Principle:Kserve Kserve TrainedModel Registration

Knowledge Sources	KServe Multi-Model Serving
Domains	MLOps, Model_Management, Kubernetes
Last Updated	2026-02-13 00:00 GMT

Overview

A model registration mechanism that binds individual trained model artifacts to a shared InferenceService using the TrainedModel custom resource.

Description

TrainedModel Registration uses the TrainedModel CRD to declare that a specific model artifact should be loaded into a shared InferenceService. Each TrainedModel specifies:

• The parent InferenceService name
• The model storage URI
• The ML framework
• The estimated memory requirement

The TrainedModel controller validates that the parent ISVC exists, is in MMS mode, and has sufficient memory. It then writes the model configuration to the shared ConfigMap, triggering the model agent to download and load the model.

Usage

Use this after deploying a shared InferenceService (without storageUri). Create one TrainedModel resource per model you want to load.

Theoretical Basis

# Registration flow (NOT implementation code)
1. Create TrainedModel with parent ISVC reference
2. Controller validates:
   a. Parent ISVC exists and is ready
   b. Parent ISVC is MMS (IsMMSPredictor)
   c. Total memory of all TrainedModels ≤ ISVC memory limit
3. Write model spec to ConfigMap models.json
4. Agent detects ConfigMap change → downloads model → calls V2 load API

Related Pages

Implemented By

• Implementation:Kserve_Kserve_TrainedModel_CRD_Spec