AWS has reportedly made SageMaker AI Training Sessions generally available. Developers control generation and the calculation and application of updates through APIs, while the service handles resource allocation, distributed execution, recovery and progress storage. Billing is based on processed tokens, with no charge while customer code evaluates results, according to the report. Training data and checkpoints remain in customer S3 storage. This offers finer control than a fixed fine-tuning job, but supported models and actual costs depend on the workflow.
Back to Tech
Tech
SageMaker sessions provide programmatic control of training loops
Developers define the procedure while the service manages GPU operations, recovery and state.