Amazon Web Services has introduced training session APIs for Amazon SageMaker AI, a programming model designed to give developers control over large language model training procedures without requiring them to manage GPU infrastructure directly.

According to AWS, developers can write Python loops that call operations including forward_backward(), optim_step(), sample(), and compute_logprobs(). The calls run against a live, stateful session in which the selected base model and Low-Rank Adaptation adapter remain loaded on managed GPUs. AWS says it handles GPU allocation, distributed execution, gradient synchronization, fault recovery, checkpoint storage, security, and infrastructure operations.

The SageMaker Python SDK model uses three clients. A ServiceClient starts sessions and holds platform settings, a TrainingClient provides training and checkpointing operations, and a SamplingClient supports generation and scoring. Active sessions also provide an OpenAI-compatible chat completions endpoint, according to the announcement.

AWS says the APIs can support supervised fine-tuning, preference optimization, reinforcement learning, and custom objectives. Its example uses Group Relative Policy Optimization on the GSM8K mathematics dataset with OpenAI’s GPT-OSS 20B and a LoRA adapter. The company also describes a cookbook containing ready-to-run supervised fine-tuning and reinforcement learning recipes, including a competitive-programming workload where generated code is tested.

The service is billed per token processed, AWS says, with GPU-bound operations incurring usage while local reward computation, data loading, logging, and debugging do not. Supported base models and regional availability are listed in the associated documentation.