SageMaker adds ordered GPU fallback for training jobs

AWS says Amazon SageMaker AI Training and Processing Jobs can now use instance preference lists, allowing customers to submit up to five acceptable instance types in priority order. SageMaker checks the list and launches on the first available type; accelerated-compute jobs can also remain in an event-driven retry queue bounded by MaxPendingTimeInSeconds. Training jobs can place Flexible Training Plan capacity first and use on-demand instances as a fallback. AWS says customers must ensure their containers are compatible across the listed instance types.