Fd serving add docker images correlation and docs (#311)

* fd serving add dockerfile * fix enable_paddle_mkldnn * delete disable_paddle_mkldnn Co-authored-by: Jason <jiangjiajun@baidu.com>
2025-10-05 08:37:06 +08:00 · 2022-10-08 03:08:07 -05:00
parent 1efc0fa6b0
commit d57e997fa0
23 changed files with 673 additions and 112 deletions
--- a/docs/api/runtime/runtime_option.md
+++ b/docs/api/runtime/runtime_option.md
@@ -66,8 +66,7 @@ use_openvino_backend()
 使用OpenVINO后端进行推理，支持CPU, 支持Paddle/ONNX模型格式

 ```
-enable_paddle_mkldnn()
-disable_paddle_mkldnn()
+set_paddle_mkldnn()
 ```
 当使用Paddle Inference后端时，通过此开关开启或关闭CPU上MKLDNN推理加速，后端默认为开启

@@ -180,8 +179,7 @@ void UseOpenVINOBackend()
 使用OpenVINO后端进行推理，支持CPU, 支持Paddle/ONNX模型格式

 ```
-void EnablePaddleMKLDNN()
-void DisablePaddleMKLDNN()
+void SetPaddleMKLDNN(bool pd_mkldnn = true)
 ```
 当使用Paddle Inference后端时，通过此开关开启或关闭CPU上MKLDNN推理加速，后端默认为开启

--- a/docs/docs_en/api/runtime/runtime_option.md
+++ b/docs/docs_en/api/runtime/runtime_option.md
@@ -73,8 +73,7 @@ use_openvino_backend()
 Inference with OpenVINO backend (CPU supported, Paddle/ONNX model format supported)

 ```
-enable_paddle_mkldnn()
-disable_paddle_mkldnn()
+set_paddle_mkldnn()
 ```

 When using the Paddle Inference backend, this parameter determines whether the MKLDNN inference acceleration on the CPU is on or off. It is on by default.
@@ -204,8 +203,7 @@ void UseOpenVINOBackend()
 Inference with OpenVINO backend (CPU supported, Paddle/ONNX model format supported)

 ```
-void EnablePaddleMKLDNN()
-void DisablePaddleMKLDNN()
+void SetPaddleMKLDNN(bool pd_mkldnn = true)
 ```

 When using the Paddle Inference backend, this parameter determines whether the MKLDNN inference acceleration on the CPU is on or off. It is on by default.
--- a/examples/vision/detection/yolov5/README.md
+++ b/examples/vision/detection/yolov5/README.md
@@ -23,7 +23,7 @@

 - [Python部署](python)
 - [C++部署](cpp)
-
+- [服务化部署](serving)

 ## 版本说明

--- a/examples/vision/detection/yolov5/serving/README.md
+++ b/examples/vision/detection/yolov5/serving/README.md
@@ -1,19 +1,53 @@
-# YOLOv5 Serving部署示例
+# YOLOv5 服务化部署示例
+
+## 启动服务

 ```bash
-#下载yolov5模型文件和测试图片
+#下载yolov5模型文件
 wget https://bj.bcebos.com/paddlehub/fastdeploy/yolov5s.onnx
-wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg

 # 将模型放入 models/infer/1目录下, 并重命名为model.onnx
 mv yolov5s.onnx models/infer/1/

 # 拉取fastdeploy镜像
-docker pull xxx
+docker pull paddlepaddle/fastdeploy:0.3.0-gpu-cuda11.4-trt8.4-21.10

-# 启动镜像和服务
-docker run xx
+# 运行容器.容器名字为 fd_serving, 并挂载当前目录为容器的 /yolov5_serving 目录
+nvidia-docker run -it --net=host --name fd_serving -v `pwd`/:/yolov5_serving paddlepaddle/fastdeploy:0.3.0-gpu-cuda11.4-trt8.4-21.10  bash

-# 客户端请求
-python yolov5_grpc_client.py
+# 启动服务(不设置CUDA_VISIBLE_DEVICES环境变量，会拥有所有GPU卡的调度权限)
+CUDA_VISIBLE_DEVICES=0 fastdeployserver --model-repository=models --backend-config=python,shm-default-byte-size=10485760
 ```
+
+服务启动成功后， 会有以下输出:
+```
+......
+I0928 04:51:15.784517 206 grpc_server.cc:4117] Started GRPCInferenceService at 0.0.0.0:8001
+I0928 04:51:15.785177 206 http_server.cc:2815] Started HTTPService at 0.0.0.0:8000
+I0928 04:51:15.826578 206 http_server.cc:167] Started Metrics Service at 0.0.0.0:8002
+```
+
+
+## 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
+```
+#下载测试图片
+wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg
+
+#安装客户端依赖
+python3 -m pip install tritonclient\[all\]
+
+# 发送请求
+python3 yolov5_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:
+```
+output_name: detction_result
+{'boxes': [[268.48028564453125, 81.05305480957031, 298.69476318359375, 169.43902587890625], [104.73116302490234, 45.66197204589844, 127.58382415771484, 93.44938659667969], [378.9093933105469, 39.75013732910156, 395.6086120605469, 84.24342346191406], [158.552978515625, 80.36149597167969, 199.18576049804688, 168.18191528320312], [414.37530517578125, 90.94805908203125, 506.3218994140625, 280.40521240234375], [364.00341796875, 56.608917236328125, 381.97857666015625, 115.96823120117188], [351.7251281738281, 42.635345458984375, 366.9103088378906, 98.04837036132812], [505.8882751464844, 114.36674499511719, 593.1248779296875, 275.99530029296875], [327.7086181640625, 38.36369323730469, 346.84991455078125, 80.89302062988281], [583.493408203125, 114.53289794921875, 612.3546142578125, 175.87353515625], [186.4706573486328, 44.941375732421875, 199.6645050048828, 61.037628173828125], [169.6158905029297, 48.01460266113281, 178.1415557861328, 60.88859558105469], [25.81019401550293, 117.19969177246094, 59.88878631591797, 152.85012817382812], [352.1452941894531, 46.71272277832031, 381.9460754394531, 106.75212097167969], [1.875, 150.734375, 37.96875, 173.78125], [464.65728759765625, 15.901412963867188, 472.512939453125, 34.11640930175781], [64.625, 135.171875, 84.5, 154.40625], [57.8125, 151.234375, 103.0, 174.15625], [165.890625, 88.609375, 527.90625, 339.953125], [101.40625, 152.5625, 118.890625, 169.140625]], 'scores': [0.8965693116188049, 0.8695310950279236, 0.8684297800064087, 0.8429877758026123, 0.8358422517776489, 0.8151364326477051, 0.8089362382888794, 0.801361083984375, 0.7947245836257935, 0.7606497406959534, 0.6325908303260803, 0.6139386892318726, 0.5906146764755249, 0.505328893661499, 0.40457233786582947, 0.3460320234298706, 0.33283042907714844, 0.3325657248497009, 0.2594234347343445, 0.25389009714126587], 'label_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 24, 0, 24, 24, 33, 24], 'masks': [], 'contain_masks': False}
+```
+
+## 配置修改
+
+当前默认配置在GPU上运行ONNXRuntime引擎， 如果要在CPU或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](../../../../../serving/docs/zh_CN/model_configuration.md)
--- a/examples/vision/detection/yolov5/serving/models/runtime/config.pbtxt
+++ b/examples/vision/detection/yolov5/serving/models/runtime/config.pbtxt
@@ -5,7 +5,6 @@ max_batch_size: 16

 # Input configuration of the model
 input [
-  # 第一个输入
  {
    # input name
    name: "images"
@@ -36,3 +35,12 @@ instance_group [
    gpus: [0]
  }
 ]
+
+optimization {
+  execution_accelerators {
+  gpu_execution_accelerator : [ {
+    # use ONNXRuntime engine
+    name: "onnxruntime",
+    parameters { key: "cpu_threads" value: "2" }
+  }]
+}}
--- a/examples/vision/detection/yolov5/serving/yolov5_grpc_client.py
+++ b/examples/vision/detection/yolov5/serving/yolov5_grpc_client.py
@@ -65,11 +65,6 @@ class SyncGRPCTritonRunner:
        """
        infer_inputs = []
        for idx, data in enumerate(inputs):
-            print("len(data):", len(data))
-            print("name:", self._input_names[idx], "  shape:", data.shape,
-                  data.dtype)
-            #data = np.array([[x.encode('utf-8')] for x in data],
-            #                dtype=np.object_)
            infer_input = InferInput(self._input_names[idx], data.shape,
                                     "UINT8")
            infer_input.set_data_from_numpy(data)
@@ -106,7 +101,7 @@ if __name__ == "__main__":
        result = runner.Run([im, ])
        for name, values in result.items():
            print("output_name:", name)
-            for i in range(len(values)):
-                value = values[i][0]
+            for j in range(len(values)):
+                value = values[j][0]
                value = json.loads(value)
                print(value)
--- a/fastdeploy/pybind/runtime.cc
+++ b/fastdeploy/pybind/runtime.cc
@@ -28,8 +28,7 @@ void BindRuntime(pybind11::module& m) {
      .def("use_trt_backend", &RuntimeOption::UseTrtBackend)
      .def("use_openvino_backend", &RuntimeOption::UseOpenVINOBackend)
      .def("use_lite_backend", &RuntimeOption::UseLiteBackend)
-      .def("enable_paddle_mkldnn", &RuntimeOption::EnablePaddleMKLDNN)
-      .def("disable_paddle_mkldnn", &RuntimeOption::DisablePaddleMKLDNN)
+      .def("set_paddle_mkldnn", &RuntimeOption::SetPaddleMKLDNN)
      .def("enable_paddle_log_info", &RuntimeOption::EnablePaddleLogInfo)
      .def("disable_paddle_log_info", &RuntimeOption::DisablePaddleLogInfo)
      .def("set_paddle_mkldnn_cache_size",
--- a/fastdeploy/runtime.cc
+++ b/fastdeploy/runtime.cc
@@ -240,9 +240,9 @@ void RuntimeOption::UseLiteBackend() {
 #endif
 }

-void RuntimeOption::EnablePaddleMKLDNN() { pd_enable_mkldnn = true; }
-
-void RuntimeOption::DisablePaddleMKLDNN() { pd_enable_mkldnn = false; }
+void RuntimeOption::SetPaddleMKLDNN(bool pd_mkldnn) {
+  pd_enable_mkldnn = pd_mkldnn;
+}

 void RuntimeOption::DeletePaddleBackendPass(const std::string& pass_name) {
  pd_delete_pass_names.push_back(pass_name);
--- a/fastdeploy/runtime.h
+++ b/fastdeploy/runtime.h
@@ -119,11 +119,8 @@ struct FASTDEPLOY_DECL RuntimeOption {
  /// Set Paddle Lite as inference backend, only support arm cpu
  void UseLiteBackend();

-  /// Enable mkldnn while using Paddle Inference as inference backend
-  void EnablePaddleMKLDNN();
-
-  /// Disable mkldnn while using Paddle Inference as inference backend
-  void DisablePaddleMKLDNN();
+  // set mkldnn switch while using Paddle Inference as inference backend
+  void SetPaddleMKLDNN(bool pd_mkldnn = true);

  /**
   * @brief Delete pass by name while using Paddle Inference as inference backend, this can be called multiple times to delete a set of passes
--- a/python/fastdeploy/runtime.py
+++ b/python/fastdeploy/runtime.py
@@ -85,11 +85,8 @@ class RuntimeOption:
    def use_lite_backend(self):
        return self._option.use_lite_backend()

-    def enable_paddle_mkldnn(self):
-        return self._option.enable_paddle_mkldnn()
-
-    def disable_paddle_mkldnn(self):
-        return self._option.disable_paddle_mkldnn()
+    def set_paddle_mkldnn(self):
+        return self._option.set_paddle_mkldnn()

    def enable_paddle_log_info(self):
        return self._option.enable_paddle_log_info()
--- a/serving/Dockerfile
+++ b/serving/Dockerfile
@@ -0,0 +1,48 @@
+# 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.
+
+FROM nvcr.io/nvidia/tritonserver:21.10-py3 as full
+FROM nvcr.io/nvidia/tritonserver:21.10-py3-min
+
+COPY --from=full /opt/tritonserver/bin/tritonserver /opt/tritonserver/bin/fastdeployserver
+COPY --from=full /opt/tritonserver/lib /opt/tritonserver/lib
+COPY --from=full /opt/tritonserver/include /opt/tritonserver/include
+COPY --from=full /opt/tritonserver/backends/python /opt/tritonserver/backends/python
+
+COPY TensorRT-8.4.1.5 /opt/
+
+ENV TZ=Asia/Shanghai \
+    DEBIAN_FRONTEND=noninteractive \
+    DCGM_VERSION=2.2.9
+RUN apt-get update \
+    && apt-key del 7fa2af80 \
+    && wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-keyring_1.0-1_all.deb \
+    && dpkg -i cuda-keyring_1.0-1_all.deb \
+    && apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/7fa2af80.pub \
+    && apt-get update && apt-get install -y --no-install-recommends datacenter-gpu-manager=1:2.2.9
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends libre2-5 libb64-0d python3 python3-pip libarchive-dev \
+    && python3 -m pip install -U pip \
+    && python3 -m pip install paddlepaddle-gpu paddlenlp faster_tokenizer
+
+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-0.0.3 /opt/fastdeploy/
+
+ENV LD_LIBRARY_PATH="/opt/TensorRT-8.4.1.5/lib/:/opt/fastdeploy/lib:/opt/fastdeploy/third_libs/install/onnxruntime/lib:/opt/fastdeploy/third_libs/install/paddle2onnx/lib:/opt/fastdeploy/third_libs/install/tensorrt/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/openvino/runtime/lib:$LD_LIBRARY_PATH"
+ENV PATH="/opt/tritonserver/bin:$PATH"
--- a/serving/README.md
+++ b/serving/README.md
@@ -0,0 +1 @@
+README_CN.md
--- a/serving/README_CN.md
+++ b/serving/README_CN.md
@@ -0,0 +1,19 @@
+简体中文 | [English](README_EN.md)
+
+# FastDeploy 服务化部署
+
+## 简介
+
+FastDeploy基于[Triton Inference Server](https://github.com/triton-inference-server/server)搭建了端到端的服务化部署。底层后端使用FastDeploy高性能Runtime模块，并串联FastDeploy前后处理模块实现端到端的服务化部署。具有快速部署、使用简单、性能卓越的特性。
+
+## 端到端部署示例
+
+- [YOLOV5 检测任务](../examples/vision/detection/yolov5/README.md)
+- [OCR ]()
+- [Erinie3.0 文本分类任务]()
+- [UIE ]()
+- [Speech ]()
+
+## 高阶文档
+- [模型仓库](docs/zh_CN/model_repository.md)
+- [模型配置](docs/zh_CN/model_configuration.md)
--- a/serving/README_EN.md
+++ b/serving/README_EN.md
@@ -0,0 +1 @@
+English | [简体中文](README_CN.md)
--- a/serving/docs/zh_CN/model_configuration.md
+++ b/serving/docs/zh_CN/model_configuration.md
@@ -0,0 +1,168 @@
+# 模型配置
+模型存储库中的每个模型都必须包含一个模型配置，该配置提供了关于模型的必要和可选信息。这些配置信息一般写在 *config.pbtxt* 文件中，[ModelConfig protobuf](https://github.com/triton-inference-server/common/blob/main/protobuf/model_config.proto)格式。
+
+## 模型通用最小配置
+详细的模型通用配置请看官网文档: [model_configuration](https://github.com/triton-inference-server/server/blob/main/docs/model_configuration.md).Triton的最小模型配置必须包括: *platform* 或 *backend* 属性、*max_batch_size* 属性和模型的输入输出.
+
+例如一个Paddle模型，有两个输入*input0* 和 *input1*，一个输出*output0*，输入输出都是float32类型的tensor，最大batch为8.则最小的配置如下:
+
+```
+  backend: "fastdeploy"
+  max_batch_size: 8
+  input [
+    {
+      name: "input0"
+      data_type: TYPE_FP32
+      dims: [ 16 ]
+    },
+    {
+      name: "input1"
+      data_type: TYPE_FP32
+      dims: [ 16 ]
+    }
+  ]
+  output [
+    {
+      name: "output0"
+      data_type: TYPE_FP32
+      dims: [ 16 ]
+    }
+  ]
+```
+
+## CPU、GPU和实例个数配置
+
+通过*instance_group*属性可以配置服务使用哪种硬件资源，分别部署多少个模型推理实例。
+
+CPU部署例子：
+```
+  instance_group [
+    {
+      # 创建两个CPU实例
+      count: 2
+      # 使用CPU部署  
+      kind: KIND_CPU
+    }
+  ]
+```
+
+在*GPU 0*上部署2个实例，在*GPU1*和*GPU*上分别部署1个实例
+
+```
+  instance_group [
+    {
+      # 创建两个GPU实例
+      count: 2
+      # 使用GPU推理
+      kind: KIND_GPU
+      # 部署在GPU卡0上
+      gpus: [ 0 ]
+    },
+    {
+      count: 1
+      kind: KIND_GPU
+      # 在GPU卡1、2都部署
+      gpus: [ 1, 2 ]
+    }
+  ]
+```
+
+### Name, Platform and Backend
+模型配置中 *name* 属性是可选的。如果模型没有在配置中指定，则使用模型的目录名；如果指定了该属性，它必须要跟模型的目录名一致。
+
+使用 *fastdeploy backend*，没有*platform*属性可以配置，必须配置*backend*属性为*fastdeploy*。
+
+```
+backend: "fastdeploy"
+```
+
+### FastDeploy Backend配置
+
+FastDeploy后端目前支持*cpu*和*gpu*推理，*cpu*上支持*paddle*、*onnxruntime*和*openvino*三个推理引擎，*gpu*上支持*paddle*、*onnxruntime*和*tensorrt*三个引擎。
+
+
+#### 配置使用Paddle引擎
+除去配置 *Instance Groups*，决定模型运行在CPU还是GPU上。Paddle引擎中，还可以进行如下配置:
+
+```
+optimization {
+  execution_accelerators {
+    # CPU推理配置， 配合KIND_CPU使用
+    cpu_execution_accelerator : [
+      {
+        name : "paddle"
+        # 设置推理并行计算线程数为4
+        parameters { key: "cpu_threads" value: "4" }
+        # 开启mkldnn加速，设置为0关闭mkldnn
+        parameters { key: "use_mkldnn" value: "1" }
+      }
+    ],
+    # GPU推理配置， 配合KIND_GPU使用
+    gpu_execution_accelerator : [
+      {
+        name : "paddle"
+        # 设置推理并行计算线程数为4
+        parameters { key: "cpu_threads" value: "4" }
+        # 开启mkldnn加速，设置为0关闭mkldnn
+        parameters { key: "use_mkldnn" value: "1" }
+      }
+    ]
+  }
+}
+```
+
+### 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:
+
+```
+optimization {
+  execution_accelerators {
+    cpu_execution_accelerator : [
+      {
+        name : "onnxruntime"
+        # 设置推理并行计算线程数为4
+        parameters { key: "cpu_threads" value: "4" }
+      }
+    ],
+    gpu_execution_accelerator : [
+      {
+        name : "onnxruntime"
+      }
+    ]
+  }
+}
+```
+
+### 配置使用OpenVINO引擎
+OpenVINO引擎只支持CPU推理，配置如下:
+
+```
+optimization {
+  execution_accelerators {
+    cpu_execution_accelerator : [
+      {
+        name : "openvino"
+        # 设置推理并行计算线程数为4
+        parameters { key: "cpu_threads" value: "4" }
+      }
+    ]
+  }
+}
+```
+
+### 配置使用TensorRT引擎
+TensorRT引擎只支持GPU推理，配置如下:
+
+```
+optimization {
+  execution_accelerators {
+    gpu_execution_accelerator : [
+      {
+        name : "tensorrt"
+        # 使用TensorRT的FP16推理,其他可选项为: trt_fp32、trt_int8
+        parameters { key: "precision" value: "trt_fp16" }
+      }
+    ]
+  }
+}
+```
--- a/serving/docs/zh_CN/model_repository.md
+++ b/serving/docs/zh_CN/model_repository.md
@@ -0,0 +1,78 @@
+# 模型仓库(Model Repository)
+
+FastDeploy启动服务时指定模型仓库中一个或多个模型部署服务。当服务运行时，可以用[Model Management](https://github.com/triton-inference-server/server/blob/main/docs/model_management.md)中描述的方式修改服务中的模型。
+从服务器启动时指定的一个或多个模型存储库中为模型提供服务
+
+## 仓库结构
+模型仓库路径通过FastDeploy启动时的*--model-repository*选项指定，可以多次指定*--model-repository*选项来加载多个仓库。例如:
+
+```
+$ fastdeploy --model-repository=<model-repository-path>
+```
+
+模型仓库的结构必须按以下的格式创建:
+```
+  <model-repository-path>/
+    <model-name>/
+      [config.pbtxt]
+      [<output-labels-file> ...]
+      <version>/
+        <model-definition-file>
+      <version>/
+        <model-definition-file>
+      ...
+    <model-name>/
+      [config.pbtxt]
+      [<output-labels-file> ...]
+      <version>/
+        <model-definition-file>
+      <version>/
+        <model-definition-file>
+      ...
+    ...
+```
+在最顶层`<model-repository-path>`模型仓库目录下，必须有0个或多个`<model-name>`模型名字的子目录。每个`<model-name>`模型名字子目录包含部署模型相应的信息，多个表示模型版本的数字子目录和一个描述模型配置的*config.pbtxt*文件。
+
+Paddle模型存在版本号子目录中，必须为`model.pdmodel`文件和`model.pdiparams`文件。
+
+## 模型版本
+每个模型在仓库中可以有一个或多个可用的版本，模型目录中以数字命名的子目录就是对应的版本，数字即版本号。没有以数字命名的子目录，或以*0*开头的子目录都会被忽略。模型配置文件中可以指定[版本策略](https://github.com/triton-inference-server/server/blob/main/docs/model_configuration.md#version-policy)，控制Triton启动模型目录中的哪个版本。
+
+## 模型仓库示例
+部署Paddle模型时需要的模型必须是2.0版本以上导出的推理模型，模型包含`model.pdmodel`和`model.pdiparams`两个文件放在版本目录中。
+
+部署Paddle模型的最小模型仓库目录示例:
+```
+  <model-repository-path>/
+    <model-name>/
+      config.pbtxt
+      1/
+        model.pdmodel
+        model.pdiparams
+
+  # 真实例子:
+  models
+  └── ResNet50
+      ├── 1
+      │   ├── model.pdiparams
+      │   └── model.pdmodel
+      └── config.pbtxt
+```
+
+部署ONNX模型，必须要在版本目录中包含`model.onnx`名字的模型。
+
+部署ONNX模型的最小模型仓库目录示例:
+```
+  <model-repository-path>/
+    <model-name>/
+      config.pbtxt
+      1/
+        model.onnx
+
+  # 真实例子:
+  models
+  └── ResNet50
+      ├── 1
+      │   ├── model.onnx
+      └── config.pbtxt
+```
--- a/serving/scripts/build.sh
+++ b/serving/scripts/build.sh
@@ -0,0 +1,18 @@
+#!/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.
+
+sh build_fd_vison.sh
+sh build_fd_runtime.sh
+sh build_fd_backend.sh
--- a/serving/scripts/build_fd_backend.sh
+++ b/serving/scripts/build_fd_backend.sh
@@ -0,0 +1,29 @@
+#!/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.
+if [ ! -d "./cmake-3.18.6-Linux-x86_64/" ]; then
+    wget https://github.com/Kitware/CMake/releases/download/v3.18.6/cmake-3.18.6-Linux-x86_64.tar.gz
+    tar -zxvf cmake-3.18.6-Linux-x86_64.tar.gz
+    rm -rf cmake-3.18.6-Linux-x86_64.tar.gz
+fi
+
+docker run -it --rm --name build_fd_backend \
+           -v`pwd`:/workspace/fastdeploy \
+           nvcr.io/nvidia/tritonserver:21.10-py3 \
+           bash -c \
+           'cd /workspace/fastdeploy/serving;
+            rm -rf build; mkdir build; cd build;
+            apt-get update; apt-get install -y --no-install-recommends rapidjson-dev;
+            export PATH=/workspace/fastdeploy/cmake-3.18.6-Linux-x86_64/bin:$PATH;
+            cmake .. -DFASTDEPLOY_DIR=/workspace/fastdeploy/build/fastdeploy-0.0.3 -DTRITON_COMMON_REPO_TAG=r21.10 -DTRITON_CORE_REPO_TAG=r21.10 -DTRITON_BACKEND_REPO_TAG=r21.10; make -j`nproc`'
--- a/serving/scripts/build_fd_runtime.sh
+++ b/serving/scripts/build_fd_runtime.sh
@@ -0,0 +1,39 @@
+#!/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.
+if [ ! -d "./cmake-3.18.6-Linux-x86_64/" ]; then
+    wget https://github.com/Kitware/CMake/releases/download/v3.18.6/cmake-3.18.6-Linux-x86_64.tar.gz
+    tar -zxvf cmake-3.18.6-Linux-x86_64.tar.gz
+    rm -rf cmake-3.18.6-Linux-x86_64.tar.gz
+fi
+
+if [ ! -d "./TensorRT-8.4.1.5/" ]; then
+    wget https://fastdeploy.bj.bcebos.com/third_libs/TensorRT-8.4.1.5.Linux.x86_64-gnu.cuda-11.6.cudnn8.4.tar.gz
+    tar -zxvf TensorRT-8.4.1.5.Linux.x86_64-gnu.cuda-11.6.cudnn8.4.tar.gz
+    rm -rf TensorRT-8.4.1.5.Linux.x86_64-gnu.cuda-11.6.cudnn8.4.tar.gz
+fi
+
+docker run -it --rm --name build_fd_runtime \
+           -v`pwd`:/workspace/fastdeploy \
+           nvcr.io/nvidia/tritonserver:21.10-py3-min \
+           bash -c \
+           'cd /workspace/fastdeploy;
+            rm -rf build; mkdir build; cd build;
+            apt-get update;
+            apt-get install -y --no-install-recommends python3-dev python3-pip;
+            ln -s /usr/bin/python3 /usr/bin/python;
+            export PATH=/workspace/fastdeploy/cmake-3.18.6-Linux-x86_64/bin:$PATH;
+            cmake .. -DENABLE_TRT_BACKEND=ON -DCMAKE_INSTALL_PREFIX=${PWD}/fastdeploy-0.0.3 -DWITH_GPU=ON -DTRT_DIRECTORY=${PWD}/../TensorRT-8.4.1.5/ -DENABLE_PADDLE_BACKEND=ON -DENABLE_ORT_BACKEND=ON -DENABLE_OPENVINO_BACKEND=ON -DENABLE_VISION=OFF -DBUILD_FASTDEPLOY_PYTHON=OFF -DENABLE_PADDLE_FRONTEND=ON -DENABLE_TEXT=OFF -DLIBRARY_NAME=fastdeploy_runtime;
+            make -j`nproc`;
+            make install'
--- a/serving/scripts/build_fd_vison.sh
+++ b/serving/scripts/build_fd_vison.sh
@@ -0,0 +1,36 @@
+#!/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.
+if [ ! -d "./cmake-3.18.6-Linux-x86_64/" ]; then
+    wget https://github.com/Kitware/CMake/releases/download/v3.18.6/cmake-3.18.6-Linux-x86_64.tar.gz
+    tar -zxvf cmake-3.18.6-Linux-x86_64.tar.gz
+    rm -rf cmake-3.18.6-Linux-x86_64.tar.gz
+fi
+
+docker run -it --rm --name build_fd_vison \
+           -v`pwd`:/workspace/fastdeploy \
+           nvcr.io/nvidia/tritonserver:21.10-py3-min \
+           bash -c \
+           'cd /workspace/fastdeploy/python;
+            rm -rf .setuptools-cmake-build dist;
+            apt-get update;
+            apt-get install -y --no-install-recommends patchelf python3-dev python3-pip;
+            ln -s /usr/bin/python3 /usr/bin/python;
+            export PATH=/workspace/fastdeploy/cmake-3.18.6-Linux-x86_64/bin:$PATH;
+            export WITH_GPU=ON;
+            export ENABLE_ORT_BACKEND=OFF;
+            export ENABLE_VISION=ON;
+            export ENABLE_TEXT=ON;
+            python setup.py build;
+            python setup.py bdist_wheel'
--- a/serving/src/fastdeploy_backend_utils.cc
+++ b/serving/src/fastdeploy_backend_utils.cc
@@ -123,6 +123,27 @@ std::string FDTypeToModelConfigDataType(fastdeploy::FDDataType data_type) {
  return "TYPE_INVALID";
 }

+TRITONSERVER_Error* FDParseShape(triton::common::TritonJson::Value& io,
+                                 const std::string& name,
+                                 std::vector<int32_t>* shape) {
+  std::string shape_string;
+  RETURN_IF_ERROR(io.MemberAsString(name.c_str(), &shape_string));
+
+  std::vector<std::string> str_shapes;
+  std::istringstream in(shape_string);
+  std::copy(std::istream_iterator<std::string>(in),
+            std::istream_iterator<std::string>(),
+            std::back_inserter(str_shapes));
+
+  std::transform(str_shapes.cbegin(), str_shapes.cend(),
+                 std::back_inserter(*shape),
+                 [](const std::string& str) -> int32_t {
+                   return static_cast<int32_t>(std::stoll(str));
+                 });
+
+  return nullptr;  // success
+}
+
 }  // namespace fastdeploy_runtime
 }  // namespace backend
 }  // namespace triton
--- a/serving/src/fastdeploy_backend_utils.h
+++ b/serving/src/fastdeploy_backend_utils.h
@@ -33,6 +33,7 @@
 #include <vector>

 #include "fastdeploy/core/fd_type.h"
+#include "triton/backend/backend_common.h"
 #include "triton/core/tritonserver.h"

 namespace triton {
@@ -67,6 +68,10 @@ fastdeploy::FDDataType ModelConfigDataTypeToFDType(

 std::string FDTypeToModelConfigDataType(fastdeploy::FDDataType data_type);

+TRITONSERVER_Error* FDParseShape(triton::common::TritonJson::Value& io,
+                                 const std::string& name,
+                                 std::vector<int32_t>* shape);
+
 }  // namespace fastdeploy_runtime
 }  // namespace backend
 }  // namespace triton
--- a/serving/src/fastdeploy_runtime.cc
+++ b/serving/src/fastdeploy_runtime.cc
@@ -26,6 +26,7 @@

 #include <stdint.h>

+#include <algorithm>
 #include <mutex>
 #include <vector>

@@ -169,83 +170,154 @@ ModelState::ModelState(TRITONBACKEND_Model* triton_model)
  // instance when creating that instance's runtime.
  runtime_options_.reset(new fastdeploy::RuntimeOption());

+  triton::common::TritonJson::Value optimization;
+  if (not ModelConfig().Find("optimization", &optimization)) {
+    return;
+  }
+
+  triton::common::TritonJson::Value eas;
+  if (not optimization.Find("execution_accelerators", &eas)) {
+    return;
+  }
+
+  // CPU execution providers
  {
-    triton::common::TritonJson::Value optimization;
-    if (ModelConfig().Find("optimization", &optimization)) {
-      triton::common::TritonJson::Value backend;
-      if (optimization.Find("onnxruntime", &backend)) {
-        runtime_options_->UseOrtBackend();
-        std::vector<std::string> param_keys;
-        THROW_IF_BACKEND_MODEL_ERROR(backend.Members(&param_keys));
-        for (const auto& param_key : param_keys) {
-          std::string value_string;
-          if (param_key == "graph_level") {
+    triton::common::TritonJson::Value cpu_eas;
+    if (eas.Find("cpu_execution_accelerator", &cpu_eas)) {
+      for (size_t idx = 0; idx < cpu_eas.ArraySize(); idx++) {
+        triton::common::TritonJson::Value ea;
+        THROW_IF_BACKEND_MODEL_ERROR(cpu_eas.IndexAsObject(idx, &ea));
+        std::string name;
+        THROW_IF_BACKEND_MODEL_ERROR(ea.MemberAsString("name", &name));
+        if (name == "onnxruntime") {
+          runtime_options_->UseOrtBackend();
+        } else if (name == "paddle") {
+          runtime_options_->UsePaddleBackend();
+        } else if (name == "openvino") {
+          runtime_options_->UseOpenVINOBackend();
+        } else if (name != "") {
+          TRITONSERVER_Error* error = TRITONSERVER_ErrorNew(
+              TRITONSERVER_ERROR_INVALID_ARG,
+              std::string("unknown cpu_execution_accelerator name '" + name +
+                          "' is provided. Available choices are [onnxruntime, "
+                          "paddle, openvino]")
+                  .c_str());
+          THROW_IF_BACKEND_MODEL_ERROR(error);
+        }
+
+        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(
-                backend.MemberAsString(param_key.c_str(), &value_string));
-            THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
-                value_string, &runtime_options_->ort_graph_opt_level));
-          } else if (param_key == "inter_op_num_threads") {
-            THROW_IF_BACKEND_MODEL_ERROR(
-                backend.MemberAsString(param_key.c_str(), &value_string));
-            THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
-                value_string, &runtime_options_->ort_inter_op_num_threads));
-          } else if (param_key == "execution_mode") {
-            THROW_IF_BACKEND_MODEL_ERROR(
-                backend.MemberAsString(param_key.c_str(), &value_string));
-            THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
-                value_string, &runtime_options_->ort_execution_mode));
+                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 == "graph_level") {
+              //   THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
+              //       value_string, &runtime_options_->ort_graph_opt_level));
+              // } else if (param_key == "inter_op_num_threads") {
+              //   THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
+              //       value_string,
+              //       &runtime_options_->ort_inter_op_num_threads));
+              // } else if (param_key == "execution_mode") {
+              //   THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
+              //       value_string, &runtime_options_->ort_execution_mode));
+              // } else if (param_key == "capacity") {
+              //     THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
+              //     value_string, &runtime_options_->pd_mkldnn_cache_size));
+            } 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 (optimization.Find("tensorrt", &backend)) {
-        runtime_options_->UseTrtBackend();
-        std::vector<std::string> param_keys;
-        THROW_IF_BACKEND_MODEL_ERROR(backend.Members(&param_keys));
-        for (const auto& param_key : param_keys) {
-          std::string value_string;
-          if (param_key == "cpu_threads") {
-            THROW_IF_BACKEND_MODEL_ERROR(
-                backend.MemberAsString(param_key.c_str(), &value_string));
-            THROW_IF_BACKEND_MODEL_ERROR(
-                ParseIntValue(value_string, &runtime_options_->cpu_thread_num));
-          }
-          // TODO(liqi): add tensorrt
+      }
+    }
+  }
+
+  // GPU execution providers
+  {
+    triton::common::TritonJson::Value gpu_eas;
+    if (eas.Find("gpu_execution_accelerator", &gpu_eas)) {
+      for (size_t idx = 0; idx < gpu_eas.ArraySize(); idx++) {
+        triton::common::TritonJson::Value ea;
+        THROW_IF_BACKEND_MODEL_ERROR(gpu_eas.IndexAsObject(idx, &ea));
+        std::string name;
+        THROW_IF_BACKEND_MODEL_ERROR(ea.MemberAsString("name", &name));
+
+        if (name == "onnxruntime") {
+          runtime_options_->UseOrtBackend();
+        } else if (name == "paddle") {
+          runtime_options_->UsePaddleBackend();
+        } else if (name == "tensorrt") {
+          runtime_options_->UseTrtBackend();
        }
-      } else if (optimization.Find("paddle", &backend)) {
-        runtime_options_->UsePaddleBackend();
-        std::vector<std::string> param_keys;
-        THROW_IF_BACKEND_MODEL_ERROR(backend.Members(&param_keys));
-        for (const auto& param_key : param_keys) {
-          std::string value_string;
-          if (param_key == "cpu_threads") {
-            THROW_IF_BACKEND_MODEL_ERROR(
-                backend.MemberAsString(param_key.c_str(), &value_string));
-            THROW_IF_BACKEND_MODEL_ERROR(
-                ParseIntValue(value_string, &runtime_options_->cpu_thread_num));
-          } else if (param_key == "capacity") {
-            THROW_IF_BACKEND_MODEL_ERROR(
-                backend.MemberAsString(param_key.c_str(), &value_string));
-            THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
-                value_string, &runtime_options_->pd_mkldnn_cache_size));
-          } else if (param_key == "use_mkldnn") {
-            THROW_IF_BACKEND_MODEL_ERROR(
-                backend.MemberAsString(param_key.c_str(), &value_string));
-            THROW_IF_BACKEND_MODEL_ERROR(ParseBoolValue(
-                value_string, &runtime_options_->pd_enable_mkldnn));
+        if (name == "min_shape" or name == "max_shape" or name == "opt_shape") {
+          triton::common::TritonJson::Value params;
+          if (ea.Find("parameters", &params)) {
+            std::vector<std::string> input_names;
+            THROW_IF_BACKEND_MODEL_ERROR(params.Members(&input_names));
+            for (const auto& input_name : input_names) {
+              std::vector<int32_t> shape;
+              FDParseShape(params, input_name, &shape);
+              if (name == "min_shape") {
+                runtime_options_->trt_min_shape[input_name] = shape;
+              } else if (name == "max_shape") {
+                runtime_options_->trt_max_shape[input_name] = shape;
+              } else {
+                runtime_options_->trt_opt_shape[input_name] = shape;
+              }
+            }
          }
-        }
-      } else if (optimization.Find("openvino", &backend)) {
-        runtime_options_->UseOpenVINOBackend();
-        std::vector<std::string> param_keys;
-        THROW_IF_BACKEND_MODEL_ERROR(backend.Members(&param_keys));
-        for (const auto& param_key : param_keys) {
-          std::string value_string;
-          if (param_key == "cpu_threads") {
-            THROW_IF_BACKEND_MODEL_ERROR(
-                backend.MemberAsString(param_key.c_str(), &value_string));
-            THROW_IF_BACKEND_MODEL_ERROR(
-                ParseIntValue(value_string, &runtime_options_->cpu_thread_num));
+        } else {
+          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 == "graph_level") {
+              //   THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
+              //       value_string, &runtime_options_->ort_graph_opt_level));
+              // } else if (param_key == "inter_op_num_threads") {
+              //   THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
+              //       value_string,
+              //       &runtime_options_->ort_inter_op_num_threads));
+              // } else if (param_key == "execution_mode") {
+              //   THROW_IF_BACKEND_MODEL_ERROR(ParseIntValue(
+              //       value_string, &runtime_options_->ort_execution_mode));
+              // }
+              if (param_key == "precision") {
+                std::transform(value_string.begin(), value_string.end(),
+                               value_string.begin(), ::tolower);
+                if (value_string == "trt_fp16") {
+                  runtime_options_->EnableTrtFP16();
+                } else if (value_string == "trt_int8") {
+                  // TODO(liqi): use EnableTrtINT8
+                  runtime_options_->trt_enable_int8 = true;
+                }
+                // } else if( param_key == "max_batch_size") {
+                //   THROW_IF_BACKEND_MODEL_ERROR(ParseUnsignedLongLongValue(
+                //       value_string, &runtime_options_->trt_max_batch_size));
+                // } else if( param_key == "workspace_size") {
+                //   THROW_IF_BACKEND_MODEL_ERROR(ParseUnsignedLongLongValue(
+                //       value_string,
+                //       &runtime_options_->trt_max_workspace_size));
+              } else if (param_key == "cache_file") {
+                runtime_options_->SetTrtCacheFile(value_string);
+              }
+            }
          }
-          // TODO(liqi): add openvino
        }
      }
    }
@@ -285,11 +357,11 @@ TRITONSERVER_Error* ModelState::LoadModel(
                        "not provided.'")
                .c_str());
      }
-      runtime_options_->model_format = fastdeploy::Frontend::PADDLE;
+      runtime_options_->model_format = fastdeploy::ModelFormat::PADDLE;
      runtime_options_->model_file = *model_path;
      runtime_options_->params_file = *params_path;
    } else {
-      runtime_options_->model_format = fastdeploy::Frontend::ONNX;
+      runtime_options_->model_format = fastdeploy::ModelFormat::ONNX;
      runtime_options_->model_file = *model_path;
    }
  }