View a markdown version of this page

Accelerated inference - Amazon SageMaker AI

Accelerated inference

Ray Serve is the open source model-serving library for Ray. You deploy a model as a Serve application, and Ray Serve handles request routing, batching, and replica scaling. On HyperPod, Ray Serve runs on KubeRay through a RayService resource, so your Serve code and deployment graph are unchanged.

Requirements

Ray Serve on HyperPod requires only the KubeRay operator. KubeRay reconciles the RayService resource, creates the Ray cluster that backs it, and manages the Serve applications on it. For more information, see Installing KubeRay on HyperPod Amazon EKS.

You do not install a separate inference operator to serve models with Ray Serve. The JumpStart deployment path additionally uses the toolkit-for-ray-on-sagemaker-ai library, which you install with pip.

What you can do

  • Deploy a model from your own code as a RayService and reach it over HTTP.

  • Deploy an JumpStart model into a Ray Serve deployment.

  • Lower time to first token for long-context and multi-turn workloads with a managed tiered KV cache and prefix-aware routing.

  • Scale serving replicas within the cluster, and scale cluster capacity with managed Karpenter.

For the Ray Serve API and deployment concepts, see Ray Serve: Scalable and Programmable Serving in the Ray documentation.