AI 市场文档

架构概述

1.1 概述、

Figure 1: AI app architecture overview

表 1. 描述 AI 服务架构中的组件

No.	Category	Component	Description	AI scope	Note
1	Client	Client	AI marketplace	N
2	Client callback	Client callback	Get results after a task has been completed	N
3	AI App API	AI App API	Receive request from client (marketplace) and process. Manage stages to solve a specific task.	Y
4	Storage	S3 storage	Store media data (Including request data from the client and AI result data). Services send media file metadata to each other	N
5		AI APP Database	Store user request information for statistics	Y
6		Redis	Store temporary results of tasks and requests: Will be deleted after a period of time	Y
7	Queue	RabbitMQ	Store pending tasks	Y
8	AI service 1	AI service 1 API	Receive requests to process tasks and push tasks into the queue	Y
9		AI service worker	Pull the task in the queue and process it	Y
10	AI service 2	AI service 2 API	Receive requests to process tasks and push tasks into the queue	Optional	Whether or not depends on the AI ​​problem
11		AI service worker	Pull the task in the queue and process it	Optional	Whether or not depends on the AI ​​problem

对于一个基本的AI问题：AI服务将需要至少6个组件（docker容器）。当AI问题复杂且需要分为多个阶段进行处理时，组件数量将会增加

简单AI问题的示例：人脸检测、人脸检测、人脸嵌入、人脸搜索……对于每个这样的问题，都需要开发和部署6个组件

对于人脸识别等更复杂的问题，需要经过许多步骤进行处理：人脸检测，然后是人脸嵌入，最后是搜索。这个问题最多有3个阶段，因此最多会有3个AI服务。因此，需要开发和部署的组件数量为10个（每增加一个服务增加2个组件）

1.2 AudioCraft服务

Audio Craft是一种根据用户描述生成声音和音乐的AI服务。它分为2个彼此独立运行的子服务。 AudioGen 用于生成声音，例如：汽车喇叭声、猫叫声……而 MusicGen 则专门用于生成音乐（带有旋律）。AudioCraft 的系统设计如图 2 所示

Figure 2: AudioCraft architecture

Figure 3: Call API sequence diagram

表2. 源代码中的序列图和可执行代码描述

No.	Description	Executable code	Note
1.1	Client create a request to AI service	Module: app_api.app_api Class:AudioCraftAppAPI Function:call Line number: 198
1.2	Service validation	Module: app_api.app_api Class:AudioCraftAppAPI Function:call Line number: 205
1.3	Create workflow with workflow_id (task_id)	Module: app_api.app_api Class:AudioCraftAppAPI Function:call: Line number: 223
1.4	Response to client status of request	Module: app_api.app_api Class:AudioCraftAppAPI Function:call: Line number: 227
1.5	Sen request to AudioGen API	Module: app_api.app_api Class:AudioCraftWorkFlow Function:trigger_request_text2audio: Line number: 132
1.6	Create a task and push to task queue	Module: audio_core_api.audio_core_api Class:AudioGenerationAPI Function:api_generate_audio Line number: 66
1.7	AudioGen worker pull task and perform	Module: audio_core_api.audio_gen_worker Class:AudioGenerationSerivce Function:run Line number: 28
1.8	Execute callback	Module: audio_core_api.audio_gen_worker Class:AudioGenerationSerivceCallback Function:run Line number: 104
1.9	Store media file on S3 storage	Module: audio_core_api.audio_gen_worker Class:AudioGenerationSerivce Function:run Line number: 54
1.10	Store temporary result on Redis	Module: audio_core_api.audio_gen_worker Class:AudioGenerationSerivce Function:run Line number: 74
1.11	Aggregate final result	Module: app_api.app_api Class:AudioCraftWorkFlow Function:trigger_text2audio_callback Line number: 145
1.12	Store final result on Redis	Module: app_api.app_api Class:AudioCraftWorkFlow Function:trigger_text2audio_callback Line number: 147
1.13	Store request information in database	Module: app_api.app_api Class:AudioCraftWorkFlow Function:process_result Line number: 103
1.14	Execute client callback	Module: app_api.app_api Class:AudioCraftWorkFlow Function:process_result Line number: 94

Figure 4: Result API sequence diagram

Table 3. Description of sequence diagram and executable code in source code

No.	Description	Executable code	Note
2.1	Client create a request to to get result based on task_id	Module: app_api.app_api Class:AudioCraftAppAPI Function:result Line number: 235
2.2	Service validation	Module: app_api.app_api Class:AudioCraftAppAPI Function:result Line number: 238
2.3	Retrieve workflow based on task_id (workflow uuid)	Module: app_api.app_api Class:AudioCraftAppAPI Function:result Line number: 258
2.4	Get URL in S3	Module: app_api.app_api Class:AudioCraftAppAPI Function:result Line number: 266
2.5	Response result to client	Module: app_api.app_api Class:AudioCraftAppAPI Function:result Line number: 288

Figure 5: Stats API sequence diagram

Table 4. Description of sequence diagram and executable code in source code

No.	Description	Executable code	Note
3.1	Client create a request to to get statistical information	Module: aimarket_core_app_base.app Class:BaseAppAPI Function:stats Line number: 112
3.2	Service validation	Module: aimarket_core_app_base.app Class:BaseAppAPI Function:stats Line number: 112
3.3	Query data in database	Module: aimarket_core_app_base.app Class:BaseAppAPI Function:get_results_by_user_id Line number: 117
3.4	Response result to client	Module: aimarket_core_app_base.app Class:BaseAppAPI Function:stats Line number: 120

2. Package

2.1 Docker image

No.	Component	Docker image name	Type
1	RabbitMQ	rabbitmq:3.8.14-management-alpine	Opensource
2	Redis	redis:6-alpine
3	AI APP database	postgres:12-alpine
4	AI APP API	your_registry_url/your_repo:version	Self-build
5	AudioGen API
6	AudioGen worker
7	MusicGen API
8	MusicGen worker

2.2 构建docker镜像

步骤1：克隆项目

$ git clone https://github.com/DopikAI-Labs/aimarket-audiocraft-service.git
# The Git repository URL may change when delivering the source code to the customer. Carefully check the provided URL.

第 2 步：更新子模块

$ cd aimarket-audiocraft-service
$ submodule update --init --recursive

如果提供的源代码是压缩文件，请跳过以上两个步骤

步骤 3：构建镜像

编辑脚本：./scripts/build_dockerfile.sh

更新 ./version.json 中的 ECR 和版本

#!/bin/bash

ECR="your_registry_url/your_repo"
TAG=$(jq -r '.api_version' version.json)
SERVICE="audio_craft"
IMAGE_NAME="${ECR}:${SERVICE}_${TAG}"

docker build -t $IMAGE_NAME -f dockerfile .

echo  image $IMAGE_NAME is built

$ cd aimarket-audiocraft-service

Build image

$ chmod +x ./scripts/build_dockerfile.sh
$ ./scripts/build_dockerfile.sh

If no errors occur, the terminal will show

$ your_registry_url/your_repo:audio_craft_0.1.4 is built

Step 4: Push image to registry

Step 4.1: Login to registry

Install awscli

$ pip install awscli
# or pip3 install awscli
# If you're using a virtual environment created with venv or conda, make sure it's activated before proceeding.

Configure AWS information

$ aws configure
# AWS Access Key ID [None]: Enter your access key ID
# AWS Secret Access Key [None]: Enter your secret  access key 
# Default region name [None]: Enter region
# Default output format [None]: Can skip it

For the next time, there is no need to do the above 2 steps in step 4.1 section

Get the password

$ aws ecr get-login-password
# The terminal will generate a string (password)

Login repository: Make sure repository has been created

$ docker login -u AWS -p <<generated_string_from_above_command>> <<repository_URL_in_ECR>>

Step 4.2: Push docker image

Push docker image

$ docker push docker_image_name:tag

3. Deployment

3.1 Infrastructure requirements

No.	Category	Item	Required
1	Hardware	CPU
2		CPU arch	x64
3		RAM	16G
4		NVIDIA GPU	12GB
5		Storage (root)	32GB
6	Software	OS	Ubuntu 20.04 or later
7		Docker	26.x.x or later
8		Docker-compose	1.29.x or later

3.2 Deployment procedure

Step 1: Pull docker image

Step 1.1 Login registry

Do the same as step 4.1 in the package section

Step 1.2 Pull image

$ docker pull docker_image_name:tag
# docker_image_name:tag => It is an pushed image in the package section 

Step 2: Deploy service

Make a deployment directory

$ cd ../../../deployment_space_dirctory_path
$ mkdir audio_craft
$ cd audio_craft
$ mkdir v0.1.4
$ cd v0.1.4
# v0.1.4: Change to match the version to be deployed

Create a docker-compose.yml file

$ vim docker-compose.yml 
# Copy ./aimarket-audiocraft-service/docker-compose.yml and past to it

Create .env file

$ vim .env 
# Copy ./aimarket-audiocraft-service/.env and past to it

With the .env file, all environment variables highlighted in red need to be updated

No.	Group	Variable	Default value	Required update
1	AI Service	APP_API_PORT	8007	Y
2		AUDIO_GEN_PORT	8081	Optional
3		MUSIC_GEN_PORT	8082	Optional
4		MAX_DURATION	30	Optional
5		MAX_PROMPT_LEN	100	Optional
6		CACHE_MODEL_DIR	../model_hub	NO
7		TRANSFORMERS_CACHE	/model_hub	NO
8		AUDIOCRAFT_CACHE_DIR	/model_hub	NO
9		AUDIO_GEN_MODEL	facebook/audiogen-medium	Optional
10		MUSIC_GEN_MODEL	facebook/musicgen-small	Optional
11		API_VERSION	0.1.4	Optional
12		API_NAME	AudioCraft	Optional
13	Database	POSTGRES_USER	aimarketAdmin	Optional
14		POSTGRES_PASSWORD	aimarketPassword	Optional
15		POSTGRES_DB	aimarketDb	Optional
16	S3 storage	AWS_REGION_NAME	<<SECRET>>	Y
17		AWS_ACCESS_KEY	<<SECRET>>	Y
18		AWS_SECRET_KEY	<<SECRET>>	Y
19	RabbitMQ	RABBITMQ_HOST	aimarket_rabbitmq	Optional
20		RABBITMQ_VHOST	aiqrbroker	Optional
21		RABBITMQ_USER	aiqradmin	Optional
22		RABBITMQ_PASS	aiqrPassword	Optional
23		RABBITMQ_PORT	5672	Optional
24		RABBITMQ_UI_PORT	15672	Optional
25	Redis	REDIS_HOST	aimarket_redis	Optional
26		REDIS_PORT	6379	Optional
27		REDIS_DB_NAME	0	Optional
28	MarketPlace	MARKETPLACE_TOKEN	<<SECRET>>	Y
29		MARKETPLACE_CALLBACK	<<SECRET>>	Y

Run service

# make sure you are in the same directory with docker-compose.yml file
$ docker-compose up

Check the terminal to determine if any errors occurred. If not, Ctrl + C to stop running service

Run service in detached mode

$ docker-compose up -d