Amazon SageMaker HyperPod now supports flexible instance groups

Amazon Web Services has enhanced its SageMaker HyperPod service with flexible instance groups, allowing customers to specify multiple instance types and subnets within a single instance group configuration. Previously, users had to create separate instance groups for every combination of instance type and availability zone, creating operational overhead in cluster management, scaling, patching, and monitoring tasks. The new feature introduces an InstanceRequirements parameter that accepts an ordered list of instance types, with HyperPod automatically provisioning the highest-priority instances first and falling back to lower-priority options when capacity is unavailable. Training workloads benefit from multi-subnet distribution within availability zones to prevent subnet exhaustion, while inference workloads gain automatic priority-based fallback without manual intervention across instance groups. The flexible instance groups are available for SageMaker HyperPod clusters using the EKS orchestrator across all AWS regions where the service operates. Customers can configure these groups through the CreateCluster and UpdateCluster APIs, AWS CLI, or the AWS Management Console, with Karpenter autoscaling integration automatically detecting supported instance types and provisioning optimal resources based on pod requirements.