Platform For AI:Features

Module

Feature

Description

Reference

AI computing resource management

Lingjun resources

PAI provides Lingjun resources for large-scale, high-density computing workloads. Lingjun resources deliver heterogeneous computing power required for high-performance AI training and computation. You can leverage Lingjun resources for training tasks in PAI.

Create resource quotas

General training resources

General training resources are deep learning training resources built on Container Service for Kubernetes (ACK). These resources provide scalable, stable, user-friendly, and high-performance runtimes for training deep learning models.

General computing resource quotas

Other big data computing resources

Big data computing resources including MaxCompute and Realtime Compute for Apache Flink.

Overview of AI computing resources

Workspaces

Resource management

Workspace administrators can associate AI computing resources from the current Alibaba Cloud account with the workspace, enabling workspace members to utilize these resources for development and training activities.

Manage workspaces

Workspace notification

PAI provides a notification mechanism for workspaces. You can create notification rules to track and monitor Deep Learning Containers (DLC) jobs or Machine Learning Designer pipelines. Notification rules can also trigger events based on model version status changes.

Create a notification rule

Workspace storage and SLS configuration

Workspace administrators can configure the default storage path for development and training within the workspace, as well as the storage lifecycle for temporary tables.

Manage workspaces

Member and permission management

PAI employs role-based access control with multiple predefined roles, including labeling administrators, algorithm developers, and algorithm operations and maintenance (O&M) personnel, facilitating efficient collaboration. You can manage AI asset visibility scope within workspaces and configure access permissions for different roles.

Manage members of a workspace

QuickStart

Model Hub

PAI provides access to diverse pre-trained models from open-source communities including ModelScope and Hugging Face.

Model deployment and training

Pre-trained model training

You can utilize the pre-trained models for training tasks in PAI.

Model deployment and training

Pre-trained model deployment

You can deploy the pre-trained models as services in PAI.

Model deployment and training

Machine Learning Designer

Pipeline building

Machine Learning Designer enables you to build and debug models using visual pipelines. You can drag and drop components onto the canvas to construct pipelines tailored to your business requirements.

Pipeline overview

Pipeline import and export

You can export pipelines as JSON files and import JSON files into workspaces to reconstruct pipelines.

Export and import pipelines

Pipeline scheduling

You can leverage DataWorks to schedule Machine Learning Designer pipelines on a periodic basis.

Use DataWorks tasks to schedule pipelines in Machine Learning Designer

Preset pipeline templates

PAI provides industry-specific pipeline templates covering domains such as product recommendation, news classification, financial risk management, weather prediction, healthcare diagnostics, agricultural lending, and demographic analysis. These templates include complete datasets and documentation for streamlined implementation.

Use cases for Designer

Custom pipeline templates

You can create custom pipeline templates based on your proprietary algorithm workflows and share them with team members. Team members can directly perform modeling, deployment, and production validation using these custom templates.

Create a pipeline from a custom template

Dashboards

Machine Learning Designer provides interactive dashboards to visualize data analysis, model performance, and prediction results.

Use dashboards to view analytical reports

Preset algorithm component library

PAI provides hundreds of built-in algorithm components spanning multiple domains including data sources, data preprocessing, feature engineering, statistical analysis, machine learning, time series analysis, recommendation systems, anomaly detection, natural language processing, network analysis, financial analytics, computer vision, speech processing, and custom algorithms.

Overview of Designer components

Custom algorithms

You can implement custom nodes using multiple programming interfaces including SQL, Python, and PyAlink scripts.

Custom algorithm components

Data Science Workshop (DSW)

Cloud-native development environment

DSW provides a flexible, stable, user-friendly, and high-performance environment for AI development, offering both CPU-accelerated and GPU-accelerated resources to support training workflows.

What is DSW?

DSW Gallery

DSW Gallery provides easy-to-use cases from various industries and technical verticals to help improve development efficiency.

Notebook Gallery

JupyterLab

DSW integrates open source JupyterLab and provides plug-ins for custom development. You can directly start Notebook to write, debug, and run Python code without O&M configurations.

Access DSW instances from the console

WebIDE

DSW provides WebIDE in which you can install open source plug-ins for modeling.

Access DSW instances from the console

Terminal

DSW supports character terminals to debug models.

Access DSW instances from the console

Persistent instance environment

You can manage the lifecycle of the development environment, save the instance environment, mount and share data, and persist the environment image.

Mount a dataset, OSS, NAS, or CPFS

Resource usage monitoring

You can view real-time resource usage in a visualized manner.

Access DSW instances from the console

Image creation

You can create an image and save the image to Container Registry for subsequent distributed training or inference.

Manage DSW instances

SSH remote connection

DSW provides the following SSH connection methods: direct connection and proxy client connection. You can select a connection method based on the resource dependencies, usage methods, and limits of the connection methods to meet your business requirements.

Connect to a DSW instance over SSH

Deep Learning Containers (DLC)

Cloud-native distributed training environment

DLC is a deep learning platform developed based on Container Service for Kubernetes (ACK) that provides stable, easy-to-use, scalable, and high-performance runtimes for training deep learning models.

Before you begin

Dataset mounting

You can mount multiple datasets, such as File Storage NAS or Object Storage Service (OSS) datasets, in DLC at the same time.

Before you begin

Public and dedicated resource groups

DLC provides public and dedicated resource groups.

Before you begin

Official and custom images

DLC allows you to use official images or custom images to submit training jobs.

Before you begin

Distributed trainings

DLC provides a distributed deployment solution for implementing data parallelism, model parallelism, and hybrid parallelism.

Create a training job

Training job management

DLC allows you to manage jobs during the entire lifecycle.

Manage training tasks

Elastic Algorithm Service (EAS)

Resource group management

EAS provides resources in resource groups for isolation. When you create a model service, you can deploy the model service in the public resource group provided by the system or a dedicated resource group that you created.

Overview of EAS resource groups

Service and application deployment

You can deploy models that you downloaded from the open source community or models that you trained as inference services or AI-powered web applications in EAS. EAS provides multiple methods that you can use to deploy models. You can use the PAI console to deploy models as API services.

Custom deployment

Service debugging and stress testing

After you deploy the service, you can use the online debugging and stress testing feature to test whether the service runs as expected.

Service debugging and stress testing

Auto scaling

You can configure automatic scaling, scheduled scaling, and elastic resource pools for EAS services.

Auto scaling

Service calls

EAS provides the following service call methods based on the network environment of the client: Internet access, VPC access, and VPC direct connection.

Service invocation

Asynchronous inference

EAS provides the asynchronous inference feature, which allows you to obtain inference results by subscribing to requests or polling.

Asynchronous inference services

Integrated resource group and service management capabilities

EAS provides standard OpenAPI and SDKs that support integration.

API overview

AI computing asset management

Datasets

PAI provides public datasets and supports dataset management during labeling and modeling. PAI also support OSS and NAS datasets and SDK calls.

Create and manage datasets

Models

PAI allows you to manage versions, lineages, evaluation metrics, and associated services of models in a centralized manner.

Register and manage models

Tasks

PAI supports management of distributed training tasks and PAIFlow pipeline runs.

Manage jobs

Images

PAI provides official images and supports image management.

View and add images

Code builds

You can register code repositories to PAI to facilitate code version management in PAI modules.

Configure code

Custom components

You can create custom algorithm components based on your business requirements. You can use custom components together with preset components in Machine Learning Designer to manage pipelines in a flexible manner.

-

AutoML

Automatic hyperparameter optimization (HPO)

HPO is used to automatically fine-tune model-related parameters and training parameters.

How AutoML works

Scenario-based solutions

Multimedia analysis

PAI provides ready-to-use image-related services such as image labeling, classification, and quality evaluation.

Overview of multimedia analysis

AI acceleration

Dataset Accelerator

DatasetAcc is a PaaS service developed by Alibaba Cloud to accelerate AI and datasets in the cloud. DatasetAcc provides dataset acceleration solutions for various cloud-native training engines by pre-analyzing and preprocessing training datasets used in machine learning training. This helps improve the overall training efficiency.

-

Easy Parallel Library (EPL)

EPL is an efficient and easy-to-use framework for distributed model training. EPL uses multiple training optimization technologies and provides easy-to-use API operations that allow you to use parallelism strategies. You can use EPL to reduce costs and improve the efficiency of distributed model training.

Use EPL to accelerate AI model training

PAI-Rapidformer

PAI-Rapidformer applies various technologies to optimize the training of PyTorch transformers and provide optimal training performance.

Pai-Megatron-Patch overview

Blade

Blade integrates various optimization technologies. You can use PAI-Blade to optimize the inference performance of a trained model.

Overview of Blade

PAI-SDK

Distributed model training

PAI SDK for Python provides an easy-to-use HighLevel API that allows you to submit training jobs to PAI and run the jobs in the cloud.

Submit a training job

Service deployment

PAI SDK for Python provides an easy-to-use HighLevel API that allows you to deploy models to PAI and create inference services.

Deploy an inference service