What's changed: Added per-section figures (cert-figure-retrofit). New AI-300 Chapter 3 (Domain 3 "GenAIOps infrastructure": Foundry environments = project/RBAC/managed identities/private networking/Bicep-CLI; foundation model deployment = serverless API endpoints vs managed compute/use-case model selection/versioning-deployment strategies/PTU; prompt versioning = design-develop/variant comparison/Git)
3.2Deploying and managing foundation models
Understand deploying foundation models via serverless API endpoints and managed compute, selecting models for use cases, model versioning and production deployment strategies, and configuring provisioned throughput units (PTU) for high-volume workloads.
Generative AI apps deploy foundation models chosen from the catalog. Deployment method, model selection, and securing throughput are the operational essentials.
3.2.1Deployment method and model selection
There are two deployment methods: serverless API endpoints (pay-as-you-go; call as an API without managing infrastructure) and managed compute (deploy to dedicated compute—when you need finer control or isolation). Model selection depends on the use case (required quality, cost, latency, context length, multimodality). The basics: "requirements first → the right model and deployment method."
3.2.2Versioning and PTU
Models also need versioning and production deployment strategies (e.g., phased switchover). Because models are updated and retired, deploy to pinned versions and plan switchovers. For high volume needing stable low latency and predictable cost, reserve capacity with provisioned throughput units (PTU) (better for heavy load than pay-as-you-go). Low/variable load fits pay-as-you-go (serverless).
| Aspect | Option | When to use |
|---|---|---|
| Deployment (lightweight) | Serverless API endpoint | Pay-as-you-go; no infra mgmt |
| Deployment (dedicated) | Managed compute | Need finer control/isolation |
| Model selection | Use-case criteria | Quality/cost/latency/context |
| Capacity for high volume | PTU (provisioned) | Stable low latency, predictable cost |
| Low/variable load | Pay-as-you-go (serverless) | Pay per use |
Watch the mix-ups: (1) serverless API endpoint (pay-as-you-go, no mgmt) vs managed compute (dedicated, fine control). (2) PTU (reserved capacity = stable under heavy load) vs pay-as-you-go (for variable load). (3) Models are updated/retired, so plan versioning and switchovers.
Map requirement → choice: "call a model via API without managing infra" = serverless API endpoint; "fine control on dedicated resources" = managed compute; "stable low latency and predictable cost at high volume" = PTU; "pay per use for low/variable load" = pay-as-you-go; "pick a model that fits the use case" = use-case-based selection.
3.2.3Section summary
- Deploy: serverless API endpoint (pay-as-you-go/no mgmt) vs managed compute (dedicated/control)
- Select models by use case (quality/cost/latency/context length/multimodality)
- High volume → reserve with PTU; variable load → pay-as-you-go; version models and plan switchovers
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which deployment method calls a foundation model as an API, pay-as-you-go, without managing infrastructure?
Q2. Which reserves capacity for stable low latency and predictable cost in high-volume workloads?
Q3. Which method deploys a model to dedicated compute when finer control or isolation is needed?
Q4. Which is the most appropriate basis for selecting a foundation model?
Q5. Because foundation models are updated and retired, what is needed in production?
Q6. Which correctly distinguishes PTU from pay-as-you-go (serverless)?