Announcing new ways to customize models with SageMaker AI, without managing GPUs
Amazon SageMaker AI serverless model customization now includes training sessions, generally available today. You write and coordinate your own post-training loop (custom loss functions, reward signals, or multi-turn reinforcement learning) while SageMaker AI handles GPU allocation, distributed execution, fault recovery, and checkpoint storage.
First seen by Oriel on .
AI summary
AWS announces general availability of SageMaker AI serverless training sessions, letting teams write custom post-training loops while SageMaker manages GPU execution and infrastructure.
- Developers control training logic, including loss functions, reward signals, and algorithm stages, while SageMaker handles GPU allocation, distributed execution, fault recovery, and checkpoints.
- The sessions support techniques including SFT, DPO, RLVR, RLAIF, and multi-turn RL, with LoRA or full fine-tuning and chained customization.
- For GRPO, a sampler generates responses, an environment scores them, and the trainer calculates gradients and updates model weights before the sampler refreshes.
- Pricing is based on tokens processed and pauses while an external reward function scores a batch; capacity cannot be reserved.
- Training sessions support gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B, with data and checkpoints stored in the customer's S3 buckets.
#sagemaker ai #reinforcement learning #model fine-tuning #llm agents
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| AWS Machine Learning Blog | 123e9274ff10d780d0521efd9bdc03cfb23cff08 | https://aws.amazon.com/blogs/machine-learning/announcing-new-ways-to-customize-models-with-sagemaker-ai-without-managing-gpus/ Read discussion |