AWS: SageMaker & Bedrock News
AWS rolled out SageMaker inference revisions including multi‑variant endpoints for A/B testing and better distributed inference patterns, while Bedrock continues to be presented as a single interface for multiple foundation models—helping teams swap and test models more easily in production summarized, summarized. These updates smooth progressive deployments and model evaluation loops for SaaS ML teams.
SageMaker’s production-variant model routing supports per-request targeting via the TargetVariant parameter, which overrides any configured traffic weights so a single InvokeEndpoint call can exercise a specific model version for logging and comparison. (docs.aws.amazon.com) The SageMaker Inference Recommender automates cross-instance benchmarking and load testing to recommend an optimized instance type and endpoint configuration, reducing what used to take weeks of manual tuning to hours. (sagemaker-examples.readthedocs.io) SageMaker hosting now codifies progressive-deployment guardrails — shadow tests, blue/green and canary traffic-shifting options, and multi‑AZ instance deployment — so teams can validate shadow variant responses and roll back at the endpoint-configuration level. (docs.aws.amazon.com) Amazon Bedrock’s Marketplace exposes over 100 foundation models through a single Bedrock API and console, with one-click enablement and consolidated billing so enterprises can compare and provision vendors like Anthropic, Meta, and Amazon Nova from the same interface. (aws.amazon.com) Bedrock surfaces model lifecycle metadata via GetFoundationModel/ListFoundationModels and keeps new model versions “Active” for at least 12 months before EOL, enabling predictable migration windows for production migrations and RAG index compatibility planning. (docs.aws.amazon.com) AWS’s Trainium + Cerebras CS‑3 partnership for Bedrock implements “inference disaggregation” (prefill on Trainium, decode on CS‑3) wired with Elastic Fabric Adapter (EFA), a design AWS says will deliver order‑of‑magnitude inference speedups for interactive LLM workloads. (press.aboutamazon.com)
