ExamPark
Course menuMLA-C02

Exams / MLA-C02 / Study guide / Deploying foundation models

Deploying foundation models: MLA-C02 study notes and practice questions

What to know about Deploying foundation models for the MLA-C02 exam, from the Deployment and Orchestration of ML and AI Workflows domain (24% of the exam guide), then 2 free questions to test it. Part of the MLA-C02 study guide.

Last verified against current AWS documentation ·How we write our questions

What to know for the exam

These questions ask how to serve a foundation model with the right throughput, cost and control.

Amazon Bedrock options

Option Use it when
On-demand Variable traffic; pay per token; subject to account quotas
Cross-Region inference (inference profiles) You need more throughput and resilience; Bedrock routes requests to other Regions in a geography
Provisioned throughput Steady, high traffic needing guaranteed capacity, under a commitment; also used to serve many customized models
Batch inference Large non-urgent jobs from S3, at a lower price than on-demand
Custom Model Import You trained or fine-tuned a supported open-weight model elsewhere and want to call it through Bedrock APIs with the same IAM, guardrails and tooling
Bedrock Marketplace Specialized models deployed to managed endpoints

SageMaker AI hosting

  • SageMaker JumpStart deploys open models (such as Llama or Mistral families) to endpoints you control, with fine-tuning options.
  • Large model inference (LMI) containers with vLLM or TensorRT-LLM handle batching and tensor parallelism.
  • Inference components pack several models onto GPU instances.
  • Choose SageMaker AI when you need custom containers, specific instances, VPC-only hosting or full control of the serving stack.

Performance tuning

  • Streaming responses improve perceived latency.
  • Quantization shrinks models to fit smaller GPUs.
  • Inferentia instances lower inference cost for compiled models.
  • Prompt caching cuts cost and latency for repeated context.

Common traps

  • Provisioned throughput for spiky, low traffic. On-demand fits.
  • Fine-tuning in Bedrock a model whose weights were created elsewhere. Use Custom Model Import.
  • Expecting batch inference or cross-Region inference to reserve capacity. Only provisioned throughput does.

Try the 2 practice questions

Practice Deploying foundation modelsTake the free practice exam

Question 1

Harder ·Link to this question (question 1)

A ski resort group fine-tuned an open-weight language model on its own server in a data center, using its guest-feedback data. It now wants the model available through Amazon Bedrock so that its applications can call it with the same APIs and access controls as its other Bedrock models.

The weights are in a private S3 bucket in the account, and the team does not want to host or patch inference servers. The model must not be retrained on AWS.

Which approach meets these requirements?

Choose one answer.

  1. A. Create a Bedrock fine-tuning job that points at the uploaded weights, so that Bedrock registers the result as a custom model it can serve.
  2. B. Use Bedrock Custom Model Import, uploading the model weights from the S3 bucket in a supported format, and invoke the imported model through Bedrock.
  3. C. Deploy the weights on an Amazon EC2 instance with a container and call that instance from the applications through a load balancer.
  4. D. Store the model in Amazon SageMaker JumpStart and invoke it through the JumpStart catalog, since JumpStart offers all open-weight models through Bedrock.
Show the answer and explanation

Answer: B

  • A. Wrong: Bedrock fine-tuning starts from a Bedrock base model and runs training on AWS, which the stem does not want. It does not import a model that was trained elsewhere.
  • B. Right: Correct. Custom Model Import brings a model whose weights were built outside AWS into Bedrock from S3, so applications use the same Bedrock APIs and access controls. The team does not run inference servers. Learn more: Use Custom Model Import to import a customized open-source model (opens in a new tab)
  • C. Wrong: This keeps the model outside Bedrock and makes the team operate the inference servers, which the stem rules out. The applications would not use Bedrock APIs.
  • D. Wrong: JumpStart is a catalog of models for SageMaker AI, and it does not put a custom model into Bedrock. The stem asks for Bedrock APIs and access controls.

Explanation

Bedrock Custom Model Import loads a model whose weights were created outside Bedrock from S3, so the model is invoked through the Bedrock APIs with the same access controls. Fine-tuning in Bedrock trains from a Bedrock base model, and hosting on EC2 or JumpStart does not put the model in Bedrock.

Study

Question 2

Medium ·Link to this question (question 2)

A regional bank runs a customer-service assistant on a Bedrock foundation model. Traffic is steady and high during business hours in every time zone the bank serves, and the bank needs predictable latency and throughput for its peak periods.

The bank wants a commitment that guarantees capacity for its model usage, and it accepts paying for that reserved capacity for a fixed term.

Which Bedrock option should the bank choose?

Choose one answer.

  1. A. On-demand inference, which charges per token and lets the bank send as many requests as the service can accept at the time.
  2. B. Cross-Region inference, which routes requests to other Regions when the home Region is busy, so that the bank never sees throttling.
  3. C. Provisioned throughput for the model with a commitment term, so that the bank has reserved model capacity for its steady peak load.
  4. D. Batch inference, which processes the requests in a file overnight so that daytime traffic does not reach the model.
Show the answer and explanation

Answer: C

  • A. Wrong: On-demand has no capacity commitment, so throughput can vary during peaks. The stem asks for a guaranteed capacity commitment.
  • B. Wrong: Cross-Region inference helps absorb bursts across Regions, but it does not reserve capacity for a commitment term. The stem asks for guaranteed throughput.
  • C. Right: Correct. Provisioned throughput reserves model capacity for a commitment term, which gives predictable throughput and latency for steady, high traffic. The bank pays for the reserved capacity. Learn more: Increase model invocation capacity with Provisioned Throughput (opens in a new tab)
  • D. Wrong: Batch inference is for offline work and does not serve the bank's interactive assistant. The option removes the requests the stem is about.

Explanation

Provisioned throughput reserves model capacity under a commitment, which suits steady, high traffic that needs predictable throughput. On-demand has no reservation, and cross-Region inference and batch inference do not reserve capacity for interactive peaks.

Study

Want more on Deploying foundation models? The 5 full-length practice sets have 10 more Deploying foundation models questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets

Study Deploying foundation models

More Deployment and Orchestration of ML and AI Workflows topics