Values_0.CreateClusterRequestSourceCreates an Amazon SageMaker HyperPod cluster. SageMaker HyperPod is a capability of SageMaker for creating and managing persistent clusters for developing large machine learning models, such as large language models (LLMs) and diffusion models. To learn more, see Amazon SageMaker HyperPod in the Amazon SageMaker Developer Guide.
type nonrec t = {clusterName : ClusterName.t;The name for the new SageMaker HyperPod cluster.
*)instanceGroups : ClusterInstanceGroupSpecifications.t option;The instance groups to be created in the SageMaker HyperPod cluster.
*)restrictedInstanceGroups : ClusterRestrictedInstanceGroupSpecifications.t
option;The specialized instance groups for training models like Amazon Nova to be created in the SageMaker HyperPod cluster.
*)vpcConfig : VpcConfig.t option;Specifies the Amazon Virtual Private Cloud (VPC) that is associated with the Amazon SageMaker HyperPod cluster. You can control access to and from your resources by configuring your VPC. For more information, see Give SageMaker access to resources in your Amazon VPC. When your Amazon VPC and subnets support IPv6, network communications differ based on the cluster orchestration platform: Slurm-orchestrated clusters automatically configure nodes with dual IPv6 and IPv4 addresses, allowing immediate IPv6 network communications. In Amazon EKS-orchestrated clusters, nodes receive dual-stack addressing, but pods can only use IPv6 when the Amazon EKS cluster is explicitly IPv6-enabled. For information about deploying an IPv6 Amazon EKS cluster, see Amazon EKS IPv6 Cluster Deployment. Additional resources for IPv6 configuration: For information about adding IPv6 support to your VPC, see to IPv6 Support for VPC. For information about creating a new IPv6-compatible VPC, see Amazon VPC Creation Guide. To configure SageMaker HyperPod with a custom Amazon VPC, see Custom Amazon VPC Setup for SageMaker HyperPod.
*)orchestrator : ClusterOrchestrator.t option;The type of orchestrator to use for the SageMaker HyperPod cluster. Currently, supported values are "Eks" and "Slurm", which is to use an Amazon Elastic Kubernetes Service or Slurm cluster as the orchestrator. If you specify the Orchestrator field, you must provide exactly one orchestrator configuration: either Eks or Slurm. Specifying both or providing an empty configuration returns a validation error.
*)nodeRecovery : ClusterNodeRecovery.t option;The node recovery mode for the SageMaker HyperPod cluster. When set to Automatic, SageMaker HyperPod will automatically reboot or replace faulty nodes when issues are detected. When set to None, cluster administrators will need to manually manage any faulty cluster instances.
*)tieredStorageConfig : ClusterTieredStorageConfig.t option;The configuration for managed tier checkpointing on the HyperPod cluster. When enabled, this feature uses a multi-tier storage approach for storing model checkpoints, providing faster checkpoint operations and improved fault tolerance across cluster nodes.
*)nodeProvisioningMode : ClusterNodeProvisioningMode.t option;The mode for provisioning nodes in the cluster. You can specify the following modes: Continuous: Scaling behavior that enables 1) concurrent operation execution within instance groups, 2) continuous retry mechanisms for failed operations, 3) enhanced customer visibility into cluster events through detailed event streams, 4) partial provisioning capabilities. Your clusters and instance groups remain InService while scaling. This mode is only supported for EKS orchestrated clusters.
*)clusterRole : RoleArn.t option;The Amazon Resource Name (ARN) of the IAM role that HyperPod assumes to perform cluster autoscaling operations. This role must have permissions for sagemaker:BatchAddClusterNodes and sagemaker:BatchDeleteClusterNodes. This is only required when autoscaling is enabled and when HyperPod is performing autoscaling operations.
*)autoScaling : ClusterAutoScalingConfig.t option;The autoscaling configuration for the cluster. Enables automatic scaling of cluster nodes based on workload demand using a Karpenter-based system.
*)}val make :
?instanceGroups:??? ->
?restrictedInstanceGroups:??? ->
?vpcConfig:??? ->
?tags:??? ->
?orchestrator:??? ->
?nodeRecovery:??? ->
?tieredStorageConfig:??? ->
?nodeProvisioningMode:??? ->
?clusterRole:??? ->
?autoScaling:??? ->
clusterName:ClusterName.t ->
unit ->
tval to_value :
t ->
[> `Structure of
(string
* [> `Enum of string
| `List of
[> `Structure of
(string
* [> `Enum of string
| `Integer of ClusterInstanceCount.t
| `List of
[> `Enum of string
| `Structure of
(string
* [> `Structure of
(string
* [> `Boolean of Boolean.t
| `Integer of ClusterEbsVolumeSizeInGB.t
| `String of KmsKeyId.t ])
list ])
list ]
list
| `String of ClusterInstanceGroupName.t
| `Structure of
(string
* [> `Enum of string
| `List of
[> `Enum of string
| `String of SecurityGroupId.t
| `Structure of
(string
* [> `Enum of string
| `String of ClusterKubernetesTaintKey.t ])
list ]
list
| `Map of
([> `String of ClusterKubernetesLabelKey.t ]
* [> `String of ClusterKubernetesLabelValue.t ])
list
| `String of S3Uri.t
| `Structure of
(string
* [> `Integer of WaitTimeIntervalInSeconds.t
| `List of
[> `Structure of
(string * [> `String of AlarmName.t ])
list ]
list
| `Structure of
(string
* [> `Structure of
(string
* [> `Enum of string
| `Integer of
NodeUnavailabilityValue.t ])
list ])
list ])
list ])
list ])
list ]
list
| `String of ClusterName.t
| `Structure of
(string
* [> `Enum of string
| `Integer of ClusterInstanceMemoryAllocationPercentage.t
| `List of [> `String of SecurityGroupId.t ] list
| `Structure of
(string * [> `Enum of string | `String of EksClusterArn.t ])
list ])
list ])
list ]