Elastic Training on Amazon SageMaker HyperPod

This course explores elastic training capabilities on Amazon SageMaker HyperPod, enabling ML workloads to automatically scale based on resource availability. You will learn how to maximize GPU utilization, reduce training costs, and accelerate model development while maintaining training quality.

• Explaining the mechanics of elastic training on SageMaker HyperPod, including automatic expansion to idle AI accelerators (e.g., GPUs) and contraction when higher-priority tasks require resources.
• Configuring elastic policies and recipes in SageMaker HyperPod (e.g., updating YAML configurations for publicly available foundation models with no code modifications needed).
• Identifying use cases and best practices for elastic training to optimize large-scale distributed workloads, such as handling dynamic cluster availability, graceful scaling, and integration with checkpointing/resilience features.
• Troubleshooting and monitoring elastic training jobs, recognizing benefits like maximized compute utilization, minimized idle time, and seamless continuation during resource fluctuations.

• Understand how elastic training transforms traditional fixed-size training runs into dynamic, efficient processes that scale up/down automatically, improving cost efficiency and throughput for foundation model training.
• Gain knowledge of integrating elastic training into SageMaker HyperPod clusters (often EKS-based) to handle variable resource pools, prioritize workloads, and achieve higher accelerator utilization without manual reconfiguration.
• Be prepared to apply elastic training in production AI development scenarios, contributing to faster iteration cycles, reduced cost overruns, and resilient large-scale model training on AWS.

• 1-hour digital course content with explanations, architecture overviews, configuration examples, and scenarios demonstrating elastic training on Amazon SageMaker HyperPod.
• Intermediate-to-advanced training in the Artificial Intelligence domain, part of broader SageMaker HyperPod and SageMaker AI learning resources (complements related new features like checkpointless training).
• Alignment with AWS announcements and documentation on elastic training (launched Dec 2025), including integration with HyperPod's managed clusters, recipes, and no-infrastructure-management approach.
• Certificate of completion issued.