Skill 3.1.4: Evaluate and select appropriate foundation model (FM) deployment options
| Question | Answer & Explanation |
|---|---|
| ### Question 1 - Deploying Externally Fine-Tuned Open-Weight Models A machine learning team has fine-tuned an open-weight foundation model (Llama 3 8B) using their own training environment on Amazon EC2. They now need to serve this model in production through a managed API without provisioning or managing any underlying compute instances. The operational overhead must be minimized. Which deployment option BEST satisfies these requirements? (Select ONE) - A) Deploy the model artifacts to a SageMaker Real-Time Endpoint using a custom container. - B) Use Amazon Bedrock Custom Model Import to import the customized model artifacts from Amazon S3. - C) Package the model as a Docker container and run it as a long-lived service on Amazon ECS. - D) Deploy the base model from Amazon SageMaker JumpStart and apply the fine-tuned weights dynamically. |
Correct Answer: B Correct Answer Analysis: - B is Correct: Amazon Bedrock Custom Model Import allows organizations to import open-weight foundation models (like Llama or Mistral) that were customized or fine-tuned externally into Bedrock. Once imported from S3, the model is hosted on Bedrock's serverless infrastructure and served via the standard InvokeModel API, completely eliminating the need to manage endpoints or instances.Distractor Analysis: - A is Incorrect: A SageMaker Real-Time Endpoint requires selecting, provisioning, and managing compute instances, violating the constraint of minimizing operational overhead and avoiding instance management. - C is Incorrect: Running a container on Amazon ECS requires the team to manage the container orchestration, cluster capacity, and load balancing, introducing high operational overhead. - D is Incorrect: While SageMaker JumpStart deploys pre-trained models quickly, dynamically applying fine-tuned weights to a base model endpoint is not a native feature of JumpStart and does not provide a managed API without endpoint management. > Exam Tip: Bedrock Custom Model Import is the go-to solution for serving externally customized open-weight models via a fully managed API without instance management. |
| ### Question 2 - Cost-Effective Capacity for Steady High-Traffic FMs An enterprise is launching a customer-facing chatbot using an Anthropic Claude foundation model on Amazon Bedrock. The application is expected to receive steady, high-volume traffic 24/7. The engineering team must ensure predictable performance while optimizing costs for this continuous load. Which capacity strategy is the MOST cost-effective and appropriate for this workload? (Select ONE) - A) Use Amazon Bedrock On-Demand capacity and configure a SageMaker Endpoint for fallback. - B) Purchase Amazon Bedrock Provisioned Throughput to reserve a set level of input/output capacity. - C) Deploy the model on an Amazon EC2 instance with attached GPU accelerators. - D) Use Amazon Bedrock On-Demand capacity with an AWS Lambda concurrency limiter. |
Correct Answer: B Correct Answer Analysis: - B is Correct: Amazon Bedrock Provisioned Throughput allows you to reserve model units (MUs) for a set level of input and output capacity. For steady, high-volume 24/7 traffic, Provisioned Throughput provides predictable performance and is more cost-effective than paying On-Demand rates for continuous high usage. Distractor Analysis: - A is Incorrect: On-Demand capacity is better suited for intermittent or unpredictable traffic. Using a SageMaker Endpoint for fallback adds unnecessary operational complexity and does not address the cost optimization for steady traffic on Bedrock. - C is Incorrect: Deploying on EC2 requires managing the underlying infrastructure, drivers, and model serving stack, which contradicts the managed service benefit of Bedrock and introduces high overhead. - D is Incorrect: On-Demand capacity for 24/7 high-volume traffic will be more expensive than Provisioned Throughput. A Lambda concurrency limiter does not solve the cost or performance predictability requirements. > Exam Tip: Use Bedrock Provisioned Throughput for steady, predictable, high-volume traffic to guarantee performance and reduce costs compared to On-Demand. |
| ### Question 3 - Rapid Deployment of Pre-trained Open-Weight FMs A data science team needs to quickly deploy a pre-trained open-weight foundation model (e.g., Mixtral 8x7B) for a new internal proof-of-concept (POC). They require a pre-configured serving environment to avoid building custom containers from scratch, but they need the flexibility to choose specific EC2 instance types for the endpoint. Which AWS service provides the FASTEST way to deploy this model with a pre-configured environment? (Select ONE) - A) Amazon Bedrock On-Demand - B) Amazon Bedrock Custom Model Import - C) Amazon SageMaker JumpStart - D) AWS Lambda with container image packaging |
Correct Answer: C Correct Answer Analysis: - C is Correct: Amazon SageMaker JumpStart provides a catalog of pre-trained foundation models that can be deployed directly to a pre-configured SageMaker endpoint with a few clicks or lines of code. It handles the container setup while allowing the user to select the specific EC2 instance type for hosting. Distractor Analysis: - A is Incorrect: While Amazon Bedrock On-Demand is fast, it does not allow you to choose specific EC2 instance types because the infrastructure is fully abstracted and managed by AWS. - B is Incorrect: Custom Model Import is for importing customized model weights into Bedrock, not for deploying pre-trained models from a catalog into a configurable endpoint. - D is Incorrect: AWS Lambda has strict memory (10 GB) and timeout (15 minutes) limits and is generally not suitable for hosting large foundation models requiring GPU acceleration. > Exam Tip: Use SageMaker JumpStart when you need a pre-trained model deployed quickly but still want control over the underlying EC2 instance types and endpoint configuration. |
| ### Question 4 - Supported Architectures for Bedrock Custom Model Import An AI engineering team has trained a custom foundation model from scratch using a novel, proprietary neural network architecture. They want to use Amazon Bedrock Custom Model Import to serve this model without managing infrastructure, ensuring the lowest operational overhead. What must the team verify before choosing this deployment path? (Select ONE) - A) The model artifacts are serialized using TorchScript and stored in an EFS volume. - B) The model architecture is currently supported by Amazon Bedrock Custom Model Import (e.g., Llama, Mistral). - C) The custom model is packaged as a Docker container compatible with Amazon EKS. - D) The model has been quantized to 4-bit precision to meet Bedrock's memory limits. |
Correct Answer: B Correct Answer Analysis: - B is Correct: Amazon Bedrock Custom Model Import only accepts specific open-weight model architectures (such as Llama and Mistral) that have been customized or fine-tuned. A completely novel, proprietary architecture is not supported by the Custom Model Import feature. Distractor Analysis: - A is Incorrect: The storage format and medium (TorchScript/EFS) are not the primary gating factors; the underlying model architecture itself must be on Bedrock's supported list. - C is Incorrect: Custom Model Import does not use Docker containers; it takes raw model weights from Amazon S3 and hosts them on Bedrock's managed infrastructure. - D is Incorrect: While quantization can help with performance, it is not the prerequisite for determining if a model can be imported into Bedrock. > Exam Tip: Custom Model Import in Bedrock is architecture-gated. You can only import models based on supported open-weight architectures (like Llama or Mistral), not entirely custom-built network architectures. |
| ### Question 5 - Deep Infrastructure Control for FM Hosting A financial services company is deploying a highly optimized foundation model that requires custom CUDA libraries, specific kernel patches, and direct access to the network stack to integrate with legacy on-premises systems. Managed hosting services do not expose the required level of OS and driver control. The team has existing Kubernetes expertise. Which deployment option provides the necessary control over the infrastructure? (Select ONE) - A) Amazon SageMaker Real-Time Endpoints - B) Amazon Bedrock Provisioned Throughput - C) Amazon Elastic Kubernetes Service (Amazon EKS) - D) Amazon SageMaker Serverless Inference |
Correct Answer: C Correct Answer Analysis: - C is Correct: Amazon EKS provides managed Kubernetes, allowing teams to run their own custom containers with full root access to the underlying worker nodes. This allows them to install custom CUDA drivers, patch kernels, and configure advanced networking, which managed ML services abstract away. Distractor Analysis: - A is Incorrect: SageMaker Real-Time Endpoints abstract the underlying EC2 instances; you cannot install custom CUDA libraries or patch the OS on the endpoint instances. - B is Incorrect: Bedrock is a fully managed API service with zero infrastructure access. - D is Incorrect: Serverless Inference completely abstracts the infrastructure and does not even allow VPC configuration or custom instance selection. > Exam Tip: When a workload requires deep OS-level control (custom drivers, kernel patches) and the team has container orchestration expertise, Amazon EKS or ECS is the correct choice over managed ML services like SageMaker or Bedrock. |
| ### Question 6 - Serving Models Customized Inside Bedrock An enterprise has used Amazon Bedrock to fine-tune a base foundation model (Anthropic Claude) on their proprietary customer support data. The fine-tuned model is now ready for production. The engineering team wants to serve this model with a predictable level of throughput to meet SLA requirements. Which invocation method is REQUIRED to serve this customized model in production? (Select ONE) - A) Invoke the customized model using standard Bedrock On-Demand capacity. - B) Export the customized model weights and import them back via Custom Model Import. - C) Purchase Amazon Bedrock Provisioned Throughput to reserve capacity for the customized model. - D) Deploy the customized model to a SageMaker Real-Time Endpoint. |
Correct Answer: C Correct Answer Analysis: - C is Correct: According to AWS best practices, a model customized directly inside Amazon Bedrock (via fine-tuning) must be served through reserved capacity using Amazon Bedrock Provisioned Throughput. Distractor Analysis: - A is Incorrect: On-Demand capacity cannot be used to serve a model that has been customized directly within Bedrock; it requires Provisioned Throughput. - B is Incorrect: You do not need to export and re-import a model that was customized natively within Bedrock; doing so is unsupported and unnecessary. - D is Incorrect: Deploying to SageMaker is not the intended path for models fine-tuned within Bedrock, as Bedrock handles the hosting natively via Provisioned Throughput. > Exam Tip: Models fine-tuned natively inside Amazon Bedrock must be served using Bedrock Provisioned Throughput, whereas models customized externally and imported via Custom Model Import can be invoked On-Demand. |
| ### Question 7 - Managed Agent Orchestration A generative AI team is building an enterprise application where a foundation model needs to autonomously orchestrate multi-step workflows, query internal databases via APIs, and maintain session memory. The team wants to minimize the operational overhead of building and managing the orchestration logic. Which AWS service is PURPOSE-BUILT for this requirement? (Select ONE) - A) Amazon SageMaker Asynchronous Inference - B) Amazon Bedrock Agents (AgentCore) - C) AWS Step Functions integrated with Bedrock InvokeModel- D) Amazon OpenSearch Serverless with Vector Search |
Correct Answer: B Correct Answer Analysis: - B is Correct: Amazon Bedrock Agents (and the underlying AgentCore services) are purpose-built to orchestrate multi-step tasks, integrate with external APIs/functions as tools, manage memory, and handle identity/observability without requiring the team to build custom orchestration logic. Distractor Analysis: - A is Incorrect: Asynchronous Inference is an endpoint deployment strategy for long-running or large payload requests, not an agent orchestration framework. - C is Incorrect: While Step Functions can orchestrate Bedrock calls, building and maintaining the state machines and logic requires significant operational overhead compared to using managed Bedrock Agents. - D is Incorrect: OpenSearch Serverless handles vector storage and retrieval (RAG), not autonomous agent task execution or API tool integration. > Exam Tip: For autonomous FM orchestration with tools, memory, and API integration with least operational overhead, choose Amazon Bedrock Agents over custom-built Step Functions workflows. |
| ### Question 8 - Cost-Effective Batch Processing with FMs A retail company needs to generate product descriptions for 1 million items from a pre-assembled dataset stored in an Amazon S3 bucket. The process runs once a week and takes several hours to complete. The team wants to deploy a foundation model for this task in the MOST cost-effective manner, ensuring they only pay for compute while the job is actively running. Which deployment option meets these requirements? (Select ONE) - A) Amazon SageMaker Real-Time Endpoint with Auto Scaling - B) Amazon SageMaker Batch Transform - C) Amazon Bedrock Provisioned Throughput - D) Amazon SageMaker Serverless Inference |
Correct Answer: B Correct Answer Analysis: - B is Correct: SageMaker Batch Transform is designed exactly for this scenario: processing large, pre-assembled datasets in bulk. It provisions the compute, processes the input from S3, writes the predictions back to S3, and terminates the instances automatically, ensuring you only pay for the duration of the job. Distractor Analysis: - A is Incorrect: A Real-Time Endpoint would remain running and incur costs during the 6 days of the week it is not processing data. - C is Incorrect: Bedrock Provisioned Throughput reserves capacity over a period of time (e.g., hours or months) and would be massively underutilized for a weekly batch job. - D is Incorrect: Serverless Inference is designed for intermittent, on-demand single requests, not for bulk processing of millions of items in a single continuous job. > Exam Tip: For pre-assembled, scheduled, bulk dataset processing where the goal is to pay only for active compute time, SageMaker Batch Transform is the correct inference strategy. |
| ### Question 9 - Hosting Multiple Adapters on a Single Foundation Model An AI team has fine-tuned a base foundation model using Low-Rank Adaptation (LoRA) for 50 different enterprise use cases. They want to host these 50 adapters on a single managed endpoint to minimize idle infrastructure costs while allowing the calling application to specify which adapter to use at inference time. The models must share the same underlying serving container. Which deployment strategy is the MOST appropriate? (Select ONE) - A) Deploy 50 separate SageMaker Real-Time Endpoints, one for each adapter. - B) Use a SageMaker Multi-Model Endpoint (MME) to host the adapters on a shared compute fleet. - C) Deploy the base model on Bedrock and import the 50 adapters using Custom Model Import. - D) Deploy 50 individual containers on an Amazon ECS cluster. |
Correct Answer: B Correct Answer Analysis: - B is Correct: A SageMaker Multi-Model Endpoint (MME) allows you to host multiple models (or LoRA adapters) on a single shared compute fleet behind one endpoint. The endpoint dynamically loads the required adapter into memory based on the invocation request, drastically reducing infrastructure costs compared to dedicated endpoints. Distractor Analysis: - A is Incorrect: Deploying 50 separate endpoints would result in massive over-provisioning and high idle infrastructure costs. - C is Incorrect: Bedrock Custom Model Import is for importing full model weights of supported architectures, not for dynamically switching between 50 custom LoRA adapters on a single base model. - D is Incorrect: Managing 50 containers on ECS introduces significant operational overhead compared to the managed scaling and dynamic loading provided by SageMaker MME. > Exam Tip: SageMaker Multi-Model Endpoints (MME) are ideal for hosting many models or adapters that share the same framework and base container, reducing costs by sharing the compute fleet. |
| ### Question 10 - Zero-Infrastructure FM Inference for Bursty Traffic A startup is building an internal coding assistant that uses a foundation model to generate code snippets. The application experiences unpredictable, highly bursty traffic with long idle periods overnight. The team has no DevOps resources and strictly requires a deployment with zero infrastructure management and no idle compute costs, while tolerating cold start latency. Which AWS service is the BEST fit for this workload? (Select ONE) - A) Amazon SageMaker Real-Time Endpoint with Auto Scaling - B) Amazon Bedrock On-Demand - C) Amazon SageMaker Batch Transform - D) Amazon EC2 with G5 instances |
Correct Answer: B Correct Answer Analysis: - B is Correct: Amazon Bedrock completely abstracts the infrastructure. Using Bedrock On-Demand allows the startup to invoke the foundation model via an API without managing endpoints, scaling, or instances. They only pay per token/invocation, resulting in zero idle compute costs and perfectly handling bursty traffic. Distractor Analysis: - A is Incorrect: SageMaker Real-Time Endpoints require instance selection and management of Auto Scaling policies. Even with scaling to zero, it introduces operational overhead (inference components) that contradicts the zero-management constraint. - C is Incorrect: Batch Transform is for pre-assembled datasets, not for handling dynamic, bursty interactive requests from a coding assistant. - D is Incorrect: EC2 instances require full OS management, driver installation, and scaling logic, which violates the zero infrastructure management constraint. > Exam Tip: For bursty, unpredictable traffic with zero infrastructure management, Amazon Bedrock (On-Demand) is the ideal choice, as it abstracts all endpoint and instance management and eliminates idle compute costs. |