SageMaker puts custom training loops on managed GPUs
AWS has opened training sessions in SageMaker AI to general use. Developers direct how a model generates answers, receives scores and updates its parameters, while the service manages the computing infrastructure. The arrangement gives teams control over their own learning procedures, including multistep agent tasks, without requiring them to operate the underlying GPUs.
Artificial Intelligence··Morning
Developers direct the learning procedure
AWS, Amazon’s cloud-computing business, made training sessions in SageMaker AI generally available on October 22. Teams can specify their own learning loops, loss functions and reward signals while the service handles GPU allocation, distributed execution, recovery and saved training progress. The division leaves algorithm design with the developer and infrastructure operation with AWS. Developers use the service to customize an existing model for their own tasks.[1], [2]
The customer’s code runs in its own environment, connecting an agent with tools or simulators. A sampler on the service side generates responses; a trainer calculates gradients and changes the model’s weights. These are separate operations within a procedure directed by the customer. The developer can define both the sequence of operations and the signals used to judge the generated behavior.[1]
Rewards can follow several agent actions
AWS illustrates reinforcement learning with several responses to the same prompt. The customer’s verifier or reward function scores those responses and sends the results back to the trainer. After an update, the sampler refreshes with the changed weights before the next batch. This arrangement supports tasks whose success becomes apparent only after tool use or several conversational turns, with the scoring point chosen by the developer.[1]
Billing is based on tokens processed by the service, according to AWS. Its meter pauses while the customer’s environment scores a batch. The announcement does not establish a particular saving or an improvement in task success for every workload. Teams retain responsibility for deciding what successful behavior means and evaluating the model they have customized before deploying it in an application.[1]
Data stays in the customer’s storage
Sessions use an access-management execution role created by the customer. Customers can add their own encryption key and private-network configuration, with data and checkpoints kept in their S3 buckets in the selected region. AWS says it does not use customer data or customized model weights to train another model. Its launch materials list supported model options and example procedures, while leaving accuracy, bias and safety evaluation within the customer’s workflow.[1]
