AWS makes training sessions available for custom training loops in Amazon SageMaker AI
AWS has made training sessions available in Amazon SageMaker AI. Developers can write custom training loops and rewards; according to the company, the service handles GPU infrastructure and bills for tokens processed, rather than time spent waiting for reward computation.
AWS announced the general availability of training sessions in Amazon SageMaker AI. The new APIs let developers control custom training workflows when adapting models without managing GPU infrastructure. At launch, the supported models are gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B.
The developer controls the training loop from their own Python process and calls forward_backward(), optim_step() and sample() to compute gradients, update weights and generate responses. The model remains loaded on the GPU between steps within a stateful session. According to AWS, the service handles GPU allocation, distributed execution, gradient synchronization, failure recovery and checkpointing.
The developer can define custom loss functions, rewards and the sequence of training steps. According to the company, the service supports both LoRA and full fine-tuning, which updates all model weights. The SageMaker training sessions cookbook is being released at the same time, with runnable examples for supervised fine-tuning (SFT), meaning further training on labeled examples, and reinforcement learning (RL), meaning learning through rewards.
According to AWS, usage is billed based on the number of tokens processed. GPU time is not billed while rewards are being computed in the customer's own environment, and there is no need to reserve capacity.
Why it matters
The service is intended for developers whose needs go beyond a predefined training workflow: they can modify rewards or the loss function and control individual steps as training progresses. For experiments where evaluating responses takes longer, the stated pause in billing during reward computation outside the service also matters.
What was added since the original report
Verified updates
-
The training loop can be controlled from your own Python process.; The API provides forward_backward(), optim_step() and sample() operations.; The model remains loaded on the GPU between steps in a stateful session.; Runnable example workflows for SFT and RL are being released at the same time.
- The training loop can be controlled from your own Python process.
- The API provides forward_backward(), optim_step() and sample() operations.
- The model remains loaded on the GPU between steps in a stateful session.
- Runnable example workflows for SFT and RL are being released at the same time.
Two audiences, two different impacts
What this means
For individuals
When fine-tuning a model, a developer can modify custom training logic directly from a Python environment without having to set up and operate a GPU cluster.
For a business
A company can delegate management of its training infrastructure to Amazon SageMaker AI. Under the billing model described, experiment budgets depend on the number of tokens processed, which changes cost planning compared with hourly GPU rental.
DevelopmentCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.