|
Today I’m excited to announce new serverless customization in Amazon SageMaker AI for popular AI models such as Amazon Nova, DeepSeek, GPT-OSS, Llama, and Qwen. The new customization option provides an easy-to-use interface for the latest fine-tuning techniques, such as reinforcement learning, so you can speed the process of customizing an AI model from months to days.
With just a few clicks, you can seamlessly select a model and fit technique and handle model evaluation and deployment—all completely serverless, so you can focus on model tuning rather than infrastructure management. When you choose serverless customization, SageMaker AI will automatically select and provision the appropriate computing resources based on the model and data size.
We start by adapting the serverless model
You can start customizing your models in Amazon SageMaker Studio. Choose Models in the left navigation bar to see your favorite AI models that you can customize.
Customize with the user interface
You can customize your AI models with just a few clicks. IN Customize the model drop-down list for a specific model, e.g Meta Llama 3.1 8B Instructionstake your pick Customize with the user interface.
You can choose the customization technique used to adapt the base model to your use case. SageMaker AI supports Fine tuning under supervision and the latest model fitting techniques including Direct preference optimization, Reinforcement learning from verifiable rewards (RLVR)and Reinforcement Learning from AI Feedback (RLAIF). Each technique optimizes models in different ways, with factors such as dataset size and quality, available computing resources, the task at hand, desired levels of accuracy, and deployment constraints influencing the choice.
Upload or select a training dataset to match the format required by the selected fitting technique. Use the values of batch size, learning rate, and number of epochs recommended by the selected technique. You can configure advanced settings such as hyperparameters, the newly introduced serverless MLflow application for monitoring experiments, and encryption of network and storage volume. Choose Submit to start training the model.
After your training task is complete, you can see the models you created in the My models table Choose View details in one of your models.
By choosing Continue to customizeyou can continue to fit the model by adjusting the hyperparameters or training with different techniques. By choosing Evaluateyou can evaluate your customized model to see how it performs against the base model.
When you complete both tasks, you can choose one of them SageMaker gold It will underlie in Deploy drop-down list to deploy your model.
For serverless derivation, you can choose Amazon Bedrock. Choose It will underlie and model name to deploy the model to Amazon Bedrock. To find your deployed models, select Imported models in the Bedrock console.
You can also deploy your model to a SageMaker AI inference endpoint if you want to control deployment resources such as instance type and number of instances. After deploying SageMaker’s AI On dutyyou can use this endpoint to perform inference. IN Playground you can test your customized model using a single challenge or chat mode.
With serverless MLflow, you can automatically log all critical experiment metrics without modifying code and access rich visualizations for further analysis.
Customize with code
When you choose to customize with code, you can see a sample notebook for fine-tuning or deploying AI models. To edit the sample notebook, open it in JupyterLab. Alternatively, you can deploy the model immediately by selecting it Deploy.
You can choose an Amazon Bedrock or SageMaker AI endpoint by selecting deployment resources from either Amazon SageMaker Inference or Amazon SageMaker Hyperpod.
When you choose Deploy at the bottom right of the page will be redirected back to the model details page. After your SageMaker AI deployment is up and running, you can use this endpoint for inference.
OK, you’ve seen how to streamline model fitting in AI SageMaker. Now you can choose your favorite way. To learn more, visit the Amazon SageMaker AI Developer Guide.
Now available
The new customization of the serverless AI model in Amazon SageMaker AI is now available in the US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland) regions. You only pay for tokens processed during training and inference. To learn more details, visit the Amazon SageMaker AI pricing page.
Try it out in Amazon SageMaker Studio and submit feedback to AWS re:Post for SageMaker or through your usual AWS support contacts.
— Channy