Understand this story
On October 23, 2026, Amazon SageMaker AI added a training session API in serverless model customization. This API turns operations such as forward_backward(), optim_step(), and sample() into API calls, keeping the model loaded on the GPU; users only need to write the training loop in a local Python process, while GPU allocation, distributed execution, fault recovery, and checkpoint storage are handled by the service. The API is billed by processed tokens, and steps such as local reward computation do not count toward GPU time. The accompanying cookbook provides SFT and RL recipes; in the example, the GRPO loop for GPT-OSS 20B trains for 10 steps on GSM8K, and held-out set accuracy rises from 56% to 75%; the Code RL recipe, after 50 training steps, raises test accuracy from 40% to 50.5%. There are currently no subsequent developments.
AI-generated from public reports
Latest development · Report
Amazon SageMaker AI launches Training Sessions API, enabling custom training loops without managing GPUsDevelopments
1 confirmed developments
SageMaker AI training session API extends serverless model customization
Source:AWS Machine Learning
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 23
AWS Machine LearningFirst-party
Amazon SageMaker AI launches Training Sessions API, enabling custom training loops without managing GPUs
Amazon SageMaker AI has added a Training Sessions API to serverless model customization, turning operations such as forward_backward(), optim_step(), and sample() into API calls. The model stays loaded on the GPU, and users only need to write the training loop in a local Python process, while GPU allocation, distributed execution, fault recovery, and checkpoint storage are handled by the service. The API is billed by tokens processed, and steps such as local reward computation do not count toward GPU time. The accompanying cookbook provides SFT and RL recipes; in the example, the GRPO loop for GPT-OSS 20B trains for 10 steps on GSM8K, raising held-out set accuracy from 56% to 75%; the Code RL recipe raises test accuracy from 40% to 50.5% after 50 training steps.