AWS launches SageMaker training sessions: write your own RL loop without managing GPUs

AWS ML Blog · rss · 2026-10-23

AWS announced general availability of training sessions in SageMaker AI serverless model customization, letting developers write their own post-training loops while AWS handles GPU allocation, distributed execution, fault recovery, and checkpoint storage.

Key points:

Developers control training logic including custom loss functions and reward signals; techniques supported span SFT, DPO, RLVR, RLAIF, and multi-turn RL, with LoRA or full fine-tuning
A training sessions cookbook plus model customization agent skills work inside SageMaker Studio, Kiro, Claude Code, Codex, or GitHub Copilot to handle data prep, technique choice, training, evaluation, and deployment
A GRPO walkthrough shows the loop: sample responses, score with your reward function, forwardbackward() and optimstep() for updates, then refresh the sampler
Billing is per token processed, with the meter paused while your reward function scores batches locally
Available now for gpt-oss-20b, Gemma, and Qwen3 models; customers like ImagineSoftware and Karini AI already use it for healthcare claim triage and agentic tool-calling fine-tuning

Original post →

AWS launches SageMaker training sessions: write your own RL loop without managing GPUs

More from coding & agent