[Serving] Support FastDeploy XPU Triton Server (#1994)

* [patchelf] fix patchelf error for inference xpu * [serving] add xpu dockerfile and support fd server * [serving] add xpu dockerfile and support fd server * [Serving] support XPU + Tritron * [Serving] support XPU + Tritron * [Dockerfile] update xpu tritron docker file -> paddle 0.0.0 * [Dockerfile] update xpu tritron docker file -> paddle 0.0.0 * [Dockerfile] update xpu tritron docker file -> paddle 0.0.0 * [Dockerfile] add comments for xpu tritron dockerfile * [Doruntime] fix xpu infer error * [Doruntime] fix xpu infer error * [XPU] update xpu dockerfile * add xpu triton server docs * add xpu triton server docs * add xpu triton server docs * add xpu triton server docs * update xpu triton server docs * update xpu triton server docs * update xpu triton server docs * update xpu triton server docs * update xpu triton server docs * update xpu triton server docs * update xpu triton server docs * update xpu triton server docs
2025-10-06 09:07:10 +08:00 · 2023-05-29 14:38:25 +08:00
parent 3a9904411a
commit 434b48dda5
10 changed files with 517 additions and 49 deletions
--- a/.gitignore
+++ b/.gitignore
@@ -43,4 +43,8 @@ examples/vision/tests_quantize
 fastdeploy/LICENSE
 fastdeploy/ThirdPartyNotices.txt
 FastDeployCSharp.cmake
-python/fastdeploy/code_version.py
+python/fastdeploy/code_version.py
+*.pdmodel
+*.pdiparams
+*.pdiparams.info
+log.txt
--- a/examples/vision/classification/paddleclas/serving/models/runtime/config.pbtxt
+++ b/examples/vision/classification/paddleclas/serving/models/runtime/config.pbtxt
@@ -10,7 +10,7 @@ input [
    name: "inputs"
    # input type such as TYPE_FP32、TYPE_UINT8、TYPE_INT8、TYPE_INT16、TYPE_INT32、TYPE_INT64、TYPE_FP16、TYPE_STRING
    data_type: TYPE_FP32
-    # input shape， The batch dimension is omitted and the actual shape is [batch, c, h, w]
+    # input shape, The batch dimension is omitted and the actual shape is [batch, c, h, w]
    dims: [ 3, 224, 224 ]
  }
 ]
@@ -31,6 +31,7 @@ instance_group [
    count: 1
    # Use GPU, CPU inference option is:KIND_CPU
    kind: KIND_GPU
+    # kind: KIND_CPU
    # The instance is deployed on the 0th GPU card
    gpus: [0]
  }
@@ -58,3 +59,32 @@ optimization {
  }
  ]
 }}
+
+# instance_group [
+#   {
+#     # The number of instances is 1
+#     count: 1
+#     # Use GPU, CPU inference option is:KIND_CPU
+#     # kind: KIND_GPU
+#     kind: KIND_CPU
+#     # The instance is deployed on the 0th GPU card
+#     # gpus: [0]
+#   }
+# ]
+
+# optimization {
+#   execution_accelerators {
+#   cpu_execution_accelerator: [{
+#     name: "paddle_xpu",
+#     parameters { key: "cpu_threads" value: "4" }
+#     parameters { key: "use_paddle_log" value: "1" }
+#     parameters { key: "kunlunxin_id" value: "0" }
+#     parameters { key: "l3_workspace_size" value: "62914560" }
+#     parameters { key: "locked" value: "0" }
+#     parameters { key: "autotune" value: "1" }
+#     parameters { key: "precision" value: "int16" }
+#     parameters { key: "adaptive_seqlen" value: "0" }
+#     parameters { key: "enable_multi_stream" value: "0" }
+#     parameters { key: "gm_default_size" value: "0" }
+#     }]
+# }}
--- a/fastdeploy/runtime/runtime.cc
+++ b/fastdeploy/runtime/runtime.cc
@@ -204,7 +204,15 @@ bool Runtime::Infer(std::vector<FDTensor>& input_tensors,
 }

 bool Runtime::Infer() {
-  bool result = backend_->Infer(input_tensors_, &output_tensors_, false);
+  bool result = false;
+  if (option.device == Device::KUNLUNXIN) {
+    // FDTensor SetExternalData is not support for Device::KUNLUNXIN
+    // now, so, we need to set copy_to_fd as 'true'.
+    result = backend_->Infer(input_tensors_, &output_tensors_, true);
+  } else {
+    result = backend_->Infer(input_tensors_, &output_tensors_, false);
+  }
+
  for (auto& tensor : output_tensors_) {
    tensor.device_id = option.device_id;
  }
--- a/scripts/patch_paddle_inference.py
+++ b/scripts/patch_paddle_inference.py
@@ -26,6 +26,7 @@ def process_paddle_inference(paddle_inference_so_file):
    rpaths = [
        "$ORIGIN", "$ORIGIN/../../third_party/install/mkldnn/lib/",
        "$ORIGIN/../../third_party/install/mklml/lib/",
+        "$ORIGIN/../../third_party/install/xpu/lib/",
        "$ORIGIN/../../../tensorrt/lib/"
    ]

--- a/serving/Dockerfile_xpu
+++ b/serving/Dockerfile_xpu
@@ -0,0 +1,47 @@
+# Copyright (c) 2022 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.
+
+ARG http_proxy
+ARG https_proxy
+ARG no_proxy
+
+FROM paddlepaddle/fastdeploy:21.10-cpu-only-min
+
+ENV TZ=Asia/Shanghai \
+    DEBIAN_FRONTEND=noninteractive \
+    http_proxy=$http_proxy \
+    https_proxy=$https_proxy \
+    no_proxy=$no_proxy
+
+# Note: Here, use nightly built of paddle for xpu tritron server image 
+# to avoid the so confilcts between paddle and fastdeploy-python.
+RUN apt-get update && apt-get install -y --no-install-recommends apt-utils libgomp1 ffmpeg libsm6 libxext6 vim wget \
+    && python3 -m pip install -U pip \
+    && python3 -m pip install paddlepaddle==0.0.0 -f https://www.paddlepaddle.org.cn/whl/linux/cpu-mkl/develop.html \
+    && python3 -m pip install paddlenlp fast-tokenizer-python
+
+COPY python/dist/*.whl /opt/fastdeploy/ 
+RUN python3 -m pip install  /opt/fastdeploy/*.whl \
+    && rm -rf /opt/fastdeploy/*.whl
+
+COPY serving/build/libtriton_fastdeploy.so /opt/tritonserver/backends/fastdeploy/
+COPY build/fastdeploy_install /opt/fastdeploy/
+COPY benchmark/cpp /opt/fastdeploy/benchmark/cpp
+
+RUN mv /opt/tritonserver/bin/tritonserver /opt/tritonserver/bin/fastdeployserver
+ENV LD_LIBRARY_PATH="/opt/fastdeploy/lib:/opt/fastdeploy/third_libs/install/opencv/lib64:/opt/fastdeploy/third_libs/install/paddle2onnx/lib:/opt/fastdeploy/third_libs/install/paddle_inference/paddle/lib:/opt/fastdeploy/third_libs/install/paddle_inference/third_party/install/mkldnn/lib:/opt/fastdeploy/third_libs/install/paddle_inference/third_party/install/mklml/lib:/opt/fastdeploy/third_libs/install/paddle_inference/third_party/install/xpu/lib:$LD_LIBRARY_PATH"
+# unset proxy
+ENV http_proxy=
+ENV https_proxy=
+ENV no_proxy=
--- a/serving/docs/zh_CN/compile.md
+++ b/serving/docs/zh_CN/compile.md
@@ -65,6 +65,19 @@ cd ../
 docker build -t paddlepaddle/fastdeploy:x.y.z-ipu-only-21.10 -f serving/Dockerfile_ipu .
 ```

+### 制作XPU镜像
+
+```
+# 进入serving目录执行脚本编译fastdeploy和服务化的backend
+cd serving
+bash scripts/build_fd_xpu.sh
+
+# 退出到FastDeploy主目录，制作镜像
+# x.y.z为FastDeploy版本号，可根据情况自己确定。比如: 1.0.6
+cd ../
+docker build -t paddlepaddle/fastdeploy:x.y.z-xpu-21.10 -f serving/Dockerfile_xpu .
+```
+
 ## 非镜像方式编译

 - [FastDeploy Serving CentOS编译教程](./compile_without_docker_centos.md)
--- a/serving/docs/zh_CN/model_configuration.md
+++ b/serving/docs/zh_CN/model_configuration.md
@@ -112,6 +112,37 @@ optimization {
 }
 ```

+#### 配置使用Paddle+XPU引擎
+```
+optimization {
+  execution_accelerators {
+    # XPU推理配置通过CPU Execution启动， 配合KIND_CPU使用
+    cpu_execution_accelerator: [
+      {
+        name: "paddle_xpu",
+        # CPU相关配置
+        # cpu_threads: CPU计算线程数
+        # use_paddle_log: 开启paddle log信息
+        parameters { key: "cpu_threads" value: "4" }
+        parameters { key: "use_paddle_log" value: "0" }
+        # XPU相关配置
+        # kunlunxin_id: 使用的XPU卡号
+        # l3_workspace_size: L3缓存size
+        parameters { key: "kunlunxin_id" value: "0" }
+        parameters { key: "l3_workspace_size" value: "0xfffc00" }
+        parameters { key: "locked" value: "0" }
+        parameters { key: "autotune" value: "1" }
+        parameters { key: "precision" value: "int16" }
+        parameters { key: "adaptive_seqlen" value: "0" }
+        parameters { key: "enable_multi_stream" value: "0" }
+        parameters { key: "gm_default_size" value: "0" }
+      }
+    ]
+  }
+}
+```
+
+
 ### 配置使用ONNXRuntime引擎
 除去配置 *Instance Groups*，决定模型运行在CPU还是GPU上。ONNXRuntime引擎中，还可以进行如下配置，具体例子可参照[YOLOv5的Runtime配置](../../../examples/vision/detection/yolov5/serving/models/runtime/config.pbtxt):

--- a/serving/docs/zh_CN/xpu.md
+++ b/serving/docs/zh_CN/xpu.md
@@ -0,0 +1,190 @@
+# FastDeploy XPU Triton Server使用文档
+FastDeploy XPU Triton Server通过Paddle Inference调用XPU进行推理，并且已经接入到 Triton Server。在FastDeploy XPU Triton Server中，使用XPU推理需要通过CPU instance_group和cpu_execution_accelerator进行配置和调用。本文档以PaddleClas为例，讲述如何把一个CPU/GPU Triton服务，改造成XPU Triton服务。
+
+## 1. 准备服务化镜像  
+
+- 下载FastDeploy XPU Triton Server镜像  
+```bash
+docker pull registry.baidubce.com/paddlepaddle/fastdeploy:1.0.7-xpu-21.10  # 稳定版
+docker pull registry.baidubce.com/paddlepaddle/fastdeploy:0.0.0-xpu-21.10  # develop版本
+```  
+
+- 下载部署示例代码
+```bash
+# 下载部署示例代码
+git clone https://github.com/PaddlePaddle/FastDeploy.git
+cd  FastDeploy/examples/vision/classification/paddleclas/serving
+
+# 下载ResNet50_vd模型文件和测试图片
+wget https://bj.bcebos.com/paddlehub/fastdeploy/ResNet50_vd_infer.tgz
+tar -xvf ResNet50_vd_infer.tgz
+
+# 将配置文件放入预处理目录
+mv ResNet50_vd_infer/inference_cls.yaml models/preprocess/1/inference_cls.yaml
+
+# 将模型放入 models/runtime/1目录下, 并重命名为model.pdmodel和model.pdiparams
+mv ResNet50_vd_infer/inference.pdmodel models/runtime/1/model.pdmodel
+mv ResNet50_vd_infer/inference.pdiparams models/runtime/1/model.pdiparams
+```
+
+## 2. 启动容器
+```bash
+docker run -itd --name fd_xpu_server -v `pwd`/:/serving --net=host --privileged registry.baidubce.com/paddlepaddle/fastdeploy:1.0.7-xpu-21.10 /bin/bash
+```
+
+## 3. 验证XPU可用性
+```bash
+docker exec -it fd_xpu_server /bin/bash
+cd /opt/fastdeploy/benchmark/cpp/build
+./benchmark --model ResNet50_infer --config_path ../config/config.xpu.paddle.fp32.txt --enable_log_info
+cd /serving
+```
+输出为：  
+```
+I0529 11:07:46.860354   222 memory_optimize_pass.cc:222] Cluster name : batch_norm_46.tmp_2_max  size: 1
+--- Running analysis [ir_graph_to_program_pass]
+I0529 11:07:46.889616   222 analysis_predictor.cc:1705] ======= optimize end =======
+I0529 11:07:46.890262   222 naive_executor.cc:160] ---  skip [feed], feed -> inputs
+I0529 11:07:46.890703   222 naive_executor.cc:160] ---  skip [save_infer_model/scale_0.tmp_1], fetch -> fetch
+[INFO] fastdeploy/runtime/runtime.cc(286)::CreatePaddleBackend	Runtime initialized with Backend::PDINFER in Device::KUNLUNXIN.
+[INFO] fastdeploy/runtime/backends/paddle/paddle_backend.cc(341)::Infer	Running profiling for Runtime without H2D and D2H, Repeats: 1000, Warmup: 200
+Runtime(ms): 0.706382ms.
+```
+显示启动的设备类型为：Device::KUNLUNXIN。FastDeploy Benchmark工具使用文档，请参考[benchmark](https://github.com/PaddlePaddle/FastDeploy/tree/develop/benchmark/cpp).
+
+## 4. 配置Triton Model Config  
+```protobuf
+# XPU服务化案例: examples/vision/classification/serving/models/runtime/config.pbtxt
+# 将XPU部分的注释撤销，并注释掉原来的GPU设置，修改为：
+# # Number of instances of the model
+# instance_group [
+#   {
+#     # The number of instances is 1
+#     count: 1
+#     # Use GPU, CPU inference option is:KIND_CPU
+#     kind: KIND_GPU
+#     # kind: KIND_CPU
+#     # The instance is deployed on the 0th GPU card
+#     gpus: [0]
+#   }
+# ]
+
+# optimization {
+#   execution_accelerators {
+#   gpu_execution_accelerator : [ {
+#     # use TRT engine
+#     name: "tensorrt",
+#     # use fp16 on TRT engine
+#     parameters { key: "precision" value: "trt_fp16" }
+#   },
+#   {
+#     name: "min_shape"
+#     parameters { key: "inputs" value: "1 3 224 224" }
+#   },
+#   {
+#     name: "opt_shape"
+#     parameters { key: "inputs" value: "1 3 224 224" }
+#   },
+#   {
+#     name: "max_shape"
+#     parameters { key: "inputs" value: "16 3 224 224" }
+#   }
+#   ]
+# }}
+
+instance_group [
+  {
+    # The number of instances is 1
+    count: 1
+    # Use GPU, CPU inference option is:KIND_CPU
+    # kind: KIND_GPU
+    kind: KIND_CPU
+    # The instance is deployed on the 0th GPU card
+    # gpus: [0]
+  }
+]
+
+optimization {
+  execution_accelerators {
+  cpu_execution_accelerator: [{
+    name: "paddle_xpu",
+    parameters { key: "cpu_threads" value: "4" }
+    parameters { key: "use_paddle_log" value: "1" }
+    parameters { key: "kunlunxin_id" value: "0" }
+    parameters { key: "l3_workspace_size" value: "62914560" }
+    parameters { key: "locked" value: "0" }
+    parameters { key: "autotune" value: "1" }
+    parameters { key: "precision" value: "int16" }
+    parameters { key: "adaptive_seqlen" value: "0" }
+    parameters { key: "enable_multi_stream" value: "0" }
+    parameters { key: "gm_default_size" value: "0" }
+    }]
+}}
+```
+
+## 5. 启动Triton服务  
+```bash
+fastdeployserver --model-repository=/serving/models --backend-config=python,shm-default-byte-size=10485760
+```  
+输出：
+```
+[INFO] fastdeploy/runtime/runtime.cc(286)::CreatePaddleBackend	Runtime initialized with Backend::PDINFER in Device::KUNLUNXIN.
+.....
+I0529 03:54:40.585326 385 server.cc:592]
+-------------+---------+--------+
+| Model       | Version | Status |
+-------------+---------+--------+
+| paddlecls   | 1       | READY  |
+| postprocess | 1       | READY  |
+| preprocess  | 1       | READY  |
+| runtime     | 1       | READY  |
+-------------+---------+--------+
+......
+I0529 03:54:40.586430 385 grpc_server.cc:4117] Started GRPCInferenceService at 0.0.0.0:8001
+I0529 03:54:40.586657 385 http_server.cc:2815] Started HTTPService at 0.0.0.0:8000
+I0529 03:54:40.627382 385 http_server.cc:167] Started Metrics Service at 0.0.0.0:8002
+```
+
+## 6. 客户端请求  
+在物理机器中执行以下命令，发送grpc请求并输出结果:
+```bash
+# 下载测试图片
+wget https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00000010.jpeg
+
+# 安装客户端依赖
+python3 -m pip install tritonclient\[all\]
+
+# 发送请求
+python3 paddlecls_grpc_client.py
+```
+
+发送请求成功后，会返回json格式的检测结果并打印输出:
+```bash
+output_name: CLAS_RESULT
+{'label_ids': [153], 'scores': [0.6858349442481995]}
+```
+以上测试结果为Paddle Inference Backend + XPU R200下的输出。
+
+## 7. 容器内自测  
+如果是想在容器内自测，则运行以下命令：  
+```bash
+cd /serving
+# 后台挂载
+nohup fastdeployserver --model-repository=/serving/models --backend-config=python,shm-default-byte-size=10485760 > log.txt 2>&1 &
+# 安装客户端依赖
+python3 -m pip install tritonclient\[all\]
+# 发送请求
+unset http_proxy
+unset https_proxy
+python3 paddlecls_grpc_client.py
+```
+
+## 8. 配置修改
+
+当前默认配置在XPU运行Paddle Inference引擎， 如果要在CPU/GPU其他推理引擎上运行。 需要修改`models/runtime/config.pbtxt`中配置，详情请参考[配置文档](./model_configuration.md).
+
+## 9. 常见问题
+- [如何编写客户端 HTTP/GRPC 请求](https://github.com/PaddlePaddle/FastDeploy/blob/develop/serving/docs/zh_CN/client.md)
+- [如何编译服务化部署镜像](https://github.com/PaddlePaddle/FastDeploy/blob/develop/serving/docs/zh_CN/compile.md)
+- [服务化部署原理及动态Batch介绍](https://github.com/PaddlePaddle/FastDeploy/blob/develop/serving/docs/zh_CN/demo.md)
+- [模型仓库介绍](https://github.com/PaddlePaddle/FastDeploy/blob/develop/serving/docs/zh_CN/model_repository.md)
--- a/serving/scripts/build_fd_xpu.sh
+++ b/serving/scripts/build_fd_xpu.sh
@@ -0,0 +1,65 @@
+#!/usr/bin/env bash
+# Copyright (c) 2022 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.
+
+
+echo "start build FD XPU library"
+
+docker run -i --rm --name build_fd_xpu \
+           -v `pwd`/..:/workspace/fastdeploy \
+           -e "http_proxy=${http_proxy}" \
+           -e "https_proxy=${https_proxy}" \
+           -e "no_proxy=${no_proxy}" \
+           --network=host --privileged \
+           paddlepaddle/fastdeploy:21.10-cpu-only-buildbase \
+           bash -c \
+           'export https_proxy_tmp=${https_proxy}
+            export http_proxy_tmp=${http_proxy}
+            cd /workspace/fastdeploy/python;
+            rm -rf .setuptools-cmake-build dist build fastdeploy/libs/third_libs;
+            ln -s /usr/bin/python3 /usr/bin/python;
+            export WITH_GPU=OFF;
+            export ENABLE_ORT_BACKEND=OFF;
+            export ENABLE_PADDLE_BACKEND=OFF;
+            export ENABLE_OPENVINO_BACKEND=OFF;
+            export ENABLE_VISION=ON;
+            export ENABLE_TEXT=ON;
+            unset http_proxy
+            unset https_proxy
+            python setup.py build;
+            python setup.py bdist_wheel;
+            cd /workspace/fastdeploy;
+            rm -rf build; mkdir build; cd build;
+            cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=${PWD}/fastdeploy_install -DWITH_KUNLUNXIN=ON -DENABLE_PADDLE_BACKEND=ON -DENABLE_VISION=ON -DENABLE_BENCHMARK=ON -DLIBRARY_NAME=fastdeploy_runtime;
+            make -j`nproc`;
+            make install;
+            cd /workspace/fastdeploy/serving;
+            rm -rf build; mkdir build; cd build;
+            export https_proxy=${https_proxy_tmp}
+            export http_proxy=${http_proxy_tmp}
+            cmake .. -DTRITON_ENABLE_GPU=OFF -DFASTDEPLOY_DIR=/workspace/fastdeploy/build/fastdeploy_install -DTRITON_COMMON_REPO_TAG=r21.10 -DTRITON_CORE_REPO_TAG=r21.10 -DTRITON_BACKEND_REPO_TAG=r21.10;
+            make -j`nproc`;
+            cd /workspace/fastdeploy/benchmark/cpp;
+            rm -rf build; mkdir build; cd build;
+            unset http_proxy
+            unset https_proxy
+            cmake .. -DFASTDEPLOY_INSTALL_DIR=/workspace/fastdeploy/build/fastdeploy_install;
+            make -j`nproc`;
+            wget https://bj.bcebos.com/paddlehub/fastdeploy/ResNet50_infer.tgz && tar -zxvf ResNet50_infer.tgz;
+            wget https://bj.bcebos.com/paddlehub/fastdeploy/000000014439.jpg;
+            rm -f ResNet50_infer.tgz;
+            rm -rf CMakeFiles;
+            '
+
+echo "build FD XPU library done"
--- a/serving/src/fastdeploy_runtime.cc
+++ b/serving/src/fastdeploy_runtime.cc
@@ -199,6 +199,9 @@ ModelState::ModelState(TRITONBACKEND_Model* triton_model)
          runtime_options_->UseOrtBackend();
        } else if (name == "paddle") {
          runtime_options_->UsePaddleBackend();
+        } else if (name == "paddle_xpu") {
+          // Note(qiuyanjun): use XPU via paddle inference backend.
+          runtime_options_->UsePaddleInferBackend();
        } else if (name == "openvino") {
          runtime_options_->UseOpenVINOBackend();
        } else if (name != "") {
@@ -212,44 +215,118 @@ ModelState::ModelState(TRITONBACKEND_Model* triton_model)
        }

        triton::common::TritonJson::Value params;
-        if (ea.Find("parameters", &params)) {
-          std::vector<std::string> param_keys;
-          THROW_IF_BACKEND_MODEL_ERROR(params.Members(&param_keys));
-          for (const auto& param_key : param_keys) {
-            std::string value_string;
-            THROW_IF_BACKEND_MODEL_ERROR(
-                params.MemberAsString(param_key.c_str(), &value_string));
-            if (param_key == "cpu_threads") {
-              int cpu_thread_num;
+        if (name == "paddle_xpu") {
+          // parse parameters for cpu host + xpu device.
+          if (ea.Find("parameters", &params)) {
+            std::vector<std::string> param_keys;
+            THROW_IF_BACKEND_MODEL_ERROR(params.Members(&param_keys));
+            // default settings for XPU.
+            int kunlunxin_id = 0;
+            int l3_workspace_size = 0xfffc00;
+            bool locked = false;
+            bool autotune = true;
+            std::string autotune_file = "";
+            std::string precision = "int16";
+            bool adaptive_seqlen = false;
+            bool enable_multi_stream = false;
+            // for future use (only support lite backend now).
+            int gm_default_size = 0;
+            // common settings for cpu host.
+            int cpu_thread_num = -1;
+            bool use_paddle_log = false;
+
+            for (const auto& param_key : param_keys) {
+              std::string value_string;
              THROW_IF_BACKEND_MODEL_ERROR(
-                  ParseIntValue(value_string, &cpu_thread_num));
-              runtime_options_->SetCpuThreadNum(cpu_thread_num);
-            } else if (param_key == "use_mkldnn") {
-              bool pd_enable_mkldnn;
+                  params.MemberAsString(param_key.c_str(), &value_string));
+              // parse common settings for cpu host.
+              if (param_key == "cpu_threads") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseIntValue(value_string, &cpu_thread_num));
+                runtime_options_->SetCpuThreadNum(cpu_thread_num);
+              } else if (param_key == "use_paddle_log") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseBoolValue(value_string, &use_paddle_log));
+                runtime_options_->paddle_infer_option.enable_log_info =
+                    use_paddle_log;
+              } else if (param_key == "is_clone") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseBoolValue(value_string, &is_clone_));
+              } else if (param_key == "encryption_key") {
+                runtime_options_->SetEncryptionKey(value_string);
+                // parse common settings for xpu device.
+              } else if (param_key == "kunlunxin_id") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseIntValue(value_string, &kunlunxin_id));
+              } else if (param_key == "l3_workspace_size") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseIntValue(value_string, &l3_workspace_size));
+              } else if (param_key == "locked") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseBoolValue(value_string, &locked));
+              } else if (param_key == "autotune") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseBoolValue(value_string, &autotune));
+              } else if (param_key == "precision") {
+                precision = value_string;
+              } else if (param_key == "adaptive_seqlen") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseBoolValue(value_string, &adaptive_seqlen));
+              } else if (param_key == "enable_multi_stream") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseBoolValue(value_string, &enable_multi_stream));
+              } else if (param_key == "gm_default_size") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseIntValue(value_string, &gm_default_size));
+              }
+            }
+            // initialize xpu device settings
+            runtime_options_->UseKunlunXin(
+                kunlunxin_id, l3_workspace_size, locked, autotune,
+                autotune_file, precision, adaptive_seqlen, enable_multi_stream,
+                int64_t(gm_default_size));
+          }
+        } else {
+          // parse parameters for cpu only
+          if (ea.Find("parameters", &params)) {
+            std::vector<std::string> param_keys;
+            THROW_IF_BACKEND_MODEL_ERROR(params.Members(&param_keys));
+            for (const auto& param_key : param_keys) {
+              std::string value_string;
              THROW_IF_BACKEND_MODEL_ERROR(
-                  ParseBoolValue(value_string, &pd_enable_mkldnn));
-              runtime_options_->SetPaddleMKLDNN(pd_enable_mkldnn);
-            } else if (param_key == "use_paddle_log") {
-              bool use_paddle_log;
-              THROW_IF_BACKEND_MODEL_ERROR(
-                  ParseBoolValue(value_string, &use_paddle_log));
-              runtime_options_->paddle_infer_option.enable_log_info =
-                  use_paddle_log;
-            } else if (param_key == "num_streams") {
-              int num_streams;
-              THROW_IF_BACKEND_MODEL_ERROR(
-                  ParseIntValue(value_string, &num_streams));
-              runtime_options_->openvino_option.num_streams = num_streams;
-            } else if (param_key == "is_clone") {
-              THROW_IF_BACKEND_MODEL_ERROR(
-                  ParseBoolValue(value_string, &is_clone_));
-            } else if (param_key == "use_ipu") {
-              // runtime_options_->UseIpu();
-            } else if (param_key == "encryption_key") {
-              runtime_options_->SetEncryptionKey(value_string);
+                  params.MemberAsString(param_key.c_str(), &value_string));
+              if (param_key == "cpu_threads") {
+                int cpu_thread_num;
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseIntValue(value_string, &cpu_thread_num));
+                runtime_options_->SetCpuThreadNum(cpu_thread_num);
+              } else if (param_key == "use_mkldnn") {
+                bool pd_enable_mkldnn;
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseBoolValue(value_string, &pd_enable_mkldnn));
+                runtime_options_->SetPaddleMKLDNN(pd_enable_mkldnn);
+              } else if (param_key == "use_paddle_log") {
+                bool use_paddle_log;
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseBoolValue(value_string, &use_paddle_log));
+                runtime_options_->paddle_infer_option.enable_log_info =
+                    use_paddle_log;
+              } else if (param_key == "num_streams") {
+                int num_streams;
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseIntValue(value_string, &num_streams));
+                runtime_options_->openvino_option.num_streams = num_streams;
+              } else if (param_key == "is_clone") {
+                THROW_IF_BACKEND_MODEL_ERROR(
+                    ParseBoolValue(value_string, &is_clone_));
+              } else if (param_key == "use_ipu") {
+                // runtime_options_->UseIpu();
+              } else if (param_key == "encryption_key") {
+                runtime_options_->SetEncryptionKey(value_string);
+              }
            }
          }
-        }
+        }  // end 'name == "paddle_xpu"'
      }
    }
  }
@@ -422,7 +499,7 @@ TRITONSERVER_Error* ModelState::LoadModel(
      }
    }

-    // GPU
+// GPU
 #ifdef TRITON_ENABLE_GPU
    if ((instance_group_kind == TRITONSERVER_INSTANCEGROUPKIND_GPU) ||
        (instance_group_kind == TRITONSERVER_INSTANCEGROUPKIND_AUTO)) {
@@ -432,8 +509,9 @@ TRITONSERVER_Error* ModelState::LoadModel(
      runtime_options_->UseCpu();
    }
 #else
-    if (runtime_options_->device != fastdeploy::Device::IPU) {
-      // If Device is set to IPU, just skip CPU setting.
+    if ((runtime_options_->device != fastdeploy::Device::IPU) &&
+        (runtime_options_->device != fastdeploy::Device::KUNLUNXIN)) {
+      // If Device is set to IPU/XPU, just skip CPU setting.
      runtime_options_->UseCpu();
    }
 #endif  // TRITON_ENABLE_GPU
@@ -972,7 +1050,7 @@ void ModelInstanceState::ProcessRequests(TRITONBACKEND_Request** requests,
      SetInputTensors(total_batch_size, requests, request_count, &responses,
                      &collector, &cuda_copy));

-  // Wait for any in-flight input tensor copies to complete.
+// Wait for any in-flight input tensor copies to complete.
 #ifdef TRITON_ENABLE_GPU
  if (cuda_copy) {
    cudaStreamSynchronize(CudaStream());
@@ -1146,15 +1224,16 @@ TRITONSERVER_Error* ModelInstanceState::ReadOutputTensors(
    const uint32_t request_count,
    std::vector<TRITONBACKEND_Response*>* responses) {
  // r22.12
-  BackendOutputResponder responder(
-      requests, request_count, responses,
-      model_state_->TritonMemoryManager(), model_state_->MaxBatchSize() > 0,
-      model_state_->EnablePinnedOutput(), CudaStream());
-  // r21.10
  // BackendOutputResponder responder(
-  //     requests, request_count, responses, StateForModel()->MaxBatchSize(),
-  //     StateForModel()->TritonMemoryManager(),
-  //     StateForModel()->EnablePinnedOutput(), CudaStream());
+  //     requests, request_count, responses,
+  //     model_state_->TritonMemoryManager(), model_state_->MaxBatchSize() > 0,
+  //     model_state_->EnablePinnedOutput(), CudaStream());
+
+  // r21.10
+  BackendOutputResponder responder(
+      requests, request_count, responses, StateForModel()->MaxBatchSize(),
+      StateForModel()->TritonMemoryManager(),
+      StateForModel()->EnablePinnedOutput(), CudaStream());

  // Use to hold string output contents
  bool cuda_copy = false;