AWS Holds MLOps Distributed Training Workshop, Sharing Large-Scale Model Training and Inference Strategies Based on SageMaker

AWS Korea ran an MLOps distributed training workshop for customers at Yeoksam Center Field from November 8 to 9, 2022. This event was the first of its kind attempted in Korea among Asian countries, and was operated as a hands-on program for field practitioners.

Practical MLOps Application and Domestic Case Studies

On the morning of the first day, along with a keynote speech by Emily Webber, advanced overseas case studies were introduced, including JP Morgan's Data Mesh construction and Fidelity's use of Feature Store. Data Mesh drew attention as a new paradigm that manages complexity by separating a centralized Data Lake by role. In the afternoon session, domestic manufacturing companies such as LG presented cases of integrated management of ETL, Model Registry, and Deployment using SageMaker, and participants directly practiced with various SageMaker features.

Large-Scale Distributed Training Trends and Technology

On the second day, the necessity of distributed training due to the emergence of large-scale applications such as Text-to-Image and Text-to-Video was emphasized. AWS confirmed that performance improves as data and model parameters increase according to Scaling Laws, and to support this, it offers Data Parallelism and Model Parallelism methods using Trainium chips. Combining SageMaker Ephemeral Training clusters with Spot Instances can reduce training costs, but caution is needed when model size grows larger.

Inference Optimization and Model Monitoring

Participants trained a GPT-2 model in SageMaker Studio while checking logs linked with CloudWatch, and experienced various inference deployment methods tailored to a data scientist's skill set. In addition, they discussed analyzing bottlenecks in the training process with Amazon SageMaker Debugger, and detecting data drift through Data Clarify and Model Monitor to maintain model performance.