A serving model is slow because it reloads large artifacts for every request. Which approach best fits?

1. Use an appropriately provisioned online prediction endpoint.
2. Load and reuse model state appropriately within the serving runtime.
3. Configure suitable autoscaling and validate peak-load behavior.
4. Use supported batch prediction.