AWS Announces Serverless Training Sessions for SageMaker AI
According to AWS. Amazon SageMaker AI introduces a new way to customize foundation models by offloading infrastructure management to the platform.
- • Users can now execute custom training loops and loss functions without managing underlying GPU clusters.
- • The platform automates GPU allocation, distributed execution, and fault recovery for model customization.
- • Supports multiple customization techniques including SFT, DPO, and multi-turn reinforcement learning.
Infrastructure-Free Customization
Amazon SageMaker AI serverless training sessions allow developers to customize models without the need to select instance types or manage clusters. By offloading the GPU-heavy work to AWS, users can focus on writing their own training logic, including custom loss functions and reward signals, while the platform handles distributed execution and checkpoint storage.
Advanced Reinforcement Learning Support
The new capability is designed to simplify complex multi-turn reinforcement learning workflows. Developers can now define their own reward functions and RL stages within a single loop, enabling more precise control over agent behavior. The service is available today with support for models including gpt-oss-20b, gemma-4-31B-it, and Qwen3 variants.
- 1. aws.amazon.comOpen source