Announcing instance preference lists for Amazon SageMaker AI training jobs
aws ·
Amazon SageMaker AI has introduced Instance Preference Lists to streamline the allocation of GPU resources for training and processing jobs. This new feature allows users to submit a single job with an ordered list of up to five acceptable instance types, enabling the platform to automatically select the first available capacity. By evaluating instances in priority order, SageMaker reduces wait times during peak demand and eliminates the need for manual retry scripts or complex monitoring tools. The solution integrates seamlessly with Flexible Training Plans, allowing reserved capacity to take precedence while falling back to on-demand options if reservations are exhausted. Users can configure per-preference instance counts to ensure throughput equivalency across different GPU architectures. Additionally, the MaxPendingTimeInSeconds parameter provides control over queue duration for accelerated computing instances. This update transforms GPU management from a manual engineering burden into a simple configuration change, helping teams focus more on model development.
aws