AWS opens SageMaker training sessions for custom training loops

AWS announced general availability of training sessions on 22 October 2026. Its example separates the work clearly: a sampler generates responses, the developer’s reward function scores them, and a trainer calculates and applies changes to the model. The code then refreshes the sampler before processing another batch. This sequence describes the announcement’s reinforcement-learning example; the actual sequence depends on the chosen algorithm.

The technical explanation describes a Python process running on the developer’s own machine. It calls individual training operations against a live session, where the model remains loaded on GPUs between steps. AWS positions this approach for researchers designing custom objectives and engineers who need more control than configured training recipes provide. Developers still write the algorithm; managing the infrastructure is the service’s responsibility.

The setup instructions require preparation before the first session: configure the calling identity, the execution role assumed by the session, and a model package group for checkpoints. Caller and execution role have different permissions. Saved checkpoints are registered as model packages in that group.

For a concrete starting task, the examples documentation describes a code-training recipe that runs generated solutions against unit tests in a local sandbox and rewards their pass rate. Recipes cover evaluation, checkpointing with resume, and metrics alongside batching. A separate deployment notebook takes a saved checkpoint to a real-time endpoint: completing the training loop is not itself deployment.