#Release Notes

#AI 2.7.0

#New and Optimized Features

#Model as a Service (MaaS)

This release adds Model as a Service (MaaS), a managed model-serving gateway based on Alauda Build of Envoy Gateway and Alauda Build of Envoy AI Gateway. MaaS uses Authorino for authentication and lets administrators publish LLMInferenceService inference services and external model services through subscriptions.

Administrators can bind multiple models to each subscription, configure token limits for different time ranges, assign subscriptions to specified users, and view subscription traffic dashboards and usage reports. Users can access My Subscriptions to create, revoke, and delete API keys and view their own usage.

#Breaking Changes

#Fresh Installation Only

AI 2.7.0 does not support upgrade from AI 2.6.x or earlier versions. Existing environment migration is outside the scope of this Fast release and requires a later LTS release with a dedicated upgrade guide.

#Fixed Issues

• When injecting SecurityContext into the inference service Pod, KServe interferes with the configuration of the sidecar injected by Knative Serving, causing the Pod to fail to run due to configuration conflicts.
• Since model-catalog-ui prioritizes the "email" as the username when extracting a username from a token, if the user's email does not match their username, it may extract an incorrect username, which affects the relevant permission checks.

#Known Issues

• When using VictoriaMetrics for monitoring data collection of inference services operating in Serverless mode, there is a known issue where the inference services cannot scale down to zero.
• Because the name of the WorkspaceKind resource in a Workbench chart includes ".Release.Revision", when the chart is deployed multiple times, multiple instances of the same WorkspaceKind resource may be deployed due to the incrementing of ".Release.Revision".
• Because the HAMi scheduler component shares the same image as kube-scheduler, it may incorrectly identify the kube-scheduler image registry in the imported cluster, failing to pull the image and thus causing startup failures.