AI Model Deployment

Overview

The model runtime workflow (.github/workflows/deploy-model-runtime.yml) deploys AI models to Azure AI Foundry and can seed the runtime model tables from the generated deployment manifest.

The environment bootstrap path invokes this workflow automatically for the stagingBootstrap promotion target. It deploys the certified staging baseline (gpt-5.3-chat, gpt-5.4-mini, and gpt-5.4-nano) before database publication, seeds the generated manifest after the DACPAC is available, and certifies the seeded deployments during normal contract validation. Manifest reconciliation disables model capabilities that are not in that baseline. Anthropic models remain available through an explicit Deploy Model Runtime run after their target tenant and regions pass certification.

Use it when:

• adding a new model to an existing environment
• updating quota allocations
• recreating model deployments after an AI Services account changes
• testing a filtered set of models before broader rollout

You do not need to run it separately when bootstrapping a supported staging environment through the Environment Bootstrap workflow.

How To Run

1. Open the Deploy Model Runtime GitHub Actions workflow.
2. Choose project, targetEnv, models, quota, deployment scope, and dry-run settings.
3. Run with dryRun=true first.
4. Run with dryRun=false and seedModelRuntime=true to deploy and seed runtime tables.

Key Inputs

Input	Description
project	Project profile such as testing, tap, k12, or sdac.
targetEnv	Target environment.
models	Comma-separated model names to deploy.
quotaPerDeployment	Quota per deployment in K TPM.
deploymentScope	Deployment naming pattern to target.
updateExistingOnly	Only update existing deployments.
seedModelRuntime	Seed runtime tables after deployment.
dryRun	Preview deployment operations without applying them.

Runtime Seeding

When seeding is enabled, the workflow downloads the model manifest, generates SQL with azure_template/scripts/generate-llm-seed.sh, and writes Core.LLMParamsDim and Core.LLMModelCapabilities rows.

After model runtime seeding, deployment contract validation should certify model parameters before the dashboard exposes editable controls.