关键词：Triton

前言

Triton Inference Server是由NVIDIA提供的一个开源推理框架，旨在为AI算法模型提供高效的部署和推理能力，目前已经成为主流的模型部署方案。本文对Triton Inference Server做简要介绍，并以一个简单的线性模型为例子来实践部署。

内容摘要

• Triton Inference Server简介
• Docker构建Triton Inference Server环境
• Triton Inference Server部署一个线性模型

Triton Inference Server简介

Triton Inference Server是一款开源的推理服务框架，它的核心库基于C++编写的，旨在在生产环境中提供快速且可扩展的AI推理能力，具有以下优势

• 支持多种深度学习框架：包括PyTorch，Tensorflow，TensorRT，ONNX，OpenVINO等产出的模型文件
• 至此多种机器学习框架：支持对树模型的部署，包括XGBoost，LightGBM等
• 支持多种推理协议：支持HTTP，GRPC推理协议
• 服务端支持模型前后处理：提供后端API，支持将数据的前处理和模型推理的后处理在服务端实现
• 支持模型并发推理：支持多个模型或者同一模型的多个实例在同一系统上并行执行
• 支持动态批处理（Dynamic batching）：支持将一个或多个推理请求合并成一个批次，以最大化吞吐量
• 支持多模型的集成流水线：支持将多个模型进行连接组合，将其视作一个整体进行调度管理

Triton Inference Server架构如下图所示，从客户端请求开始，到模型调度处理，模型仓库管理和推理，响应返回，服务状态监控等。

Docker构建Triton Inference Server环境

Triton Inference Server官方推荐使用Docker来构建环境，在Docker镜像仓库下载已经构建好tritonserver镜像，根据cuda版本下载对应版本的镜像。笔者cuda版本为11.2，对应下载
nvcr.io/nvidia/tritonserver:21.02-py3镜像

docker pull nvcr.io/nvidia/tritonserver:21.02-py3

启动容器并进入，它以Python3作为服务端语言，tritonserver是推理服务的启动命令

root@de4977a12c14:/opt/tritonserver# which tritonserver
/opt/tritonserver/bin/tritonserver
root@de4977a12c14:/opt/tritonserver# which python3
/usr/bin/python3

模型部署还需要一些常用的Python算法包，为了后续测试，我们以该镜像作为源，安装PyTorch，transformers依赖到环境中形成一个新的镜像

# vim Dockerfile 
From nvcr.io/nvidia/tritonserver:21.02-py3
Maintainer xxx
COPY ./torch-1.12.1+cu113-cp38-cp38-linux_x86_64.whl /home
RUN pip install /home/torch-1.12.1+cu113-cp38-cp38-linux_x86_64.whl -i https://pypi.tuna.tsinghua.edu.cn/simple
RUN pip install transformers==4.28.0 -i https://pypi.tuna.tsinghua.edu.cn/simple

构建新镜像triton_server:v1

docker build -t triton_server:v1 .

使用Triton Inference Server部署一个线性模型

本节实践使用Triton Inference Server部署一个线性模型成为一个API服务，包含PyTorch线性模型训练，Triton模型仓库构建，模型推理配置构建，服务端代码构建，服务端启动，客户端服务调用这六个步骤。

（1）PyTorch线性模型训练

我们使用PyTorch的nn.Linear直接构建一个线性层，舍去训练部分，直接使用初始化的权重保存为一个模型文件

import torch
import torch.nn as nn

model = nn.Linear(3, 1)
nn.init.xavier_normal_(model.weight.data)

torch.save(model, "./pytorch_model.bin")

（2）Triton模型仓库构建

Triton Inference Server将模型统一放置在一个目录下进行管理，该目录下每个记录了模型的名称，配置，后端代码，模型文件等，我们创建该目录为model_repository，并且在其下创建一个线性模型目录，命名为linear，该命名是自定义的，但是在后续的配置文件和请求中要保持命名统一。

mkdir model_repository
cd model_repository
mkdir linear

（3）模型推理配置构建

进入linear目录，创建一个文件夹命名为1，代表是linear的第一个版本，将步骤1中训练得到PyTorch模型文件放置在1目录的linear目录下

mkdir 1
cd 1
mkdir linear
cd linear
mv /home/model/pytorch_model.bin ./

回到1目录下创建模型的推理配置文件命名为config.pbtxt，所有版本共用该配置，配置如下

# vim config.pbtxt
name: "linear"
backend: "python"

max_batch_size: 4
input [
  {
    name: "x"
    data_type: TYPE_FP32
    dims: [ 3 ]
  }
]
output [
  {
    name: "y"
    data_type: TYPE_FP32
    dims: [ 1 ]
  }
]

该配置定义了模型的输入和输出，在此做简要的说明

• name：模型名称，必须和步骤二中的文件名一致
• max_batch_size：一个批次下的最大大小，4代表一次请求最大推理4条样本
• input：模型的输入信息，array格式，其中每个输入是一个json格式
• input-name：一个输入的名称，该名称自定义，但是在服务端代码必须和其保持一致
• input-data_type：一个输入的数据类型，本例中采用32位浮点
• input-data_dims：一个输入的维度，代表一条样本的维度，若max_batch_size不为0，它和max_batch_size一起构成了最终的输入大小，本例中最终输入最大是[4, 3]的矩阵，若max_batch_size为0，则dims就是最终的维度

此处输入的x维度为3，和我们训练PyTorch模型时的nn.Linear(3, 1)保持对应。

（4）服务端代码构建

服务端代码负责读取模型，并且在其中实现数据的前处理，模型推理，后处理的逻辑。Triton Inference Server提供了服务端代码模板TritonPythonModel，只需要略微修改即可，本例的服务端代码如下

import os

os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:32'

os.environ['TRANSFORMERS_CACHE'] = os.path.dirname(os.path.abspath(__file__)) + "/work/"
os.environ['HF_MODULES_CACHE'] = os.path.dirname(os.path.abspath(__file__)) + "/work/"

import json

# triton_python_backend_utils is available in every Triton Python model. You
# need to use this module to create inference requests and responses. It also
# contains some utility functions for extracting information from model_config
# and converting Triton input/output types to numpy types.
import triton_python_backend_utils as pb_utils
import sys
import gc
import time
import logging
import torch
import numpy as np

gc.collect()
torch.cuda.empty_cache()

logging.basicConfig(format='%(asctime)s - %(filename)s[line:%(lineno)d] - %(levelname)s: %(message)s',
                    level=logging.INFO)


class TritonPythonModel:
    """Your Python model must use the same class name. Every Python model
    that is created must have "TritonPythonModel" as the class name.
    """

    def initialize(self, args):
        """`initialize` is called only once when the model is being loaded.
        Implementing `initialize` function is optional. This function allows
        the model to intialize any state associated with this model.

        Parameters
        ----------
        args : dict
          Both keys and values are strings. The dictionary keys and values are:
          * model_config: A JSON string containing the model configuration
          * model_instance_kind: A string containing model instance kind
          * model_instance_device_id: A string containing model instance device ID
          * model_repository: Model repository path
          * model_version: Model version
          * model_name: Model name
        """
        # You must parse model_config. JSON string is not parsed here
        self.model_config = json.loads(args['model_config'])
        output_y_config = pb_utils.get_output_config_by_name(self.model_config, "y")

        # Convert Triton types to numpy types
        self.output_response_dtype = pb_utils.triton_string_to_numpy(output_y_config['data_type'])
        # TODO 读取模型
        model_path = os.path.dirname(os.path.abspath(__file__)) + "/linear/pytorch_model.bin"
        self.model = torch.load(model_path)
        self.model = self.model.eval()
        logging.info("model init success")

    def execute(self, requests):
        """`execute` MUST be implemented in every Python model. `execute`
        function receives a list of pb_utils.InferenceRequest as the only
        argument. This function is called when an inference request is made
        for this model. Depending on the batching configuration (e.g. Dynamic
        Batching) used, `requests` may contain multiple requests. Every
        Python model, must create one pb_utils.InferenceResponse for every
        pb_utils.InferenceRequest in `requests`. If there is an error, you can
        set the error argument when creating a pb_utils.InferenceResponse

        Parameters
        ----------
        requests : list
          A list of pb_utils.InferenceRequest

        Returns
        -------
        list
          A list of pb_utils.InferenceResponse. The length of this list must
          be the same as `requests`

        """
        # output_response_dtype = self.output_response_dtype
        # output_history_dtype = self.output_history_dtype

        # output_dtype = self.output_dtype
        responses = []
        # Every Python backend must iterate over everyone of the requests
        # and create a pb_utils.InferenceResponse for each of them.
        for request in requests:
            # TODO 拿到输入
            x = pb_utils.get_input_tensor_by_name(request, "x").as_numpy()

            in_log_info = {
                "x": x,
            }
            logging.info(in_log_info)
            # TODO 推理结果
            y = self.model(torch.tensor(x).float())

            out_log_info = {
                "y": y
            }
            logging.info(out_log_info)
            y = y.detach().cpu().numpy()

            y_tensor = pb_utils.Tensor("y", y.astype(self.output_response_dtype))

            final_inference_response = pb_utils.InferenceResponse(
                output_tensors=[y_tensor])
            responses.append(final_inference_response)
            # Create InferenceResponse. You can set an error here in case
            # there was a problem with handling this inference request.
            # Below is an example of how you can set errors in inference
            # response:
            #
            # pb_utils.InferenceResponse(
            #    output_tensors=..., TritonError("An error occured"))

        # You should return a list of pb_utils.InferenceResponse. Length
        # of this list must match the length of `requests` list.
        return responses

    def finalize(self):
        """`finalize` is called only once when the model is being unloaded.
        Implementing `finalize` function is OPTIONAL. This function allows
        the model to perform any necessary clean ups before exit.
        """
        print('Cleaning up...')

其中initialize方法实现了模型的初始化，execute方法实现了模型的推理逻辑，并且包转了返回。我们将该Python脚本保存命名为model.py，放置在文件1下。

服务端工作准备完成，此时linear目录结构如下

~/model_repository/linear$ tree
.
├── 1
│   ├── linear
│   │   └── pytorch_model.bin
│   └── model.py
└── config.pbtxt

（5）服务端启动

接下来通过Docker来启动Triton Inference Server镜像，将模型仓库目录model_repository整个挂载到容器内部

docker run --rm \
-p18999:8000 -p18998:8001 -p18997:8002 \
-v /home/model_repository/:/models \
triton_server:v1 \
tritonserver --model-repository=/models

服务分别占用三个端口，将宿主机的端口和其进行映射，其中8000端口用于客户端请求推理结果。tritonserver是容器中的服务启动命令，--model-repository是参数，指定了模型仓库位置。

容器启动日志如下

=============================
== Triton Inference Server ==
=============================

NVIDIA Release 21.02 (build 20174689)

Copyright (c) 2018-2021, NVIDIA CORPORATION.  All rights reserved.

...

I0322 09:41:22.286694 1 server.cc:538] 
+--------------+---------+--------+
| Model        | Version | Status |
+--------------+---------+--------+
| linear       | 1       | READY  |
+--------------+---------+--------+

I0322 09:41:22.286937 1 tritonserver.cc:1642] 
+----------------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------+
| Option                           | Value                                                                                                                                              |
+----------------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------+
| server_id                        | triton                                                                                                                                             |
| server_version                   | 2.7.0                                                                                                                                              |
| server_extensions                | classification sequence model_repository schedule_policy model_configuration system_shared_memory cuda_shared_memory binary_tensor_data statistics |
| model_repository_path[0]         | /models                                                                                                                                            |
| model_control_mode               | MODE_NONE                                                                                                                                          |
| strict_model_config              | 1                                                                                                                                                  |
| pinned_memory_pool_byte_size     | 268435456                                                                                                                                          |
| min_supported_compute_capability | 6.0                                                                                                                                                |
| strict_readiness                 | 1                                                                                                                                                  |
| exit_timeout                     | 30                                                                                                                                                 |
+----------------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------+

I0322 09:41:22.289210 1 grpc_server.cc:3979] Started GRPCInferenceService at 0.0.0.0:8001
I0322 09:41:22.289923 1 http_server.cc:2717] Started HTTPService at 0.0.0.0:8000
I0322 09:41:22.331775 1 http_server.cc:2736] Started Metrics Service at 0.0.0.0:8002

提示模型linear版本1已经READY，则启动成功。

（6）客户端服务调用

本例中客户端采用HTTP请求的方式进行调用，使用Python的requests即可进行请求，代码如下

import numpy as np
import requests
import json


if __name__ == "__main__":
    model_name = "linear"
    model_version = "1"

    raw_data = {
        "inputs": [
            {
                "name": "x",
                "datatype": "FP32",
                "shape": [3, 3],
                "data": [[2.0, 3.0, 4.0], [1.1, 2.3, 10.3], [2.0, 3.0, 4.0]]
            }
        ],
        "outputs": [
            {
                "name": "y"
            }
        ]
    }

    url = "http://0.0.0.0:18999/v2/models/linear/infer"
    import requests
    import json

    response = requests.post(url=url,
                             data=json.dumps(raw_data, ensure_ascii=True),
                             headers={"Content_Type": "application/json"},
                             timeout=2000)
    res = json.loads(response.text)
    print(res)

在客户端我们创建了三条样本，形成[3, 3]的矩阵输入给Triton服务，请求url指定了要请求的模型，url中的模型名称linear必须和配置和文件夹命名一致。请求结果打印如下

{'model_name': 'linear', 'model_version': '1', 'outputs': [{'name': 'y', 'datatype': 'FP32', 'shape': [3, 1], 'data': [-0.21258652210235596, -1.3153640031814575, -0.21258652210235596]}]}

其中data就是模型推理的输出，维度和训练过程中的nn.Linear(3, 1)一致。同步的在服务端日志会打印出请求的数据和输出的数据

I0322 09:54:17.211954 1 grpc_server.cc:3979] Started GRPCInferenceService at 0.0.0.0:8001
I0322 09:54:17.212302 1 http_server.cc:2717] Started HTTPService at 0.0.0.0:8000
I0322 09:54:17.253904 1 http_server.cc:2736] Started Metrics Service at 0.0.0.0:8002
2024-03-22 09:54:24,068 - model.py[line:98] - INFO: {'x': array([[ 2. ,  3. ,  4. ],
       [ 1.1,  2.3, 10.3],
       [ 2. ,  3. ,  4. ]], dtype=float32)}
2024-03-22 09:54:24,073 - model.py[line:105] - INFO: {'y': tensor([[-0.2126],
        [-1.3154],
        [-0.2126]], grad_fn=<AddmmBackward0>)}

至此，基于Triton Inference Server搭建的一个简单的线性模型的预测推理服务完成。