Model Booster Controller
Overview
As an optional component of the Kthena project, the Model Booster Controller primarily provides users with a convenient deployment form - ModelBooster. Based on the ModelBooster CR (ModelBooster Custom Resource) information provided by users, this component can automatically configure and deploy the components required for inference services, such as router routing rules, inference engine instances, and dynamic scaling configurations.
Feature Description
The Model Booster Controller provides one-stop deployment capabilities, enabling users to only need to be aware of ModelBooster information while ignoring the complex interaction specifications between various components within ModelServing, such as the use of some built-in labels and the binding relationships between CRDs. The following diagram illustrates the interaction principle of the Model Booster Controller:
The Model Booster Controller subscribes to the creation/modification/deletion events of the ModelBooster CR and synchronizes them to the CRs of ModelServer, ModelRoute, ModelServing, AutoscalingPolicy, and AutoscalingPolicyBinding that have the same semantic content as the corresponding events. All these CRs will carry the label registry.volcano.sh/managed-by=registry.volcano.sh.
Example
Read the examples to learn more.
Limitations
The ModelBooster can only cover most inference service scenarios and will be continuously updated. Some scenarios are not yet supported and require manual configuration of relevant CRDs.
Please note the following limitations when using the Model Booster:
- Each
ModelBoostercan create only oneModelRoute. - Rate limiting for
ModelRouteis not supported. - Topology configuration for
ModelServingis not supported. - The
panicPolicyconfiguration forAutoscalingPolicyis not supported. - Behavior configuration for
AutoscalingPolicyis not supported.
In these cases, you can manually create ModelServing, ModelServer, ModelRoute, AutoscalingPolicy, AutoscalingPolicyBinding resources as needed.