Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 21

Question 200

A proposed Bedrock batch inference JSONL file expects every record to call a live inventory tool and then continue the model conversation with the returned tool result. What should the developer do?

Enable a larger S3 batch output bucket so toolUse/toolResult exchanges become available.
Use batch inference records as a direct substitute for an interactive multi-turn tool dialogue.
Put the inventory access token in the prompt and assume the batch worker executes the tool itself.
Use a supported interactive tool-use API and application/tool orchestrator, because Bedrock batch inference does not support that back-and-forth tool calling.

Explanation

Correct Answer

Client-side Converse tool use, for example, supports a model tool request, an application-executed tool, and a toolResult message. Batch inference cannot perform that loop.

Incorrect Answers

Output storage capacity does not add a feature that the batch record protocol does not support.

Batch records are independent asynchronous inference operations, not ongoing interactive exchanges that wait for client tool execution.

A token in a prompt is neither safe nor an instruction for AWS batch inference to invoke an external business API.

Question 201

A deployment requires a customized open-weight model on GPU. Which deployment or capacity decision is most appropriate?

Select an arbitrary Bedrock foundation-model ID as the custom GPU container target.
Choose the lowest-CPU SageMaker instance without measuring model weight memory needs.
Put model checkpoints in S3 and assume this alone creates a live HTTPS inference endpoint.
Evaluate compatible SageMaker real-time hosting and memory/accelerator requirements.

Explanation

Correct Answer

For a custom open-weight model, verify SageMaker serving container, accelerator memory, instance and model packaging compatibility before deployment.

Incorrect Answers

Catalog model IDs do not automatically deploy arbitrary model binaries.

GPU VRAM and architecture compatibility may dominate hosting requirements.

Artifact storage and an operational serving deployment are distinct steps.

Question 202

A candidate model requires more accelerator memory than allocated. Which operational correction addresses the deployment failure?

Increase API Gateway timeout to compensate for GPU out-of-memory failures.
Scale HTTP worker threads while keeping the same undersized model GPU allocation.
Disable model error monitoring once a quantized image loads once.
Assess model loading/quantization/sharding and hardware fit before rollout.

Explanation

Correct Answer

Estimate model weight/activation/KV-cache memory and check supported quantization, tensor sharding and accelerator limits before rollout.

Incorrect Answers

Longer request time does not make insufficient accelerator memory available.

More concurrency may worsen rather than repair model-loading memory pressure.

A single startup does not validate sustained memory, accuracy or serving stability.

Question 203

An endpoint wastes GPU memory keeping idle model replicas. What should be verified before choosing this model-serving path?

Keep every GPU replica warm indefinitely even when traffic is near zero.
Reevaluate minimum replicas and auto scaling versus cold-start acceptance.
Scale replicas to zero without evaluating initialization delay for interactive callers.
Reduce output tokens only and ignore idle serving instance costs.

Explanation

Correct Answer

Tune minimum replicas and scale-to-zero where supported against measured cold-start and latency objectives, not idle GPU count alone.

Incorrect Answers

Idle replicas waste resources if cold-start latency can be tolerated.

Aggressive zero-minimum scaling may violate response-time commitments.

Shorter generation does not eliminate baseline replica expense.

Question 204

A model endpoint misses p95 while GPU utilization is low. Which deployment or capacity decision is most appropriate?

Add expensive accelerator instances solely because p95 latency increased.
Profile token preprocessing/queue/network before adding accelerators blindly.
Measure only average model-compute time and omit request queue and tokenization.
Increase max output tokens to mask intermittent input-processing delays.

Explanation

Correct Answer

Decompose end-to-end latency into tokenization, network, queueing and generation; low GPU utilization suggests the bottleneck may be outside compute.

Incorrect Answers

Adding GPUs cannot repair client, network or preprocessing delays.

End-to-end p95 can be dominated by unmeasured stages.

Longer outputs add serving time rather than remove the underlying bottleneck.

Question 205

A model endpoint is saturated at high output tokens. Which operational correction addresses the deployment failure?

Plan token throughput capacity rather than counting HTTP requests alone.
Size endpoints by calls per second while assuming every request generates the same tokens.
Increase S3 upload concurrency to address token-generation throughput.
Use input token count only and ignore long completion generation.

Explanation

Correct Answer

Capacity depends on sustained input/output tokens, output lengths and concurrency, not the number of HTTP requests alone.

Incorrect Answers

Unequal output lengths produce different compute demand and saturation.

Object upload capacity does not control model decode tokens per second.

Output token decoding can dominate the saturated endpoint.

Question 206

A workload has bursty traffic with no predictable peak. What should be verified before choosing this model-serving path?

Benchmark on-demand/auto scaling rather than reserving a fixed oversized pool.
Commit a fixed large allocation based solely on the largest imagined peak.
Disable request admission controls and assume autoscaling reacts instantly to every burst.
Use a single quiet-hour benchmark to size the unpredictable workload.

Explanation

Correct Answer

For unpredictable bursts, benchmark supported on-demand and auto-scaling policies with real token traffic before reserving oversized steady capacity.

Incorrect Answers

Unverified peaks can create expensive idle capacity.

Provisioning and model warmup have limits; overload requires bounded queues/fallback.

A quiet sample cannot represent burst concurrency and tail latency.

Question 207

A serving container loads model weights on every request. Which deployment or capacity decision is most appropriate?

Keep model loaded across requests or adopt supported serving runtime for amortization.
Download and deserialize the entire model checkpoint on every single inference request.
Increase the number of HTTP frontend proxies without changing model lifecycle.
Cache only the generated text while leaving the container to reload weights per call.

Explanation

Correct Answer

Avoid repeated weight loading by keeping a compatible model resident in a warm serving process or using a serving stack designed for amortized initialization.

Incorrect Answers

Repeated loading dominates latency and can exhaust storage/network resources.

More proxies cannot remove model initialization in each backend request.

Output caching misses new requests and does not fix model process initialization cost.

Question 208

A deployment targets a Region without selected model availability. Which operational correction addresses the deployment failure?

Assume the model runs in every region because its provider appears in the Bedrock catalog.
Create a provisioning stack using a guessed model ARN and approve it after compilation.
Verify supported deployment type, Region and quotas before provisioning.
Change only the IAM policy without checking whether the selected model exists in that Region.

Explanation

Correct Answer

Check the exact model/Region/deployment option and quotas before provisioning, including supported inference profiles where applicable.

Incorrect Answers

Provider listing is not proof of model-version/Region support.

Template syntax does not establish runtime availability of an ARN.

Permissions do not create unavailable regional model capacity.

Question 209

A team wants to add a fallback FM to an app. What should be verified before choosing this model-serving path?

Test behavioral/schema compatibility and business result quality on alternate model.
Switch immediately when the alternate model responds to a single hello prompt.
Compare only catalog prices and ignore the fallback's supported API and tool behavior.
Assume model ID aliases guarantee identical JSON schemas and safety behavior.

Explanation

Correct Answer

Test the fallback model's output schema, tool support, language and accepted answer quality before switching production traffic.

Incorrect Answers

A health check does not prove functional equivalence for business cases.

A cheap model may lack mandatory modality or tool capabilities.

Provider/version changes can alter response formats and quality.

Question 210

A deployment uses an end-of-life model ID. Which deployment or capacity decision is most appropriate?

Continue invoking a retired model indefinitely and assume Bedrock silently maps it to another version.
Replace the ID in configuration without testing output contract and user-case correctness.
Plan supported model-version upgrade and decommission according to Bedrock lifecycle.
Archive the application's prompt code and treat that as a substitute for a supported model deployment.

Explanation

Correct Answer

An EOL model requires planned migration to a published supported version with regression testing, not a silent identifier substitution.

Incorrect Answers

Lifecycle transitions can end availability rather than transparently preserving API behavior.

A new version can change semantics or unsupported settings.

Source archiving does not keep an EOL inference resource operational.

Pass Your Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Exam Easy!

Question 200

Explanation

Correct Answer

Incorrect Answers

Question 201

Explanation

Correct Answer

Incorrect Answers

Question 202

Explanation

Correct Answer

Incorrect Answers

Question 203

Explanation

Correct Answer

Incorrect Answers

Question 204

Explanation

Correct Answer

Incorrect Answers

Question 205

Explanation

Correct Answer

Incorrect Answers

Question 206

Explanation

Correct Answer

Incorrect Answers

Question 207

Explanation

Correct Answer

Incorrect Answers

Question 208

Explanation

Correct Answer

Incorrect Answers

Question 209

Explanation

Correct Answer

Incorrect Answers

Question 210

Explanation

Correct Answer

Incorrect Answers

AWS Certified Generative AI Developer - Professional AIP-C01 Practice Tests