Amazon SageMaker AI @amazon-sagemaker-ai :: Fully managed service to build, train and deploy machine learning models, including foundation models > SageMaker AI is a fully managed ML service with tools for data preparation, training, tuning, deployment and operations. It was renamed from Amazon SageMaker on December 3, 2024, and API names and namespaces still use the sagemaker prefix. > Choose it when you need to train or fine-tune your own models, bring your own algorithms, or control the hosting of an endpoint. Choose Amazon Bedrock when you only need to call and ground existing foundation models. > Several features are closed to new customers (Ground Truth, Model Monitor, Clarify and Role Manager), so check each feature page before you plan around it. + What is Amazon SageMaker AI? + SageMaker AI features + Amazon SageMaker AI pricing Prepare data :: Label, clean and store the data that models learn from > Data preparation covers labeling, transformation, feature management and processing jobs. The Prepare data section of the guide also covers SQL in Studio, Amazon EMR and AWS Glue interactive sessions. + Prepare data Ground Truth @amazon-sagemaker-ground-truth :: Data labeling with a private or vendor workforce; closed to new customers > Ground Truth builds labeled datasets using workers from a private workforce or an AWS Marketplace vendor. Automated data labeling uses ML to decide which data humans must label. Data Wrangler :: Data preparation, now part of SageMaker Canvas > In the updated Studio experience you reach Data Wrangler through SageMaker Canvas, which adds a natural language interface alongside the visual one. Feature Store :: Create, store and share ML features > Feature Store keeps an online store for real-time inference and an offline store for model training and batch inference. Using the same feature processing for both reduces training-serving skew. Processing jobs :: Run preprocessing, feature engineering and evaluation on managed infrastructure > You use built-in containers or bring your own, and submit jobs through the CreateProcessingJob API. Build :: Environments for developing models > You can develop in an IDE, a no-code tool, or start from a pretrained model. + Machine learning environments SageMaker Studio :: Web-based ML experience with Code Editor, JupyterLab and RStudio > The previous Studio experience is now called Studio Classic. Code Editor is based on Code-OSS (Visual Studio Code - Open Source). SageMaker Canvas :: No-code ML for predictions, with chat with LLMs and ready-to-use models > Canvas lets business users build predictions such as churn or inventory plans without writing code. JumpStart :: Pretrained open-source models, solution templates and example notebooks > You can deploy, fine-tune and evaluate pretrained models from popular model hubs on the Models page in Studio. Train :: Managed training jobs and clusters > SageMaker Training containerizes the workload and manages the compute. Pick a feature based on whether you train a small model, a large distributed model or a persistent cluster. + Model training Training jobs :: Fully managed training that sets up and tears down infrastructure for you Managed Spot Training :: Use EC2 Spot instances, with checkpoints, to cut training cost > Managed Spot Training can reduce training cost by up to 90 percent compared with on-demand instances. Spot instances can be interrupted, so configure checkpoints that SageMaker copies to Amazon S3 so a job resumes where it stopped. Distributed training :: Data parallel and model parallel libraries for deep learning > The libraries scale training jobs, and you can also use PyTorch DistributedDataParallel, torchrun, MPI or a parameter server. HyperPod :: Resilient clusters for large-scale ML and foundation model development > HyperPod provisions clusters with thousands of accelerators such as AWS Trainium and NVIDIA GPUs. It monitors instances and automatically detects and replaces faulty hardware. Automatic model tuning :: Hyperparameter search with grid, random, Bayesian or Hyperband strategies > Tuning can run on managed Spot training to lower cost. Deploy :: Host models for inference > Match the endpoint type to the traffic: persistent for steady low-latency traffic, serverless for spiky traffic, asynchronous for large or slow requests, and batch transform when no endpoint is needed. + Deploy models for inference Real-time inference :: Persistent, fully managed endpoints for interactive, low-latency requests > Real-time endpoints support autoscaling and enhanced metrics for instances and containers. Serverless inference :: Scales to zero and fits intermittent traffic that tolerates cold starts > You do not choose instance types or scaling policies. Optional provisioned concurrency keeps endpoints warm for predictable bursts. Asynchronous inference :: Queues requests with payloads up to 1 GB and processing up to one hour > You can autoscale the instance count to zero when no requests are waiting, so you pay only while processing. Batch transform :: Offline predictions over a whole dataset without a persistent endpoint MLOps :: Automate and track the ML lifecycle > MLOps features automate workflows, version models and track runs so that results are reproducible. + Implement MLOps Pipelines :: Workflow orchestration for ML with a visual editor, SDK or JSON > SageMaker provisions, scales and shuts down the orchestration compute for you. Model Registry :: Catalog models, manage versions and approval status, and deploy > Model groups hold versions of a model. You can manage approval status and automate deployment with CI/CD. Experiments and MLflow :: Track experiments and monitor performance with managed MLflow > Managed MLflow targets generative AI development, helping you track experiments and observe models and AI applications. Govern and monitor :: Watch quality, document models and control access > Several governance features are closed to new customers, while Model Cards remain integrated with Model Registry. Model Monitor :: Monitors production model quality; closed to new customers > AWS points existing and new users to open-source SageMaker AI monitoring solutions with AWS Quick Sight and Amazon CloudWatch as a replacement. Existing customers can keep using Model Monitor. + Model Monitor availability change Clarify :: Fairness, bias detection and explainability; closed to new customers Model Cards :: Document model details, intended uses and risk ratings in one place > Model Cards integrate with Model Registry, so you can add auditing information when you register a model. Role Manager :: Persona-based IAM roles for ML work; closed to new customers

Amazon SageMaker AI mindmap