Environment

根据Triton 环境对应表 ，Colab 环境缺少 tensorrt-8.6.1，cudnn9-cuda-12，triton-server 版本应该选择 r23.10。

apt update && apt install -y --no-install-recommends \
    ca-certificates autoconf automake build-essential docker.io git libre2-dev libssl-dev libtool libboost-dev \
    libcurl4-openssl-dev libb64-dev patchelf python3-dev python3-pip python3-setuptools rapidjson-dev scons \
    software-properties-common unzip wget zlib1g-dev libarchive-dev pkg-config uuid-dev libnuma-dev curl \
    libboost-all-dev datacenter-gpu-manager cudnn9-cuda-12

pip3 install --upgrade pip && pip3 install --upgrade wheel setuptools tritonclient[all] diffusers>=0.27.0 transformers accelerate safetensors optimum["onnxruntime"]

upgrade boost

wget https://boostorg.jfrog.io/artifactory/main/release/1.84.0/source/boost_1_84_0.tar.gz
tar -zxvf boost_1_84_0.tar.gz 
cd boost_1_84_0
chmod -R 777 .
./bootstrap.sh --with-libraries=all --with-toolset=gcc
./b2 -j20 toolset=gcc
./b2 install 

install libarchive

wget https://github.com/libarchive/libarchive/releases/download/v3.6.2/libarchive-3.6.2.tar.gz
tar -zxvf libarchive-3.6.2.tar.gz 
cd libarchive-3.6.2
./configure
make
sudo make install

install tensorrt-8.6.1

# 方法一
wget https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/secure/8.6.1/tars/TensorRT-8.6.1.6.Linux.x86_64-gnu.cuda-12.0.tar.gz
tar -xvf TensorRT-8.6.1.6.Linux.x86_64-gnu.cuda-12.0.tar.gz 
sudo mv TensorRT-8.6.1.6/ /usr/local/
vim ~/.bashrc 
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/TensorRT-8.6.1.6/lib
source ~/.bashrc 

# 方法二
wget https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/secure/8.6.1/local_repos/nv-tensorrt-local-repo-ubuntu2204-8.6.1-cuda-12.0_1.0-1_amd64.deb
sudo cp /var/nv-tensorrt-local-repo-ubuntu2204-8.6.1-cuda-12.0/nv-tensorrt-local-42B2FC56-keyring.gpg /usr/share/keyrings/
sudo dpkg -i nv-tensorrt-local-repo-ubuntu2204-8.6.1-cuda-12.0_1.0-1_amd64.deb

Building Triton server

编译 Triton

git clone -b r23.10 https://github.com/triton-inference-server/server.git

# enable-all 编译失败了，原因可能为编译某个 backend 导致的，解决方法未知
./build.py -v --no-container-build --build-dir=`pwd`/build --enable-all

# 自定义参数且只编译 python 后端，成功
./build.py -v --no-container-build --build-dir=$(pwd)/build --enable-logging --enable-stats --enable-tracing --enable-gpu --endpoint http --endpoint grpc  --backend python --extra-core-cmake-arg j=0

设置软链接

ln -s /content/server/build/opt/tritonserver /opt/tritonserver

Deploying HuggingFace models

克隆 python_backend，因为我们要使用 python_backend 中的 triton_python_backend_utils

git clone https://github.com/triton-inference-server/python_backend.git -b r23.02
cd python_backend

配置模型库
部署非常能打的文生图大模型 playground-v2.5

mkdir -p models/playground-v2.5/1/
# 配置文件
touch models/playground-v2.5/config.pbtxt
# 模型文件
touch models/playground-v2.5/1/model.py
# 客户端文件
touch models/playground-v2.5/client.py

config.pbtxt

name: "playground-v2.5"
backend: "python"
max_batch_size: 0
input [
  {
    name: "prompt"
    data_type: TYPE_STRING
    dims: [-1, -1]
  }
]
output [
  {
    name: "generated_image"
    data_type: TYPE_FP32
    dims: [-1, -1, -1]
  }
]
instance_group [
  {
    kind: KIND_GPU
  }
]

model.py

import numpy as np
import triton_python_backend_utils as pb_utils
from transformers import ViTImageProcessor, ViTModel
from diffusers import DiffusionPipeline
import torch
import time
import os
import shutil
import json
import numpy as np

class TritonPythonModel:
    def initialize(self, args):
        self.model = DiffusionPipeline.from_pretrained(
            "playgroundai/playground-v2.5-1024px-aesthetic",
            torch_dtype=torch.float16,
            variant="fp16"
        ).to("cuda")

    def execute(self, requests):
        responses = []
        for request in requests:
            inp = pb_utils.get_input_tensor_by_name(request, "prompt")
            prompt = inp.as_numpy()[0][0].decode()
            print(prompt)
            # prompt = "sailing ship in storm by Leonardo da Vinci, detailed, 8k"
            image = self.model(prompt=prompt, num_inference_steps=50, guidance_scale=3).images[0]
            pixel_values = np.asarray(image)
            inference_response = pb_utils.InferenceResponse(
                output_tensors=[
                    pb_utils.Tensor(
                        "generated_image",
                        pixel_values,
                    )
                ]
            )
            responses.append(inference_response)
        return responses

启动 Triton 服务

/opt/tritonserver/bin/tritonserver --model-repository /content/python_backend/models

client.py

import time
import os
import numpy as np
import tritonclient.http as httpclient

from PIL import Image
from tritonclient.utils import *

IMAGES_SAVE_DIR = "/content/images/"

def text2image(prompt):
	if not os.path.exists(IMAGES_SAVE_DIR):
	    os.makedirs(IMAGES_SAVE_DIR)
	    
	client = httpclient.InferenceServerClient(url="localhost:8000")
	text_obj = np.array([prompt], dtype="object").reshape((-1, 1))
	
	input_text = httpclient.InferInput(
	    "prompt", text_obj.shape, np_to_triton_dtype(text_obj.dtype)
	)
	input_text.set_data_from_numpy(text_obj)
	
	output_img = httpclient.InferRequestedOutput("generated_image")
	timestamp = str(int(time.time()))
	filename = timestamp + ".png"
	output_path = IMAGES_SAVE_DIR + filename
	
	query_response = client.infer(
	    model_name="playground-v2.5", inputs=[input_text], outputs=[output_img]
	)
	image = query_response.as_numpy("generated_image")
	im = Image.fromarray(np.squeeze(image.astype(np.uint8)))
	im.save(output_path)
	return output_path

if __name__ == '__main__':
	start = time.time()
	prompt = "A beautiful Asian girl is sitting in a rocking chair in a beautiful garden, holding a cute kitten, admiring the beautiful scenery, with willow trees and a river."
    image_path = text2image(prompt)
    end = time.time()
	print("Time taken:", end - start)

客户端

python client.py

更多示例
Space ship.

The West Lake

推荐阅读

• 一. Triton Server Python 后端性能优化

参考

• Triton Server - Conceptual Guides

• Building Triton Without Docker

• Deploying HuggingFace models

• Triton 支持的数据类型汇总

• Deploy Stable Diffusion 不支持高并发

• Optimize and deploy model on Nvidia Triton server

• 深度学习怎么模型部署？ - 李稀敏的回答 - 知乎

• How to Run a Stable Diffusion Server on Google Cloud Platform (GCP)

• 深度学习模型如何部署？部署可以试试triton~

• TensorRT 官网

• Tensorrt的安装、模型转换、推理demo编写

• datacenter-gpu-manager

• Export huggingface SDXL model to ONNX

• Replicate.com 公开部署的模型

• Replicate Playgroundv2.5 推理接口源代码

• stable-diffusion-webui

• Error: install include <NvCaffeParser.h>

• ubuntu22.04 cuda cudnn tensorRT安装

• NVIDIA Deep Learning TensorRT Documentation