SageMaker AI training session APIs let teams write custom post-training loops on serverless GPUs

Until now, serverless fine-tuning services have mostly offered fixed recipes such as supervised fine-tuning or DPO with preset knobs. SageMaker's training sessions open that up. Developers write the training loop themselves in Python, with arbitrary loss functions, reward models or multi-turn reinforcement learning, while AWS manages the GPU fleet underneath.

The infrastructure work AWS absorbs is significant: provisioning accelerators, sharding across nodes, recovering from failures and checkpointing state. AWS demonstrates the APIs with a GRPO loop on the GSM8K math benchmark and an RL workload on competitive programming problems, the two canonical examples of the post-training wave of the last year.

This competes with services like Tinker-style hosted training APIs and Together and Fireworks fine-tuning, and it pairs with research trends visible on Hugging Face this week. Microsoft's ThinkingBox, for example, converts database grading into RL reward signals for post-training agents. The pitch to enterprises: you can build domain-specialized agents with RL without hiring a cluster-operations team.

What to watch: per-session pricing compared with reserved capacity, which base models are supported, and limits on session length. With Samsung reportedly sold out of 2027 HBM capacity and GPU prices expected to rise, serverless access that avoids long reservations may become more attractive, assuming AWS can guarantee capacity.