Amazon SageMaker AI announces training session APIs for serverless model customization

Amazon SageMaker AI announces training session APIs for serverless model customization, allowing customers to build custom training loops without provisioning GPU infrastructure.

• Write custom training loops for SFT, DPO, reinforcement learning, or proprietary algorithms
• API calls handle forward_backward(), optim_step(), and sample() operations on live stateful sessions
• SageMaker manages GPU allocation, model loading, distributed execution, fault recovery, and checkpointing
• Billing per token processed; local computation like reward scoring not charged as GPU time
• SageMaker training sessions cookbook provides ready-to-run SFT and RL recipes as blueprints
• Generally available in US East (N. Virginia), US West (Oregon), and Asia Pacific (Mumbai)

Training session APIs enable flexible model customization with serverless infrastructure management and transparent token-based billing.