Category: Amazon SageMaker

Amazon SageMaker AI now supports EAGLE-based adaptive speculative decoding, a technique that accelerates large language model inference by up to 2.5x while maintaining output quality. In this post, we explain how to use EAGLE 2 and EAGLE 3 speculative decoding in Amazon SageMaker AI, covering the solution architecture, optimization workflows using your own datasets or SageMaker’s built-in data, and benchmark results demonstrating significant improvements in throughput and latency.

In this post, we demonstrate how to migrate computer vision workloads from Amazon Lookout for Vision to Amazon SageMaker AI by training custom defect detection models using pre-trained models available on AWS Marketplace. We provide step-by-step guidance on labeling datasets with SageMaker Ground Truth, training models with flexible hyperparameter configurations, and deploying them for real-time or batch inference—giving you greater control and flexibility for automated quality inspection use cases.

We’re introducing bidirectional streaming for Amazon SageMaker AI Inference, which transforms inference from a transactional exchange into a continuous conversation. This post shows you how to build and deploy a container with bidirectional streaming capability to a SageMaker AI endpoint. We also demonstrate how you can bring your own container or use our partner Deepgram’s pre-built models and containers on SageMaker AI to enable bi-directional streaming feature for real-time inference.

In this post, we explore how Amazon SageMaker HyperPod now supports NVIDIA Multi-Instance GPU (MIG) technology, enabling you to partition powerful GPUs into multiple isolated instances for running concurrent workloads like inference, research, and interactive development. By maximizing GPU utilization and reducing wasted resources, MIG helps organizations optimize costs while maintaining performance isolation and predictable quality of service across diverse machine learning tasks.

Amazon SageMaker HyperPod clusters with Amazon Elastic Kubernetes Service (EKS) orchestration now support creating and managing interactive development environments such as JupyterLab and open source Visual Studio Code, streamlining the ML development lifecycle by providing managed environments for familiar tools to data scientists. This post shows how HyperPod administrators can configure Spaces for their clusters, and how data scientists can create and connect to these Spaces.

In this post, we introduce the Multi-Provider Generative AI Gateway reference architecture, which provides guidance for deploying LiteLLM into an AWS environment to streamline the management and governance of production generative AI workloads across multiple model providers. This centralized gateway solution addresses common enterprise challenges including provider fragmentation, decentralized governance, operational complexity, and cost management by offering a unified interface that supports Amazon Bedrock, Amazon SageMaker AI, and external providers while maintaining comprehensive security, monitoring, and control capabilities.

In this post, you’ll learn how to deploy geospatial AI agents that can answer complex spatial questions in minutes instead of months. By combining Foursquare Spatial H3 Hub’s analysis-ready geospatial data with reasoning models deployed on Amazon SageMaker AI, you can build agents that enable nontechnical domain experts to perform sophisticated spatial analysis through natural language queries—without requiring geographic information system (GIS) expertise or custom data engineering pipelines.

In this post you will learn how to use Spectrum to optimize resource use and shorten training times without sacrificing quality, as well as how to implement Spectrum fine-tuning with Amazon SageMaker AI training jobs. We will also discuss the tradeoff between QLoRA and Spectrum fine-tuning, showing that while QLoRA is more resource efficient, Spectrum results in higher performance overall.

RoboTic-Tac-Toe is an interactive game where two physical robots move around a tic-tac-toe board, with both the gameplay and robots’ movements orchestrated by LLMs. Players can control the robots using natural language commands, directing them to place their markers on the game board. In this post, we explore the architecture and prompt engineering techniques used to reason about a tic-tac-toe game and decide the next best game strategy and movement plan for the current player.

This blog post introduces two major enhancements to Amazon SageMaker HyperPod that strengthen security and storage capabilities for large-scale machine learning infrastructure. The new features include customer managed key (CMK) support for encrypting EBS volumes with organization-controlled encryption keys, and Amazon EBS CSI driver integration that enables dynamic storage management for Kubernetes volumes in AI workloads.