import * as AWS from "@alchemy.run/aws";
/**
* A minimal SageMaker HyperPod training cluster:
*
* - a lifecycle-script bucket + instance execution role (`src/infra.ts`),
* - a deploy-time Action that uploads `on_create.sh` to the bucket
* (`src/lifecycle.ts`) — bucket → script → cluster ordering is inferred
* from the data flow,
* - a Slurm-orchestrated `AWS.SageMaker.Cluster` with one instance group.
*
* The demo instance group is a single ml.t3.medium to keep the example
* cheap; swap in ml.g5/ml.p5 groups (and more of them) for real training.
* Provisioning takes ~5 minutes at this size, 10-25 minutes for large GPU
* groups.
*/
import * as Alchemy from "alchemy";
import * as Effect from "effect/Effect";
import { HyperPodInfra } from "./src/infra.ts";
import { UploadLifecycleScript } from "./src/lifecycle.ts";
export default Alchemy.Stack(
"AwsHyperPodExample",
{
providers: AWS.providers(),
state: Alchemy.localState(),
},
Effect.gen(function* () {
const { bucket, role } = yield* HyperPodInfra;
// Upload the on_create script; the Action's output feeds the cluster's
// LifeCycleConfig, so the script is in place before the cluster exists.
const script = yield* UploadLifecycleScript({
bucketName: bucket.bucketName,
});
const cluster = yield* AWS.SageMaker.Cluster("TrainingCluster", {
instanceGroups: {
controller: {
InstanceType: "ml.t3.medium",
InstanceCount: 1,
ExecutionRole: role.roleArn,
LifeCycleConfig: {
SourceS3Uri: script.sourceS3Uri,
OnCreate: script.onCreate,
},
},
},
// Automatic recovery replaces faulty nodes; "None" is fine for a demo
// controller group and avoids replacement churn on tiny instances.
nodeRecovery: "None",
tags: { app: "aws-hyperpod-example" },
});
return {
clusterName: cluster.clusterName,
clusterArn: cluster.clusterArn,
clusterStatus: cluster.clusterStatus,
lifecycleBucket: bucket.bucketName,
};
}),
);