The Pulse
AWS Opens Serverless Training Sessions for Custom LLM Loops
AWS made SageMaker AI Serverless Training Sessions generally available on October 22, letting developers write custom post-training loops without provisioning GPU infrastructure. The service handles distributed execution and checkpoints, ch
AI.info Team · Written with AI · fact-checked ·
AWS is making custom post-training loops available without requiring teams to provision or manage GPU infrastructure. Announced October 22, SageMaker AI Serverless Training Sessions reached general availability, giving developers control over training logic while AWS manages GPU allocation, distributed execution, fault recovery and checkpoint storage.
The distinction is between writing the algorithm and operating the machines that run it. Developers can specify custom loss functions, reward signals and reinforcement-learning stages; SageMaker AI handles the compute-heavy work. AWS says users pay for tokens processed, and billing pauses while their own environment scores a batch of responses.
Custom loops, managed GPU work
The launch extends SageMaker AI’s existing serverless customization options, which include supervised fine-tuning, direct preference optimization and reinforcement-learning methods. AWS already offers managed multi-turn reinforcement learning, but says some customers wanted to define their own losses, rewards and training stages within one loop. Training Sessions are aimed at that need, rather than replacing the managed techniques.
That control matters for agent training, where a model may need to call tools, receive feedback and try again before a task can be judged. Teams can run the training loop and reward function in their own environment, connecting an agent to tools or simulators. The session supplies a sampler to generate responses and a trainer to calculate gradients and update model weights.
Sampling, scoring and updating
AWS illustrates the workflow with Group Relative Policy Optimization, or GRPO. The developer’s loop requests several responses to a prompt, scores them with a verifier, reward model or test harness, then passes the scores to the trainer. Calls to forward_backward() calculate gradients, while optim_step() applies the weight update; the loop then refreshes the sampler with the updated weights and repeats.
The arrangement brings generation and model updates into a managed session, while leaving the scoring logic with the customer. That separation can help teams that need rewards tied to a private database or a bespoke test environment, but it also means they still have to build and operate that environment and write the loop that coordinates it.
Four models at launch
Training Sessions launch with support for gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B. Developers can start from an IDE by installing the SageMaker Python SDK on Linux or macOS with Python 3.10 through 3.12, then cloning AWS’s cookbook. In SageMaker Studio, AWS says a ready-to-use development environment can launch in under 20 seconds.
The announcement also points to agent skills that can help with data preparation, technique selection, training, evaluation and deployment. AWS lists SageMaker Studio, Kiro, Claude Code, Codex and GitHub Copilot as environments where the agent can work. Those tools guide the workflow; they do not remove the need to decide what counts as a successful model response.
Customer control comes with a narrow first roster
Each session uses an AWS Identity and Access Management execution role created by the customer. Teams can configure their own encryption key and virtual private cloud settings, while training data and checkpoints remain in their chosen Amazon S3 buckets and AWS Region. AWS says it does not use customer data or customized weights to train another model.
ImagineSoftware’s Rob Gontarek, senior vice president of data science, said the service’s API gave scientists direct control over the forward and backward passes, sampling and optimizer step while building a healthcare claim-denial triage model. Karini AI CEO Nitin Wagh said the team used it to fine-tune gpt-oss-20b for multi-turn tool calling against an Amazon Neptune knowledge graph: “Setup just worked, with no manual tuning and no out-of-memory wrangling.”
The announcement does not publish per-token rates or a timetable for adding models, leaving two practical questions for teams weighing a move from self-managed training. For now, the supported roster is four models, and the bill is tied to processed tokens rather than reserved capacity.