Detailed Explanation

Based on AWS documentation and best practices for Amazon Bedrock, the correct answer is A: Purchase Provisioned Throughput for the custom model.

Why Option A is Correct

1. AWS Documentation Requirement: According to the official AWS Bedrock documentation, "Before you can use a customized model, you need to purchase Provisioned Throughput for it." This is a mandatory step for using any custom model (fine-tuned or imported) through Amazon Bedrock.

2. Provisioned Throughput Purpose: Provisioned Throughput in Amazon Bedrock guarantees a specific level of inference capacity for your custom model. It ensures predictable performance and availability when invoking the model through Bedrock's APIs.

3. Prerequisite for Inference: Without purchasing Provisioned Throughput, you cannot invoke or use your custom model through Amazon Bedrock, regardless of whether the model was fine-tuned within Bedrock or imported externally.

Why Other Options Are Incorrect

Option B: Deploy the custom model in an Amazon SageMaker endpoint for real-time inference

• This is unnecessary when using Amazon Bedrock. Bedrock manages the deployment and hosting of custom models internally. Deploying to SageMaker would create a separate inference endpoint outside of Bedrock's managed service.

Option C: Register the model with the Amazon SageMaker Model Registry

• The SageMaker Model Registry is for managing models within SageMaker workflows. For Bedrock custom models, registration happens within Bedrock itself, not through SageMaker.

Option D: Grant access to the custom model in Amazon Bedrock

• While access control is important for security, it is not the fundamental requirement for using a custom model. Access permissions can be configured, but they don't enable the model's availability for inference. The mandatory step is purchasing Provisioned Throughput, after which access can be managed through IAM policies.

Key Distinction

The critical distinction is that Provisioned Throughput (Option A) is the enabling requirement that makes the custom model available for inference through Bedrock. Access control (Option D) is a security configuration that can be applied afterward but doesn't by itself make the model usable.

Best Practice Consideration

For production workloads using custom models in Bedrock, companies should:

1. First purchase Provisioned Throughput for the custom model
2. Then configure appropriate access controls via IAM
3. Use the model ARN to invoke inference through Bedrock APIs

This approach aligns with AWS's managed service model where Bedrock handles the infrastructure, scaling, and deployment complexities.