Skip to content

Getting started

Subscribe to the model package, configure your own SageMaker resources and run a small fictional passage pool before using your application data.

Subscribe and select the regional package

  1. Open the Marketplace listing. Review software pricing, trial, EULA and refund terms before accepting the offer.
  2. After subscribing, choose Continue to configuration and select us-east-1.
  3. Copy the product/model-package ARN supplied for that region. Use the subscribed package ARN for deployment.
  4. Download the buyer notebook. Use a current boto3 version and your own SageMaker execution role. Do not put access keys in the notebook.

Set the ARN in your own notebook environment, or assign it to PACKAGE_ARN in the configuration cell:

export SAGEMAKER_MARKETPLACE_MODEL_PACKAGE_ARN='<regional-subscribed-model-package-arn>'

All creation, invocation, batch and cleanup flags are disabled by default. The notebook validates the completed package, its Marketplace marker and published image identity before creating resources. You need the corresponding SageMaker instance quotas; quota does not guarantee capacity.

Create a real-time endpoint

Use one ml.g5.xlarge, network isolation and this explicit setting on its production variant:

InferenceAmiVersion="al2023-ami-sagemaker-inference-gpu-4-1"

The default legacy g5 driver cannot run this CUDA 13 runtime. Follow the notebook's deployment cell and set CREATE_RESOURCES=True only after checking your role, package ARN and costs. After the endpoint is ready, enable INVOKE_ENDPOINT for the fictional request.

Run batch inference

Download sample_batch.jsonl and upload it to a private encrypted S3 input location in your account. Configure a separate output prefix and your role's S3 permissions.

The notebook includes a disabled batch example for ml.g6e.xlarge. Its required settings are:

BatchStrategy="SingleRecord"
MaxConcurrentTransforms=1
TransformAmiVersion="al2-ami-sagemaker-batch-gpu-535"
Environment={"SAGEMAKER_CUDA_COMPAT_DISABLED": "true"}

TransformAmiVersion belongs in TransformResources. Use SplitType="Line", AssembleWith="Line" and JSON Lines content/accept types. Set START_BATCH=True only after configuring your input and output locations. Preserve input/output line order and inspect failed jobs.

Clean up

A waiter timeout does not stop AWS compute. Inspect the exact endpoint or batch job after a timeout. The notebook includes explicit cleanup flags and identity checks for the resources it creates. Endpoint deletion releases hosting compute; stopping an unfinished batch job releases its running compute. S3 inputs and outputs are preserved by cleanup.

Qualification measurements are for us-east-1. Other regions shown in the catalog have not been qualified by Software Sushi.

Last updated November 15, 2026 · Suggest an improvement