From Model to Service: Alibaba Cloud Machine Learning PAI One‑Stop Model Development and Deployment Practice

The talk, titled “From Model to Service: Alibaba Cloud Machine Learning PAI One‑Stop Model Development and Deployment Practice,” was delivered by senior Alibaba expert Luo Yiyun and provides a comprehensive overview of the end‑to‑end machine‑learning workflow on Alibaba Cloud.

The machine‑learning process is divided into three stages: data preparation, model development, and model deployment, with data scientists, algorithm scientists, and development/operations engineers respectively responsible for each stage.

Model development faces three major challenges: a steep learning curve, sub‑optimal performance, and high computational cost, which together reduce development efficiency and delay business value delivery.

Users often spend weeks or months to deploy a model to production; pre‑trained models and open‑source model communities such as HuggingFace, ModelScope, and others dramatically reduce this time by allowing fine‑tuning on user data instead of training from scratch.

Pre‑trained models and open‑source communities have democratized AI development, offering large collections of state‑of‑the‑art models and easy‑to‑use APIs that lower cost and improve results.

Alibaba Lingjie is an integrated big‑data + AI platform built on Alibaba Cloud infrastructure, featuring four key attributes—Scale, Simplicity, Speed, and Scenario—to support large‑model training, deployment, and industry‑specific solutions.

The PAI architecture consists of four modules: (1) machine‑learning frameworks and cloud‑computing infrastructure, (2) the PAI core engine (PAI‑Whale/EPL, PAI‑Blade, PAI‑EAS), (3) PAI pre‑trained model development, and (4) the AI application layer, supporting frameworks such as TensorFlow, PyTorch, EasyTransfer, and Transformers.

In the detailed pre‑trained model development flow, Lingjie reads massive training datasets, PAI‑Whale performs distributed training to produce a large model, which can then be fine‑tuned and deployed via the EAS service.

The PAI console provides three sub‑modules: DSW for interactive modeling with small data, DLC for container‑based distributed training on full datasets, and EAS for online model serving.

A practical demo walks through using ModelScope’s PALM2.0 Chinese text‑generation model, fine‑tuning it for couplet generation, training with DSW and DLC, and deploying the final model with EAS, showing significant quality improvement after fine‑tuning.

The discussion on MLOps highlights the need for end‑to‑end efficiency across data preparation, model development, and deployment, emphasizing tools for dataset management, experiment tracking, debugging, model evaluation, packaging, A/B testing, and monitoring.

In summary, PAI serves as a bridge between algorithmic innovation and business production, enabling rapid model creation, scalable training, and seamless deployment, thereby accelerating both AI research and real‑world applications.