 |
As we speak, we’re happy to announce Amazon Elastic Kubernetes Service (EKS) assist in Amazon SageMaker HyperPod — purpose-built infrastructure engineered with resilience at its core for basis mannequin (FM) growth. This new functionality allows clients to orchestrate HyperPod clusters utilizing EKS, combining the facility of Kubernetes with Amazon SageMaker HyperPod‘s resilient atmosphere designed for coaching massive fashions. Amazon SageMaker HyperPod helps effectively scale throughout greater than a thousand synthetic intelligence (AI) accelerators, decreasing coaching time by as much as 40%.
Amazon SageMaker HyperPod now allows clients to handle their clusters utilizing a Kubernetes-based interface. This integration permits seamless switching between Slurm and Amazon EKS for optimizing varied workloads, together with coaching, fine-tuning, experimentation, and inference. The CloudWatch Observability EKS add-on supplies complete monitoring capabilities, providing insights into CPU, community, disk, and different low-level node metrics on a unified dashboard. This enhanced observability extends to useful resource utilization throughout all the cluster, node-level metrics, pod-level efficiency, and container-specific utilization knowledge, facilitating environment friendly troubleshooting and optimization.
Launched at re:Invent 2023, Amazon SageMaker HyperPod has turn into a go-to resolution for AI startups and enterprises seeking to effectively practice and deploy massive scale fashions. It’s appropriate with SageMaker’s distributed coaching libraries, which supply Mannequin Parallel and Knowledge Parallel software program optimizations that assist scale back coaching time by as much as 20%. SageMaker HyperPod robotically detects and repairs or replaces defective cases, enabling knowledge scientists to coach fashions uninterrupted for weeks or months. This enables knowledge scientists to concentrate on mannequin growth, slightly than managing infrastructure.
The combination of Amazon EKS with Amazon SageMaker HyperPod makes use of some great benefits of Kubernetes, which has turn into fashionable for machine studying (ML) workloads because of its scalability and wealthy open-source tooling. Organizations typically standardize on Kubernetes for constructing functions, together with these required for generative AI use instances, because it permits reuse of capabilities throughout environments whereas assembly compliance and governance requirements. As we speak’s announcement allows clients to scale and optimize useful resource utilization throughout greater than a thousand AI accelerators. This flexibility enhances the developer expertise, containerized app administration, and dynamic scaling for FM coaching and inference workloads.
Amazon EKS assist in Amazon SageMaker HyperPod strengthens resilience by way of deep well being checks, automated node restoration, and job auto-resume capabilities, making certain uninterrupted coaching for giant scale and/or long-running jobs. Job administration might be streamlined with the non-compulsory HyperPod CLI, designed for Kubernetes environments, although clients may use their very own CLI instruments. Integration with Amazon CloudWatch Container Insights supplies superior observability, providing deeper insights into cluster efficiency, well being, and utilization. Moreover, knowledge scientists can use instruments like Kubeflow for automated ML workflows. The combination additionally contains Amazon SageMaker managed MLflow, offering a sturdy resolution for experiment monitoring and mannequin administration.
At a excessive degree, Amazon SageMaker HyperPod cluster is created by the cloud admin utilizing the HyperPod cluster API and is totally managed by the HyperPod service, eradicating the undifferentiated heavy lifting concerned in constructing and optimizing ML infrastructure. Amazon EKS is used to orchestrate these HyperPod nodes, just like how Slurm orchestrates HyperPod nodes, offering clients with a well-recognized Kubernetes-based administrator expertise.
Let’s discover the way to get began with Amazon EKS assist in Amazon SageMaker HyperPod
I begin by making ready the situation, checking the stipulations, and creating an Amazon EKS cluster with a single AWS CloudFormation stack following the Amazon SageMaker HyperPod EKS workshop, configured with VPC and storage assets.
To create and handle Amazon SageMaker HyperPod clusters, I can use both the AWS Administration Console or AWS Command Line Interface (AWS CLI). Utilizing the AWS CLI, I specify my cluster configuration in a JSON file. I select the Amazon EKS cluster created beforehand because the orchestrator of the SageMaker HyperPod Cluster. Then, I create the cluster employee nodes that I name “worker-group-1”, with a personal Subnet, NodeRecovery set to Automated to allow automated node restoration and for OnStartDeepHealthChecks I add InstanceStress and InstanceConnectivity to allow deep well being checks.
cat > eli-cluster-config.json << EOL
{
"ClusterName": "example-hp-cluster",
"Orchestrator": {
"Eks": {
"ClusterArn": "${EKS_CLUSTER_ARN}"
}
},
"InstanceGroups": [
{
"InstanceGroupName": "worker-group-1",
"InstanceType": "ml.p5.48xlarge",
"InstanceCount": 32,
"LifeCycleConfig": {
"SourceS3Uri": "s3://${BUCKET_NAME}",
"OnCreate": "on_create.sh"
},
"ExecutionRole": "${EXECUTION_ROLE}",
"ThreadsPerCore": 1,
"OnStartDeepHealthChecks": [
"InstanceStress",
"InstanceConnectivity"
],
},
....
],
"VpcConfig": {
"SecurityGroupIds": [
"$SECURITY_GROUP"
],
"Subnets": [
"$SUBNET_ID"
]
},
"ResilienceConfig": {
"NodeRecovery": "Automated"
}
}
EOL
You’ll be able to add InstanceStorageConfigs to provision and mount a further Amazon EBS volumes on HyperPod nodes.
To create the cluster utilizing the SageMaker HyperPod APIs, I run the next AWS CLI command:
aws sagemaker create-cluster
--cli-input-json file://eli-cluster-config.jsonThe AWS command returns the ARN of the brand new HyperPod cluster.
{
"ClusterArn": "arn:aws:sagemaker:us-east-2:ACCOUNT-ID:cluster/wccy5z4n4m49"
}I then confirm the HyperPod cluster standing within the SageMaker Console, awaiting till the standing modifications to InService.
And I can monitor cluster efficiency and well being metrics utilizing Amazon CloudWatch Container Insights.
Issues to know
Listed here are some key issues you need to learn about Amazon EKS assist in Amazon SageMaker HyperPod:
Resilient Setting – This integration supplies a extra resilient coaching atmosphere with deep well being checks, automated node restoration, and job auto-resume. SageMaker HyperPod robotically detects, diagnoses, and recovers from faults, permitting you to repeatedly practice basis fashions for weeks or months with out disruption. This will scale back coaching time by as much as 40%.
Enhanced GPU Observability – Amazon CloudWatch Container Insights supplies detailed metrics and logs in your containerized functions and microservices. This permits complete monitoring of cluster efficiency and well being.
Scientist-Pleasant Software – This launch features a customized HyperPod CLI for job administration, Kubeflow Coaching Operators for distributed coaching, Kueue for scheduling, and integration with SageMaker Managed MLflow for experiment monitoring. It additionally works with SageMaker’s distributed coaching libraries, which give Mannequin Parallel and Knowledge Parallel optimizations to considerably scale back coaching time. These libraries, mixed with auto-resumption of jobs, allow environment friendly and uninterrupted coaching of enormous fashions.
Versatile Useful resource Utilization – This integration enhances developer expertise and scalability for FM workloads. Knowledge scientists can effectively share compute capability throughout coaching and inference duties. You should use your current Amazon EKS clusters or create and connect new ones to HyperPod compute, deliver your personal instruments for job submission, queuing and monitoring.
To get began with Amazon SageMaker HyperPod on Amazon EKS, you possibly can discover assets such because the SageMaker HyperPod EKS Workshop, the aws-do-hyperpod challenge, and the awsome-distributed-training challenge. This launch is usually out there within the AWS Areas the place Amazon SageMaker HyperPod is obtainable besides Europe(London). For pricing info, go to the Amazon SageMaker Pricing web page.
This weblog submit was a collaborative effort. I want to thank Manoj Ravi, Adhesh Garg, Tomonori Shimomura, Alex Iankoulski, Anoop Saha, and all the workforce for his or her important contributions in compiling and refining the knowledge introduced right here. Their collective experience was essential in creating this complete article.
– Eli.