Infrastructure options
Amazon SageMaker AI offers three infrastructure options for model customization. Choose based on your requirements for control, scale, and capacity access. For example, if you have reserved compute capacity through SageMaker AI Flexible Training Plans, you can use the customization Recipes to run customization jobs on your reserved capacity. The difference is in how compute resources are managed.
Recipes are pre-configured training configurations
that work with both SageMaker AI Training Jobs and HyperPod. They automate dataset loading,
distributed training, and checkpoint management. See
SageMaker Recipes
| Option | What it is | Interfaces | Best for |
|---|---|---|---|
| Serverless model customization | Job APIs: Fully managed, no instance selection. Infrastructure provisioned and released automatically. | Studio UI, Python SDK (SFTTrainer, DPOTrainer, etc.) | Quick experimentation, production jobs without ops overhead |
| Training session APIs: Interactive, long-lived sessions that expose low-level training primitives as API calls, so you drive the training loop from your own code. | Python SDK (ServiceClient), API |
Custom training loops, multi-turn reinforcement learning (RL), multi-stage pipelines | |
| SageMaker AI Training Jobs | Ephemeral jobs where you select instance type and count. Instances released after training. | Recipes, Python SDK (ModelTrainer) | Custom instance configs, advanced distributed setups |
| HyperPod | Persistent clusters with advanced fault detection and recovery. Clusters remain available across jobs. | Recipes, HP-CLI | Large-scale training, long-running jobs, multi-job workflows |