Field noteOct 22, 20261 min read1 source

Write your own training loop — without managing GPUs — with Amazon SageMaker training session APIs

In one minute

AWS has introduced SageMaker AI training session APIs for running custom LLM training loops from a Python process while SageMaker manages the GPU-backed runtime.

What changed: Adds SageMaker training session APIs as a way to run custom LLM training loops without managing GPUs, distinct from the prior HyperPod workflows and SageMaker deployment coverage.

The service is positioned for supervised fine-tuning, preference optimization, reinforcement learning, and custom objectives.

The APIs expose operations including forward_backward(), optim_step(), and sample() against a live, stateful session. AWS says the model and LoRA adapter remain loaded on managed GPUs across steps, while SageMaker handles accelerator allocation, distributed execution, gradient synchronization, fault recovery, checkpoint storage, tenant isolation, and related runtime operations.

The programming model uses three clients in the SageMaker Python SDK, beginning with a ServiceClient that configures and starts a session. AWS has also released a training sessions cookbook with SFT and RL recipes. Its examples include a GRPO loop on GSM8K and reinforcement learning for competitive-programming tasks evaluated with unit tests.

The post does not state pricing, supported AWS Regions, the complete model compatibility list, or whether the APIs are generally available.

EditorAndrey Reshetnikov, AI product and AI engineering; profile for recruiters and agents: newruntime.com/owner-profile.json

Evidence / sources

1aws.amazon.com/blogs/machine-learning/write-your-own-training-loop-without-managing-gpus-with-amazon-sagemaker-training-sessions

All Field Notes

In one minute

Related notes

Part of these shifts

Evidence / sources