Databricks Generative AI Engineer Associate Question 175
Select 3You are tasked with deploying a generative AI model that was fine-tuned on Databricks. The model will serve real-time predictions via an API. Which of the following steps must be included to successfully assemble and deploy this application?
- A
Package the fine-tuned model and its dependencies into a containerized environment.
- B
Configure a Databricks MLflow Model Serving endpoint for the fine-tuned model.
- C
Train the model from scratch to optimize it for container deployment.
- D
Set up an API layer to communicate with the deployed model for real-time predictions.
- E
Deploy the model directly without adding any dependency management.
Show answer and explanation
Correct answers: A, B, D
Explanation
Deploying a generative AI model involves preparing the model and its dependencies for a production environment, hosting it for inference, and providing an API layer for real-time communication. Packaging the model into a container, configuring MLflow Model Serving, and setting up an API layer are critical steps in this process. Training the model again or skipping dependency management are not required or advisable.
- A. Correct.
Correct. Packaging the model and its dependencies into a containerized environment, such as a Docker image, ensures that the application is portable and can run in various deployment environments.
- B. Correct.
Correct. Configuring an MLflow Model Serving endpoint enables the model to be hosted and accessed for real-time inference on Databricks.
- C. Incorrect.
Incorrect. Training the model from scratch is unnecessary for deployment, as the model has already been fine-tuned. Deployment focuses on operationalizing the existing model.
- D. Correct.
Correct. Setting up an API layer is essential for enabling external applications to send requests and receive predictions from the deployed model.
- E. Incorrect.
Incorrect. Deploying the model without dependency management would likely lead to runtime failures due to missing libraries or mismatched versions.