The customization lifecycle: adapters, registry, rollback
Lesson 4 of 4 in Designing GenAI Solutions and Selecting Models.
When prompting and retrieval stop being enough, models get customized — and what AIP-C01 tests is the lifecycle, not the training (how tuning works lives on the MLA path; here it is out of scope). The professional vocabulary: parameter-efficient adaptation — PEFT, headlined by LoRA — trains small low-rank matrices against a frozen base model, producing adapter artifacts measured in megabytes that are cheap to train, quick to deploy, and swappable per task. Deployment lands on SageMaker AI endpoints for domain-specific fine-tuned models (or back into Bedrock via Custom Model Import for compatible open weights).
Around the artifact sits the machinery the exam actually asks about: the SageMaker Model Registry versions every candidate with an approval status; automated pipelines promote approved versions; rollback strategies keep the previous version deployable in minutes; and lifecycle management retires and replaces models on schedule instead of by surprise.
| Option | Artifact & cost shape | Serving | Rollback story |
|---|---|---|---|
Prompt + retrieval only | No training; templates versioned in Prompt Management | Any hosted FM, unchanged | Instant — switch template version |
MB-scale adapters; low training cost; one per task on a shared base | SageMaker AI endpoints for domain-tuned models | Fast — registry keeps every adapter version; redeploy the last good one | |
Full fine-tune | Full model copy; highest cost; risks forgetting | Bedrock custom model or SageMaker endpoint | Heavier — versions are whole models, so storage and deploys cost more |
Open weights you already own, brought to Bedrock | Bedrock serverless API, like any hosted FM | Re-point to the previously imported model version |
Key terms: LoRA, PEFT, SageMaker Model Registry, Custom Model Import, Fine-tuning
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.