Amazon SageMaker AI Enables Full Control via Training Session APIs

Previously managed training techniques like supervised fine-tuning (SFT) and direct preference optimization (DPO) were extended with training session APIs that expose operations such as forward/backward passes and optimization steps through API calls. This empowers users to implement custom training algorithms while SageMaker AI manages GPU infrastructure, distributed execution, checkpointing, and fault recovery.

The APIs support reinforcement learning, custom loss functions, and novel training curricula without infrastructure provisioning. Billing is token-based, separating compute cost from local training logic. The launch includes ready-made training loop recipes and deployment in several AWS regions, facilitating flexible development of machine learning workflows at scale.