Amazon SageMaker AI announces training session APIs for serverless model customization
Amazon SageMaker AI introduces training session APIs for serverless model customization, enabling direct control over post-training algorithms without managing GPU infrastructure.
Amazon SageMaker AI now offers training session APIs for serverless model customization, allowing customers to run supervised fine-tuning, reinforcement learning, or custom training procedures without provisioning or managing GPU infrastructure. Customers can select techniques such as SFT or DPO and execute training through SageMaker AI, which handles GPU allocation, model loading, distributed execution, and fault recovery. The APIs enable writing training loops in Python, including custom loss functions, reward signals, or data curricula, while SageMaker AI manages underlying infrastructure. This approach shifts operational overhead from infrastructure management to algorithm design, reducing setup complexity for teams.
Billing for training sessions is based on tokens processed, excluding local computations like reward scoring, which are not billed as GPU time. Customers can resume training from checkpoints, ensuring continuity without additional setup. The service also provides a cookbook with ready-to-run recipes for SFT and RL, serving as templates for custom training loops. These features aim to streamline experimentation and deployment of advanced training techniques.
Training session APIs are available in US East (N. Virginia), US West (Oregon), and Asia Pacific (Mumbai), with general availability announced today. Customers can access the APIs via the SageMaker Python SDK, a SageMaker Studio notebook, or by cloning the provided cookbook. Documentation and the SageMaker AI webpage offer additional resources for implementation and learning.
The announcement emphasizes flexibility in model customization, enabling teams to implement research-backed algorithms or proprietary techniques without infrastructure constraints. By abstracting GPU management, SageMaker AI allows focus on algorithmic innovation while maintaining operational efficiency. The integration of checkpointing and token-based billing further supports scalable and cost-effective training workflows.