLAS_MS_RUNTIME_PATH ${ATLAS_RUNTIME_DIR} ${ATLAS_ACL_DIR} ${ATLAS_ATC_DIR})
+
+  set(atlas_graph_lib ${ATLAS_RUNTIME_DIR}/libgraph.so)
+  set(atlas_ge_runner_lib ${ATLAS_RUNTIME_DIR}/libge_runner.so)
+  set(atlas_acl_lib ${ATLAS_RUNTIME_DIR}/libascendcl.so)
+  INCLUDE_DIRECTORIES(${ATLAS_RUNTIME_INC_DIR})
+
+
+  ADD_LIBRARY(ascend_ge SHARED IMPORTED GLOBAL)
+  SET_PROPERTY(TARGET ascend_ge PROPERTY IMPORTED_LOCATION ${atlas_ge_runner_lib})
+
+  ADD_LIBRARY(ascend_graph SHARED IMPORTED GLOBAL)
+  SET_PROPERTY(TARGET ascend_graph PROPERTY IMPORTED_LOCATION ${atlas_graph_lib})
+
+  ADD_LIBRARY(atlas_acl SHARED IMPORTED GLOBAL)
+  SET_PROPERTY(TARGET atlas_acl PROPERTY IMPORTED_LOCATION ${atlas_acl_lib})
+
+  add_custom_target(extern_ascend DEPENDS ascend_ge ascend_graph atlas_acl)
+endif()
+
+if(WITH_ASCEND_CL)
+  set(ASCEND_CL_DIR ${ASCEND_DIR}/ascend-toolkit/latest/fwkacllib/lib64)
+
+  set(ascend_hccl_lib ${ASCEND_CL_DIR}/libhccl.so)
+  set(ascendcl_lib ${ASCEND_CL_DIR}/libascendcl.so)
+  set(acl_op_compiler_lib ${ASCEND_CL_DIR}/libacl_op_compiler.so)
+  set(FWKACLLIB_INC_DIR ${ASCEND_DIR}/ascend-toolkit/latest/fwkacllib/include)
+  set(ACLLIB_INC_DIR ${ASCEND_DIR}/ascend-toolkit/latest/acllib/include)
+
+  message(STATUS "FWKACLLIB_INC_DIR ${FWKACLLIB_INC_DIR}")
+  message(STATUS "ASCEND_CL_DIR ${ASCEND_CL_DIR}")
+  INCLUDE_DIRECTORIES(${FWKACLLIB_INC_DIR})
+  INCLUDE_DIRECTORIES(${ACLLIB_INC_DIR})
+
+  ADD_LIBRARY(ascendcl SHARED IMPORTED GLOBAL)
+  SET_PROPERTY(TARGET ascendcl PROPERTY IMPORTED_LOCATION ${ascendcl_lib})
+
+  ADD_LIBRARY(ascend_hccl SHARED IMPORTED GLOBAL)
+  SET_PROPERTY(TARGET ascend_hccl PROPERTY IMPORTED_LOCATION ${ascend_hccl_lib})
+
+  ADD_LIBRARY(acl_op_compiler SHARED IMPORTED GLOBAL)
+  SET_PROPERTY(TARGET acl_op_compiler PROPERTY IMPORTED_LOCATION ${acl_op_compiler_lib})
+  add_custom_target(extern_ascend_cl DEPENDS ascendcl acl_op_compiler)
+
+endif()
--- a/cmake/hip.cmake
+++ b/cmake/hip.cmake
+if(NOT WITH_ROCM)
+    return()
+endif()
+
+if(NOT DEFINED ENV{ROCM_PATH})
+    set(ROCM_PATH "/opt/rocm" CACHE PATH "Path to which ROCm has been installed")
+    set(HIP_PATH ${ROCM_PATH}/hip CACHE PATH "Path to which HIP has been installed")
+    set(HIP_CLANG_PATH ${ROCM_PATH}/llvm/bin CACHE PATH "Path to which clang has been installed")
+else()
+    set(ROCM_PATH $ENV{ROCM_PATH} CACHE PATH "Path to which ROCm has been installed")
+    set(HIP_PATH ${ROCM_PATH}/hip CACHE PATH "Path to which HIP has been installed")
+    set(HIP_CLANG_PATH ${ROCM_PATH}/llvm/bin CACHE PATH "Path to which clang has been installed")
+endif()
+set(CMAKE_MODULE_PATH "${HIP_PATH}/cmake" ${CMAKE_MODULE_PATH})
+
+find_package(HIP REQUIRED)
+include_directories(${ROCM_PATH}/include)
+message(STATUS "HIP version: ${HIP_VERSION}")
+message(STATUS "HIP_CLANG_PATH: ${HIP_CLANG_PATH}")
+
+macro(find_package_and_include PACKAGE_NAME)
+  find_package("${PACKAGE_NAME}" REQUIRED)
+  include_directories("${ROCM_PATH}/${PACKAGE_NAME}/include")
+  message(STATUS "${PACKAGE_NAME} version: ${${PACKAGE_NAME}_VERSION}")
+endmacro()
+
+find_package_and_include(miopen)
+find_package_and_include(rocblas)
+find_package_and_include(hiprand)
+find_package_and_include(rocrand)
+find_package_and_include(rccl)
+find_package_and_include(rocthrust)
+find_package_and_include(hipcub)
+find_package_and_include(rocprim)
+find_package_and_include(hipsparse)
+find_package_and_include(rocsparse)
+find_package_and_include(rocfft)
+
+# set CXX flags for HIP
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__HIP_PLATFORM_HCC__")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_HIP")
+set(THRUST_DEVICE_SYSTEM THRUST_DEVICE_SYSTEM_HIP)
+
+# define HIP_CXX_FLAGS
+list(APPEND HIP_CXX_FLAGS -fPIC)
+list(APPEND HIP_CXX_FLAGS -D__HIP_PLATFORM_HCC__=1)
+# Note(qili93): HIP has compile conflicts of float16.h as platform::float16 overload std::is_floating_point and std::is_integer
+list(APPEND HIP_CXX_FLAGS -D__HIP_NO_HALF_CONVERSIONS__=1)
+list(APPEND HIP_CXX_FLAGS -Wno-macro-redefined)
+list(APPEND HIP_CXX_FLAGS -Wno-inconsistent-missing-override)
+list(APPEND HIP_CXX_FLAGS -Wno-exceptions)
+list(APPEND HIP_CXX_FLAGS -Wno-shift-count-negative)
+list(APPEND HIP_CXX_FLAGS -Wno-shift-count-overflow)
+list(APPEND HIP_CXX_FLAGS -Wno-unused-command-line-argument)
+list(APPEND HIP_CXX_FLAGS -Wno-duplicate-decl-specifier)
+list(APPEND HIP_CXX_FLAGS -Wno-implicit-int-float-conversion)
+list(APPEND HIP_CXX_FLAGS -Wno-pass-failed)
+list(APPEND HIP_CXX_FLAGS -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_HIP)
+list(APPEND HIP_CXX_FLAGS -std=c++14)
+
+if(CMAKE_BUILD_TYPE MATCHES Debug)
+  list(APPEND HIP_CXX_FLAGS -g2)
+  list(APPEND HIP_CXX_FLAGS -O0)
+  list(APPEND HIP_HIPCC_FLAGS -fdebug-info-for-profiling)
+endif(CMAKE_BUILD_TYPE MATCHES Debug)
+
+set(HIP_HCC_FLAGS ${HIP_CXX_FLAGS})
+set(HIP_CLANG_FLAGS ${HIP_CXX_FLAGS})
+# Ask hcc to generate device code during compilation so we can use
+# host linker to link.
+list(APPEND HIP_HCC_FLAGS -fno-gpu-rdc)
+list(APPEND HIP_HCC_FLAGS --amdgpu-target=gfx906)
+list(APPEND HIP_CLANG_FLAGS -fno-gpu-rdc)
+list(APPEND HIP_CLANG_FLAGS --amdgpu-target=gfx906)
+
+
+if(HIP_COMPILER STREQUAL clang)
+  set(hip_library_name amdhip64)
+else()
+  set(hip_library_name hip_hcc)
+endif()
+message(STATUS "HIP library name: ${hip_library_name}")
+
+# set HIP link libs
+find_library(ROCM_HIPRTC_LIB ${hip_library_name} HINTS ${HIP_PATH}/lib)
+message(STATUS "ROCM_HIPRTC_LIB: ${ROCM_HIPRTC_LIB}")
+
+#include(thrust)
--- a/cmake/paddlepaddle.cmake
+++ b/cmake/paddlepaddle.cmake
@@ -62,10 +62,27 @@ elseif (WITH_LITE)
        elseif(CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
            SET(PADDLE_LIB_VERSION "arm64_gcc7.3_openblas")
        endif()
+    elseif (WITH_ASCEND_CL)
+        if (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+             MESSAGE("paddle lite lib is unknown.")
+             SET(PADDLE_LIB_VERSION "paddle-lite-unknown")
+         elseif(CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+             SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}/cxx_c/Linux/ASCEND/arm64_gcc7.5_openblas_lite2.10")
+         endif()
    else()
        MESSAGE("paddle lite lib is unknown.")
        SET(PADDLE_LIB_VERSION "paddle-lite-unknown")
    endif()
+elseif (WITH_ROCM)
+    SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}/rocm")
+elseif (WITH_ASCEND_CL)
+    message("cpu arch: ${CMAKE_SYSTEM_PROCESSOR}")
+    if (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+        MESSAGE("paddle lib is unknown.")
+        SET(PADDLE_LIB_VERSION "paddle-unknown")
+    elseif(CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+        SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}/cxx_c/Linux/ASCEND/arm64_gcc8.2_openblas")
+    endif()
 else()
    if (WITH_AVX)
        if (WITH_MKLML)
@@ -79,7 +96,13 @@ else()
 endif()

 if(WITH_LITE)
+    if (WITH_XPU)
        SET(PADDLE_LIB_PATH "https://paddle-inference-lib.bj.bcebos.com/2.2.0-rc0/cxx_c/Linux/XPU/${PADDLE_LIB_VERSION}/paddle_inference_install_dir.tar.gz ")
+    elseif (WITH_ASCEND_CL)
+        SET(PADDLE_LIB_PATH "http://paddle-serving.bj.bcebos.com/inferlib/${PADDLE_LIB_VERSION}/paddle_inference_install_dir.tgz ")
+    endif()
+elseif(WITH_ASCEND_CL)
+    SET(PADDLE_LIB_PATH "http://paddle-serving.bj.bcebos.com/inferlib/${PADDLE_LIB_VERSION}/paddle_inference.tgz ")
 else()
    SET(PADDLE_LIB_PATH "http://paddle-inference-lib.bj.bcebos.com/${PADDLE_LIB_VERSION}/paddle_inference.tgz")
 endif()
@@ -150,6 +173,9 @@ endif()

 ADD_LIBRARY(paddle_inference STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET paddle_inference PROPERTY IMPORTED_LOCATION ${PADDLE_INSTALL_DIR}/lib/libpaddle_inference.a)
+if (WITH_ASCEND_CL)
+    SET_PROPERTY(TARGET paddle_inference PROPERTY IMPORTED_LOCATION ${PADDLE_INSTALL_DIR}/lib/libpaddle_inference.so)
+endif()

 if (WITH_TRT)
    ADD_LIBRARY(nvinfer SHARED IMPORTED GLOBAL)

--- a/core/configure/proto/server_configure.proto
+++ b/core/configure/proto/server_configure.proto
@@ -47,6 +47,7 @@ message EngineDesc {
  optional bool combined_model = 18;
  optional bool encrypted_model = 19;
  optional bool gpu_multi_stream = 20;
+  optional bool use_ascend_cl = 21;

  /*
   * "runtime_thread_num": n == 0 means don`t use Asynchronous task scheduling

--- a/core/general-client/CMakeLists.txt
+++ b/core/general-client/CMakeLists.txt
@@ -18,7 +18,7 @@ add_executable(simple_client example/simple_client.cpp)
 add_dependencies(simple_client utils sdk-cpp client)

 target_link_libraries(simple_client -Wl,--whole-archive
-        -Wl,--no-whole-archive -lpthread -lcrypto -lm -lrt -lssl -ldl -lz -Wl,-rpath,'$ORIGIN'/lib)
+        -Wl,--no-whole-archive -Wl,--start-group -lpthread -lcrypto -lm -lrt -lssl -ldl -lz -Wl,-rpath,'$ORIGIN'/lib)

 target_link_libraries(simple_client utils)
 target_link_libraries(simple_client sdk-cpp)

--- a/core/general-server/CMakeLists.txt
+++ b/core/general-server/CMakeLists.txt
@@ -31,6 +31,10 @@ target_link_libraries(serving pdserving)
 target_link_libraries(serving cube-api)
 target_link_libraries(serving utils)
 target_link_libraries(serving utf8proc)
+if(WITH_ASCEND_CL)
+    target_link_libraries(serving ascendcl acl_op_compiler)
+endif()
+

 if(WITH_GPU)
    target_link_libraries(serving ${CUDA_LIBRARIES})
@@ -63,3 +67,7 @@ install(FILES
        DESTINATION
        ${PADDLE_SERVING_INSTALL_DIR}/demo/serving/bin)
 endif()
+
+if (WITH_ROCM)
+    target_link_libraries(serving ${ROCM_HIPRTC_LIB})
+endif()
--- a/doc/Serving_Configure_CN.md
+++ b/doc/Serving_Configure_CN.md
@@ -84,8 +84,8 @@ workdir_9393
 | Argument                                       | Type | Default | Description                                           |
 | ---------------------------------------------- | ---- | ------- | ----------------------------------------------------- |
 | `thread`                                       | int  | `2`     | Number of brpc service thread                         |
-| `op_num`                                       | int[]| `0`     | Thread Number for each model in asynchronous mode     |
-| `op_max_batch`                                 | int[]| `32`    | Batch Number for each model in asynchronous mode      |
+| `runtime_thread_num`                           | int[]| `0`     | Thread Number for each model in asynchronous mode     |
+| `batch_infer_size`                             | int[]| `32`    | Batch Number for each model in asynchronous mode      |
 | `gpu_ids`                                      | str[]| `"-1"`  | Gpu card id for each model                            |
 | `port`                                         | int  | `9292`  | Exposed port of current service to users              |
 | `model`                                        | str[]| `""`    | Path of paddle model directory to be served           |
@@ -98,6 +98,7 @@ workdir_9393
 | `precision`                                    | str  | FP32    | Precision Mode, support FP32, FP16, INT8              |
 | `use_calib`                                    | bool | False   | Use TRT int8 calibration                              |
 | `gpu_multi_stream`                             | bool | False   | EnableGpuMultiStream to get larger QPS                |
+| `use_ascend_cl`                                | bool | False   | Enable for ascend910; Use with use_lite for ascend310 |

 #### 当您的某个模型想使用多张GPU卡部署时.
 ```BASH
@@ -249,6 +250,7 @@ engines {
  use_gpu: false
  combined_model: false
  gpu_multi_stream: false
+  use_ascend_cl: false
  runtime_thread_num: 0
  batch_infer_size: 32
  enable_overrun: false
@@ -286,6 +288,7 @@ gpu_ids: 2
 - use_gpu:是否使用GPU
 - combined_model: 是否使用组合模型文件
 - gpu_multi_stream: 是否开启gpu多流模式
+- use_ascend_cl: 是否使用昇腾,单独开启适配昇腾910，同时开启lite适配310
 - runtime_thread_num: 若大于0， 则启用Async异步模式，并创建对应数量的predictor实例。
 - batch_infer_size: Async异步模式下的最大batch数
 - enable_overrun: Async异步模式下总是将整个任务放入任务队列
@@ -357,7 +360,7 @@ op:
            #Fetch结果列表，以client_config中fetch_var的alias_name为准
            fetch_list: ["concat_1.tmp_0"]

-            # device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu
+            # device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu, 5=arm ascend310, 6=arm ascend910
            device_type: 0

            #计算硬件ID，当devices为""或不写时为CPU预测；当devices为"0", "0,1,2"时为GPU预测，表示使用的GPU卡
@@ -395,7 +398,7 @@ op:
            #Fetch结果列表，以client_config中fetch_var的alias_name为准
            fetch_list: ["ctc_greedy_decoder_0.tmp_0", "softmax_0.tmp_0"]

-            # device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu
+            # device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu, 5=arm ascend310, 6=arm ascend910
            device_type: 0

            #计算硬件ID，当devices为""或不写时为CPU预测；当devices为"0", "0,1,2"时为GPU预测，表示使用的GPU卡
@@ -434,10 +437,12 @@ Python Pipeline除了支持CPU、GPU之外，还支持多种异构硬件部署
 - TensorRT : 2
 - CPU(Arm) : 3
 - XPU : 4
+- Ascend310(Arm) : 5
+- Ascend910(Arm) : 6

 config.yml中硬件配置：
 ```YAML
-#计算硬件类型: 空缺时由devices决定(CPU/GPU)，0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu
+#计算硬件类型: 空缺时由devices决定(CPU/GPU)，0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu, 5=arm ascend310, 6=arm ascend910
 device_type: 0
 #计算硬件ID，优先由device_type决定硬件类型。devices为""或空缺时为CPU预测；当为"0", "0,1,2"时为GPU预测，表示使用的GPU卡
 devices: "" # "0,1"

--- a/doc/Serving_Configure_EN.md
+++ b/doc/Serving_Configure_EN.md
@@ -84,8 +84,8 @@ More flags:
 | Argument                                       | Type | Default | Description                                           |
 | ---------------------------------------------- | ---- | ------- | ----------------------------------------------------- |
 | `thread`                                       | int  | `2`     | Number of brpc service thread                         |
-| `op_num`                                       | int[]| `0`     | Thread Number for each model in asynchronous mode     |
-| `op_max_batch`                                 | int[]| `32`    | Batch Number for each model in asynchronous mode      |
+| `runtime_thread_num`                           | int[]| `0`     | Thread Number for each model in asynchronous mode     |
+| `batch_infer_size`                             | int[]| `32`    | Batch Number for each model in asynchronous mode      |
 | `gpu_ids`                                      | str[]| `"-1"`  | Gpu card id for each model                            |
 | `port`                                         | int  | `9292`  | Exposed port of current service to users              |
 | `model`                                        | str[]| `""`    | Path of paddle model directory to be served           |
@@ -98,6 +98,7 @@ More flags:
 | `precision`                                    | str  | FP32    | Precision Mode, support FP32, FP16, INT8              |
 | `use_calib`                                    | bool | False   | Use TRT int8 calibration                              |
 | `gpu_multi_stream`                             | bool | False   | EnableGpuMultiStream to get larger QPS                |
+| `use_ascend_cl`                                | bool | False   | Enable for ascend910; Use with use_lite for ascend310 |

 #### Serving model with multiple gpus.
 ```BASH
@@ -258,6 +259,7 @@ engines {
  use_gpu: false
  combined_model: false
  gpu_multi_stream: false
+  use_ascend_cl: false
  runtime_thread_num: 0
  batch_infer_size: 32
  enable_overrun: false
@@ -293,6 +295,7 @@ gpu_ids: 2
 - use_gpu: Enbale GPU.
 - combined_model: Enable combined model.
 - gpu_multi_stream: Enable gpu multiple stream mode.
+- use_ascend_cl: Enable Ascend, use individually for ascend910, use with lite for ascend310
 - runtime_thread_num: Enable Async mode when num greater than 0 and creating predictors.
 - batch_infer_size: The max batch size of Async mode.
 - enable_overrun: Enable over running of Async mode which means putting the whole task into the task queue.
@@ -380,7 +383,7 @@ op:
            #Fetch data list
            fetch_list: ["concat_1.tmp_0"]

-            # device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu
+            # device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu, 5=arm ascend310, 6=arm ascend910
            device_type: 0

            #Device ID
@@ -418,7 +421,7 @@ op:
            #Fetch data list
            fetch_list: ["ctc_greedy_decoder_0.tmp_0", "softmax_0.tmp_0"]

-            # device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu
+            # device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu, 5=arm ascend310, 6=arm ascend910
            device_type: 0
            
            #Device ID
@@ -459,10 +462,12 @@ In addition to supporting CPU and GPU, Pipeline also supports the deployment of
 - TensorRT : 2
 - CPU(Arm) : 3
 - XPU : 4
+- Ascend310(Arm) : 5
+- Ascend910(Arm) : 6

 Reference config.yaml:
 ```YAML
-# device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu
+# device_type, 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu, 5=arm ascend310, 6=arm ascend910
 device_type: 0
 devices: "" # "0,1"
 ```

--- a/paddle_inference/paddle/include/paddle_engine.h
+++ b/paddle_inference/paddle/include/paddle_engine.h
@@ -41,6 +41,9 @@ using paddle_infer::CreatePredictor;
 DECLARE_int32(gpuid);
 DECLARE_string(precision);
 DECLARE_bool(use_calib);
+DECLARE_string(nnadapter_device_names);
+DECLARE_string(nnadapter_context_properties);
+DECLARE_string(nnadapter_model_cache_dir);

 static const int max_batch = 32;
 static const int min_subgraph_size = 3;
@@ -237,6 +240,7 @@ class PaddleInferenceEngine : public EngineCore {

    if (engine_conf.has_use_lite() && engine_conf.use_lite()) {
      config.EnableLiteEngine(precision_type, true);
+      config.SwitchIrOptim(true);
    }

    if ((!engine_conf.has_use_lite() && !engine_conf.has_use_gpu()) ||
@@ -269,6 +273,33 @@ class PaddleInferenceEngine : public EngineCore {
      config.SetXpuDeviceId(gpu_id);
    }

+    if (engine_conf.has_use_ascend_cl() &&
+        engine_conf.use_ascend_cl()) {
+      if (engine_conf.has_use_lite() && engine_conf.use_lite()) {
+        // for ascend 310 
+        FLAGS_nnadapter_device_names = "huawei_ascend_npu";
+        FLAGS_nnadapter_context_properties =
+                "HUAWEI_ASCEND_NPU_SELECTED_DEVICE_IDS=" +
+                std::to_string(gpu_id);
+        FLAGS_nnadapter_model_cache_dir = "";
+        config.NNAdapter()
+        .Enable()
+        .SetDeviceNames({FLAGS_nnadapter_device_names})
+        .SetContextProperties(FLAGS_nnadapter_context_properties)
+        .SetModelCacheDir(FLAGS_nnadapter_model_cache_dir);
+        LOG(INFO) << "Enable Lite NNAdapter for Ascend,"
+                  << "nnadapter_device_names="
+                  << FLAGS_nnadapter_device_names
+                  << ",nnadapter_context_properties="
+                  << FLAGS_nnadapter_context_properties
+                  << ",nnadapter_model_cache_dir="
+                  << FLAGS_nnadapter_model_cache_dir;
+      } else {
+        // for ascend 910
+	      config.EnableNpu(gpu_id);
+      }
+    }
+
    if (engine_conf.has_enable_memory_optimization() &&
        engine_conf.enable_memory_optimization()) {
      config.EnableMemoryOptim();

--- a/paddle_inference/paddle/src/paddle_engine.cpp
+++ b/paddle_inference/paddle/src/paddle_engine.cpp
@@ -22,6 +22,11 @@ namespace inference {
 DEFINE_int32(gpuid, 0, "GPU device id to use");
 DEFINE_string(precision, "fp32", "precision to deploy, default is fp32");
 DEFINE_bool(use_calib, false, "calibration mode, default is false");
+DEFINE_string(nnadapter_device_names, "", "Names of nnadapter device");
+DEFINE_string(nnadapter_context_properties,
+              "",
+              "Properties of nnadapter context");
+DEFINE_string(nnadapter_model_cache_dir, "", "Cache dir of nnadapter model");

 REGIST_FACTORY_OBJECT_IMPL_WITH_NAME(
    ::baidu::paddle_serving::predictor::FluidInferEngine<PaddleInferenceEngine>,

--- a/python/CMakeLists.txt
+++ b/python/CMakeLists.txt
@@ -12,6 +12,10 @@ if (SERVER)
    set(SERVER_PACKAGE_NAME "paddle-serving-server-gpu")
  elseif(WITH_XPU)
    set(SERVER_PACKAGE_NAME "paddle-serving-server-xpu")
+  elseif(WITH_ROCM)
+    set(SERVER_PACKAGE_NAME "paddle-serving-server-rocm")
+  elseif(WITH_ASCEND_CL)
+    set(SERVER_PACKAGE_NAME "paddle-serving-server-npu")
  endif()
  file(INSTALL pipeline DESTINATION paddle_serving_server)
  file(GLOB_RECURSE SERVING_SERVER_PY_FILES paddle_serving_server/*.py)

--- a/python/gen_version.py
+++ b/python/gen_version.py
@@ -43,6 +43,10 @@ if package_name.endswith('gpu'):
    update_info("paddle_serving_server/version.py", "device_type", "1")
 elif package_name.endswith('xpu'):
    update_info("paddle_serving_server/version.py", "device_type", "2")
+elif package_name.endswith('rocm'):
+    update_info("paddle_serving_server/version.py", "device_type", "3")
+elif package_name.endswith('npu'):
+    update_info("paddle_serving_server/version.py", "device_type", "4")

 path = "paddle_serving_" + sys.argv[1]
 commit_id = subprocess.check_output(['git', 'rev-parse', 'HEAD'])

--- a/python/paddle_serving_app/local_predict.py
+++ b/python/paddle_serving_app/local_predict.py
@@ -86,7 +86,8 @@ class LocalPredictor(object):
                          mkldnn_cache_capacity=0,
                          mkldnn_op_list=None,
                          mkldnn_bf16_op_list=None,
-                          use_feed_fetch_ops=False):
+                          use_feed_fetch_ops=False,
+                          use_ascend_cl=False):
        """
        Load model configs and create the paddle predictor by Paddle Inference API.
   
@@ -108,6 +109,7 @@ class LocalPredictor(object):
            mkldnn_op_list: op list accelerated using MKLDNN, None default.
            mkldnn_bf16_op_list: op list accelerated using MKLDNN bf16, None default.
            use_feed_fetch_ops: use feed/fetch ops, False default.
+            use_ascend_cl: run predict on Huawei Ascend, False default
        """
        gpu_id = int(gpu_id)
        client_config = "{}/serving_server_conf.prototxt".format(model_path)
@@ -146,11 +148,12 @@ class LocalPredictor(object):
            "gpu_id:{}, use_profile:{}, thread_num:{}, mem_optim:{}, ir_optim:{}, "
            "use_trt:{}, use_lite:{}, use_xpu:{}, precision:{}, use_calib:{}, "
            "use_mkldnn:{}, mkldnn_cache_capacity:{}, mkldnn_op_list:{}, "
-            "mkldnn_bf16_op_list:{}, use_feed_fetch_ops:{}, ".format(
+            "mkldnn_bf16_op_list:{}, use_feed_fetch_ops:{}, "
+            "use_ascend_cl:{} ".format(
                model_path, use_gpu, gpu_id, use_profile, thread_num, mem_optim,
                ir_optim, use_trt, use_lite, use_xpu, precision, use_calib,
                use_mkldnn, mkldnn_cache_capacity, mkldnn_op_list,
-                mkldnn_bf16_op_list, use_feed_fetch_ops))
+                mkldnn_bf16_op_list, use_feed_fetch_ops, use_ascend_cl))

        self.feed_names_ = [var.alias_name for var in model_conf.feed_var]
        self.fetch_names_ = [var.alias_name for var in model_conf.fetch_var]
@@ -215,11 +218,28 @@ class LocalPredictor(object):
                zero_copy=True,
                passes_filter=[],
                ops_filter=[])
+            config.switch_ir_optim(True)
        # set xpu
        if use_xpu:
            # 2MB l3 cache
            config.enable_xpu(8 * 1024 * 1024)
            config.set_xpu_device_id(gpu_id)
+        # set ascend cl
+        if use_ascend_cl:
+            if use_lite:
+                # for ascend 310
+                nnadapter_device_names = "huawei_ascend_npu"
+                nnadapter_context_properties = \
+                    "HUAWEI_ASCEND_NPU_SELECTED_DEVICE_IDS={}".format(gpu_id)
+                nnadapter_model_cache_dir = ""
+                config.nnadapter() \
+                .enable() \
+                .set_device_names([nnadapter_device_names]) \
+                .set_context_properties(nnadapter_context_properties) \
+                .set_model_cache_dir(nnadapter_model_cache_dir)
+            else:
+                # for ascend 910
+                config.enable_npu(gpu_id)
        # set cpu low precision
        if not use_gpu and not use_lite:
            if precision_type == paddle_infer.PrecisionType.Int8:

--- a/python/paddle_serving_server/serve.py
+++ b/python/paddle_serving_server/serve.py
@@ -183,6 +183,11 @@ def serve_args():
        "--use_lite", default=False, action="store_true", help="Use PaddleLite")
    parser.add_argument(
        "--use_xpu", default=False, action="store_true", help="Use XPU")
+    parser.add_argument(
+        "--use_ascend_cl",
+        default=False,
+        action="store_true",
+        help="Use Ascend CL")
    parser.add_argument(
        "--product_name",
        type=str,
@@ -299,13 +304,15 @@ def start_gpu_card_model(gpu_mode, port, args):  # pylint: disable=doc-string-mi
    server.set_device(device)
    if args.use_xpu:
        server.set_xpu()
+    if args.use_ascend_cl:
+        server.set_ascend_cl()

    if args.product_name != None:
        server.set_product_name(args.product_name)
    if args.container_id != None:
        server.set_container_id(args.container_id)

-    if gpu_mode == True:
+    if gpu_mode == True or args.use_xpu or args.use_ascend_cl:
        server.set_gpuid(args.gpu_ids)
    server.load_model_config(model)
    server.prepare_server(

--- a/python/paddle_serving_server/server.py
+++ b/python/paddle_serving_server/server.py
@@ -88,6 +88,7 @@ class Server(object):
        self.gpu_multi_stream = False
        self.use_lite = False
        self.use_xpu = False
+        self.use_ascend_cl = False
        self.model_config_paths = collections.OrderedDict()
        self.product_name = None
        self.container_id = None
@@ -195,6 +196,9 @@ class Server(object):
    def set_xpu(self):
        self.use_xpu = True

+    def set_ascend_cl(self):
+        self.use_ascend_cl = True
+
    def _prepare_engine(self, model_config_paths, device, use_encryption_model):
        self.device = device
        if self.model_toolkit_conf == None:
@@ -208,6 +212,8 @@ class Server(object):
            if self.device == "gpu" or self.use_trt or self.gpu_multi_stream:
                self.gpuid = ["0"]
                self.device = "gpu"
+            elif self.use_xpu or self.use_ascend_cl:
+                self.gpuid = ["0"]
            else:
                self.gpuid = ["-1"]

@@ -244,6 +250,7 @@ class Server(object):
            engine.gpu_multi_stream = self.gpu_multi_stream
            engine.use_lite = self.use_lite
            engine.use_xpu = self.use_xpu
+            engine.use_ascend_cl = self.use_ascend_cl
            engine.use_gpu = False

            if len(self.gpuid) == 0:
@@ -436,6 +443,13 @@ class Server(object):
                device_version = "gpu-cuda" + version_suffix
        elif device_type == "2":
            device_version = "xpu-" + platform.machine()
+        elif device_type == "3":
+            device_version = "rocm-" + platform.machine()
+        elif device_type == "4":
+            if self.use_lite:
+                device_version = "ascendcl-lite-" + platform.machine()
+            else:
+                device_version = "ascendcl-" + platform.machine()
        return device_version

    def download_bin(self):

--- a/python/pipeline/local_service_handler.py
+++ b/python/pipeline/local_service_handler.py
@@ -86,6 +86,7 @@ class LocalServiceHandler(object):
        self._use_trt = False
        self._use_lite = False
        self._use_xpu = False
+        self._use_ascend_cl = False
        self._use_mkldnn = False
        self._mkldnn_cache_capacity = 0
        self._mkldnn_op_list = None
@@ -129,6 +130,17 @@ class LocalServiceHandler(object):
            devices = [int(x) for x in devices.split(",")]
            self._use_lite = True
            self._use_xpu = True
+        elif device_type == 5:
+            # Ascend 310 ARM CPU
+            self._device_name = "arm"
+            devices = [int(x) for x in devices.split(",")]
+            self._use_lite = True
+            self._use_ascend_cl = True
+        elif device_type == 6:
+            # Ascend 910 ARM CPU
+            self._device_name = "arm"
+            devices = [int(x) for x in devices.split(",")]
+            self._use_ascend_cl = True
        else:
            _LOGGER.error(
                "LocalServiceHandler initialization fail. device_type={}"
@@ -163,13 +175,14 @@ class LocalServiceHandler(object):
            "mem_optim:{}, ir_optim:{}, use_profile:{}, thread_num:{}, "
            "client_type:{}, fetch_names:{}, precision:{}, use_mkldnn:{}, "
            "mkldnn_cache_capacity:{}, mkldnn_op_list:{}, "
-            "mkldnn_bf16_op_list:{}".format(
+            "mkldnn_bf16_op_list:{}, use_ascend_cl:{}".format(
                model_config, self._device_name, self._use_gpu, self._use_trt,
                self._use_lite, self._use_xpu, device_type, self._devices,
                self._mem_optim, self._ir_optim, self._use_profile,
                self._thread_num, self._client_type, self._fetch_names,
                self._precision, self._use_mkldnn, self._mkldnn_cache_capacity,
-                self._mkldnn_op_list, self._mkldnn_bf16_op_list))
+                self._mkldnn_op_list, self._mkldnn_bf16_op_list,
+                self._use_ascend_cl))

    def get_fetch_list(self):
        return self._fetch_names
@@ -225,7 +238,8 @@ class LocalServiceHandler(object):
                use_mkldnn=self._use_mkldnn,
                mkldnn_cache_capacity=self._mkldnn_cache_capacity,
                mkldnn_op_list=self._mkldnn_op_list,
-                mkldnn_bf16_op_list=self._mkldnn_bf16_op_list)
+                mkldnn_bf16_op_list=self._mkldnn_bf16_op_list,
+                use_ascend_cl=self._use_ascend_cl)
        return self._local_predictor_client

    def get_client_config(self):
@@ -284,6 +298,8 @@ class LocalServiceHandler(object):
                server.set_xpu()
            if self._use_lite:
                server.set_lite()
+            if self._use_ascend_cl:
+                server.set_ascend_cl()

        server.set_op_sequence(op_seq_maker.get_op_sequence())
        server.set_num_threads(thread_num)