Getting started¶
Subscribe and choose your own regional package and execution role before creating GPU resources. Begin with a small fictional source segment.
Subscribe and choose deployment¶
- Review the Marketplace offer, version 0.1.0, region, software terms and separate GPU infrastructure charges.
- Subscribe, choose Continue to configuration, select us-east-1 and copy the subscribed model-package ARN.
- Create a model in your account with that ARN, your execution role and
EnableNetworkIsolation=True. Use ml.g6e.4xlarge for a real-time endpoint or ml.g6e.2xlarge for Batch Transform. - Invoke the ready endpoint with the bounded request, or use line-split JSON Lines for batch.
Your account needs the corresponding instance quotas; quota does not guarantee capacity. Allow for model startup: the qualification cold deployment took about 21 minutes. A waiter timeout does not stop compute.
Buyer examples¶
The public usage guide, notebook and fictional examples explain the released contract. Review every configuration and creation cell before running it. Set GROUNDED_SUMMARY_MODEL_PACKAGE_ARN to your subscribed regional ARN. Default Run All is offline. After reviewing permissions, costs and cleanup, set RUN_DEMO = True to enable the real-time demonstration. Enable RUN_BATCH = True only if you want the batch example too. Keep RUN_DEMO enabled until cleanup finishes. The notebook verifies the exact release-approved image digest; a buyer-account subscription transaction has not been rehearsed by the seller. See the shared guide for caller and execution-role responsibilities.
The notebook's demonstration can create a fresh private encrypted S3 bucket, or validate a supplied bucket without reconfiguring it. These examples do not remove your responsibility for AWS costs and storage retention. Do not put credentials or private source text in public notebooks or support messages.
Batch and cleanup¶
Use one complete JSON object per line, BatchStrategy="SingleRecord", SplitType="Line" and AssembleWith="Line". Keep S3 locations private and authorized to your execution role.
Delete the exact temporary endpoint, endpoint configuration and model after testing. Stop unfinished batch jobs when appropriate. Closing a notebook or timing out a waiter does not stop hosting. Retain or remove input/output evidence according to your policy; preserve shared buckets.
Last updated November 15, 2026 · Suggest an improvement