Topic 06 Practice Test 2 covers FM Deployment, Provisioned Throughput, Containers, GPUs, Cascading, and Hybrid Endpoints in Domain 2 (Task 2.2). For broader exam preparation, review the AIP-C01 Exam Dumps. Every option includes focused technical reasoning tied to the tested decision.
Question 1
A validated LLM must run at its current precision, but its parameter set is larger than any single available GPU memory footprint. Which approach best satisfies the requirement? Choose ONE.
- Model sharding across accelerators
- Load-test maximum context behavior
- Lambda as thin on-demand invocation client
- Latency-first model selection
Correct Answer: A
Correct Answer
Answer A is correct because Model sharding across accelerators: partitions large model state across accelerators so no single device must store every parameter. T06P2Q01 needs place portions of one model across multiple GPUs; it directly fits because it will use a supported multi-GPU model-sharding strategy.
Incorrect Answers
Answer B is incorrect because Load-test maximum context behavior: tests long prompts and outputs because memory use and token latency can be far worse than short synthetic requests. Best used when production traffic includes large contexts that materially change capacity needs; T06P2Q01 instead needs place portions of one model across multiple GPUs, so it misses the deciding constraint.
Answer C is incorrect because Lambda as thin on-demand invocation client: runs short-lived application logic that can call a managed FM API without hosting the model inside the function. Best used when bursty event-driven requests need lightweight orchestration around a managed FM; T06P2Q01 instead needs place portions of one model across multiple GPUs, so it misses the deciding constraint.
Answer D is incorrect because Latency-first model selection: chooses a model whose response-time characteristics satisfy the interactive experience even if a slower model has marginally higher benchmark quality. Best used when user experience has a hard response-time target; T06P2Q01 instead needs place portions of one model across multiple GPUs, so it misses the deciding constraint.
Question 2
A SageMaker LLM endpoint is costly and slow, but the team does not know whether GPUs are saturated or mostly idle waiting on preprocessing. What should the developer implement? Choose ONE.
- Quality-first model selection
- GPU utilization monitoring
- SageMaker network-isolated endpoint container
- Separate interactive and batch deployment paths
Correct Answer: B
Correct Answer
Answer B is correct because GPU utilization monitoring: tracks accelerator use to reveal whether expensive GPU capacity is saturated, idle, or blocked by another bottleneck. T06P2Q02 needs identify whether accelerator capacity is actually the bottleneck; it directly fits because it will collect and review endpoint GPU utilization and serving metrics before resizing.
Incorrect Answers
Answer A is incorrect because Quality-first model selection: chooses the higher-quality model when task errors are materially more costly than added latency or inference expense. Best used when the application can tolerate slower responses but not weaker reasoning on complex cases; T06P2Q02 instead needs identify whether accelerator capacity is actually the bottleneck, so it misses the deciding constraint.
Answer C is incorrect because SageMaker network-isolated endpoint container: hosts inference code with controlled VPC placement and can restrict the model container’s network access. Best used when custom inference must run under tighter network controls than a generic public-service integration; T06P2Q02 instead needs identify whether accelerator capacity is actually the bottleneck, so it misses the deciding constraint.
Answer D is incorrect because Separate interactive and batch deployment paths: keeps latency-sensitive online traffic from competing directly with throughput-oriented offline generation. Best used when online and offline workloads have different latency and scheduling objectives; T06P2Q02 instead needs identify whether accelerator capacity is actually the bottleneck, so it misses the deciding constraint.
Question 3
Latency rises sharply with concurrent users even though each model replica has stable per-request performance and enough memory. Which design most directly addresses the constraint? Choose ONE.
- Cost-aware routing threshold
- SageMaker endpoint startup timeout tuning
- Explicit model version pinning
- Separate model copies for concurrency
Correct Answer: D
Correct Answer
Answer D is correct because Separate model copies for concurrency: runs multiple serving replicas so independent requests can be processed concurrently when one model copy becomes a bottleneck. T06P2Q03 needs increase parallel serving capacity without changing model architecture; it directly fits because it will scale the number of hosted model replicas or endpoint instances.
Incorrect Answers
Answer A is incorrect because Cost-aware routing threshold: uses measured task value and model cost to decide when an expensive model is justified. Best used when premium inference should be reserved for requests where added capability changes the outcome; T06P2Q03 instead needs increase parallel serving capacity without changing model architecture, so it misses the deciding constraint.
Answer B is incorrect because SageMaker endpoint startup timeout tuning: extends model-download and container health-check windows so large model loading can complete before the platform declares failure. Best used when large model artifacts legitimately need more time to download and initialize; T06P2Q03 instead needs increase parallel serving capacity without changing model architecture, so it misses the deciding constraint.
Answer C is incorrect because Explicit model version pinning: keeps production behavior tied to a validated model version until a new version passes evaluation and rollout controls. Best used when uncontrolled model changes could alter output quality or compatibility; T06P2Q03 instead needs increase parallel serving capacity without changing model architecture, so it misses the deciding constraint.
Question 4
A customer-facing assistant has a strict first-response latency SLO and cannot tolerate multi-minute model initialization after idle periods. What is the most appropriate implementation? Choose ONE.
- AWS Large Model Inference container
- Warm capacity for strict latency SLO
- Fallback model for availability
- Bedrock Provisioned Throughput
Correct Answer: B
Correct Answer
Answer B is correct because Warm capacity for strict latency SLO: keeps serving resources initialized so production requests do not pay model-loading cold-start time. T06P2Q04 needs avoid large-model cold starts on user traffic; it directly fits because it will maintain warm hosted inference capacity for the interactive model.
Incorrect Answers
Answer A is incorrect because AWS Large Model Inference container: provides specialized serving components and libraries for deploying and optimizing LLM inference on SageMaker. Best used when large language model serving needs purpose-built loading and runtime optimizations; T06P2Q04 instead needs avoid large-model cold starts on user traffic, so it misses the deciding constraint.
Answer C is incorrect because Fallback model for availability: switches to an approved alternate model when the preferred model is unavailable or throttled, preserving a degraded but functional service. Best used when continuity matters more than using one exact model for every request; T06P2Q04 instead needs avoid large-model cold starts on user traffic, so it misses the deciding constraint.
Answer D is incorrect because Bedrock Provisioned Throughput: reserves model invocation capacity at a defined throughput level and fixed hourly cost for supported Bedrock models. Best used when steady, business-critical traffic needs predictable capacity; T06P2Q04 instead needs avoid large-model cold starts on user traffic, so it misses the deciding constraint.
Question 5
A pipeline only classifies support tickets into six stable categories, and a smaller tested model matches the large model’s accuracy. Which solution best fits this production requirement? Choose ONE.
- Smaller specialist model
- Tensor parallelism
- Benchmark before model downsizing
- Bedrock on-demand inference
Correct Answer: A
Correct Answer
Answer A is correct because Smaller specialist model: uses a lower-cost or lower-latency model that is sufficient for a narrow, well-bounded task. T06P2Q05 needs reduce compute and latency without sacrificing required task quality; it directly fits because it will route this bounded classification task to the validated smaller model.
Incorrect Answers
Answer B is incorrect because Tensor parallelism: splits model tensor computations across multiple accelerators so model execution is distributed beyond one GPU. Best used when a model is too large or compute-heavy for efficient execution on a single accelerator; T06P2Q05 instead needs reduce compute and latency without sacrificing required task quality, so it misses the deciding constraint.
Answer C is incorrect because Benchmark before model downsizing: compares candidate models against representative task data and acceptance criteria before replacing a larger model. Best used when cost reduction must not silently lower task quality; T06P2Q05 instead needs reduce compute and latency without sacrificing required task quality, so it misses the deciding constraint.
Answer D is incorrect because Bedrock on-demand inference: uses shared Bedrock capacity without a reserved throughput commitment and is billed according to actual model use. Best used when traffic is intermittent or uncertain and dedicated capacity would sit idle; T06P2Q05 instead needs reduce compute and latency without sacrificing required task quality, so it misses the deciding constraint.
Question 6
A support assistant can answer common FAQs reliably with a small model, while ambiguous cases need a premium model. Which approach best satisfies the requirement? Choose ONE.
- SageMaker real-time endpoint
- Confidence-based model cascade
- Quantized model serving
- Hybrid Bedrock and SageMaker deployment
Correct Answer: B
Correct Answer
Answer B is correct because Confidence-based model cascade: lets a cheaper model handle routine cases and escalates uncertain or difficult cases to a stronger model. T06P2Q06 needs spend premium-model capacity only on cases that need it; it directly fits because it will use a confidence or policy threshold to cascade difficult requests to the larger model.
Incorrect Answers
Answer A is incorrect because SageMaker real-time endpoint: hosts a model behind a managed low-latency endpoint with configurable instances, containers, networking, and autoscaling. Best used when a team needs control of model artifacts or custom inference code; T06P2Q06 instead needs spend premium-model capacity only on cases that need it, so it misses the deciding constraint.
Answer C is incorrect because Quantized model serving: reduces numerical precision of model weights or activations to lower memory use and often improve inference efficiency. Best used when the selected model nearly exceeds accelerator memory and an acceptable lower-precision variant is validated; T06P2Q06 instead needs spend premium-model capacity only on cases that need it, so it misses the deciding constraint.
Answer D is incorrect because Hybrid Bedrock and SageMaker deployment: combines managed Bedrock model APIs with SageMaker-hosted custom models when one hosting approach does not fit every model. Best used when an application needs both managed foundation models and a proprietary model with custom serving code; T06P2Q06 instead needs spend premium-model capacity only on cases that need it, so it misses the deciding constraint.
Question 7
An application handles translation, code generation, and short entity extraction, and one expensive general model is currently used for every request. What should the developer implement? Choose ONE.
- Provisioned model ARN invocation
- Dedicated endpoint for regulated custom model
- Continuous batching
- Task-based model routing
Correct Answer: D
Correct Answer
Answer D is correct because Task-based model routing: selects different models according to request type so each workload uses an appropriate capability and cost profile. T06P2Q07 needs match each task class to an appropriately capable model; it directly fits because it will route requests by task type to specialized or right-sized models.
Incorrect Answers
Answer A is incorrect because Provisioned model ARN invocation: identifies a purchased Bedrock Provisioned Throughput resource that can be supplied as the model identifier for inference. Best used when a provisioned resource is already InService and applications must use that reserved capacity; T06P2Q07 instead needs match each task class to an appropriately capable model, so it misses the deciding constraint.
Answer B is incorrect because Dedicated endpoint for regulated custom model: keeps a proprietary model on controlled SageMaker hosting where the team selects compute, networking, and deployment configuration. Best used when deployment policy requires direct control over the custom inference environment; T06P2Q07 instead needs match each task class to an appropriately capable model, so it misses the deciding constraint.
Answer C is incorrect because Continuous batching: dynamically groups token-generation work from concurrent requests to improve accelerator utilization during LLM decoding. Best used when many concurrent generation requests leave GPU execution inefficient when processed independently; T06P2Q07 instead needs match each task class to an appropriately capable model, so it misses the deciding constraint.
Question 8
A voice assistant must answer within a tight conversational delay, and two models both meet correctness requirements but one is consistently faster. Which design most directly addresses the constraint? Choose ONE.
- Provisioned Throughput readiness check
- Model preload before readiness
- Latency-first model selection
- Inference profile versus Provisioned Throughput distinction
Correct Answer: C
Correct Answer
Answer C is correct because Latency-first model selection: chooses a model whose response-time characteristics satisfy the interactive experience even if a slower model has marginally higher benchmark quality. T06P2Q08 needs prioritize the model that meets the required latency SLO; it directly fits because it will select the lower-latency model that still clears the quality threshold.
Incorrect Answers
Answer A is incorrect because Provisioned Throughput readiness check: uses the provisioning status to distinguish a usable reserved-capacity resource from one still being created or failed. Best used when deployment automation must wait until reserved capacity is available before cutting traffic; T06P2Q08 instead needs prioritize the model that meets the required latency SLO, so it misses the deciding constraint.
Answer B is incorrect because Model preload before readiness: loads required model weights and runtime state before the endpoint advertises itself as ready to receive production traffic. Best used when large model initialization would otherwise cause cold requests to fail or time out; T06P2Q08 instead needs prioritize the model that meets the required latency SLO, so it misses the deciding constraint.
Answer D is incorrect because Inference profile versus Provisioned Throughput distinction: separates cross-Region on-demand routing from dedicated Bedrock provisioned capacity because the two mechanisms solve different capacity problems. Best used when architects must choose between regional routing breadth and reserved model capacity; T06P2Q08 instead needs prioritize the model that meets the required latency SLO, so it misses the deciding constraint.
Question 9
A legal research workflow runs asynchronously and human reviewers value citation and reasoning quality far more than response speed. What is the most appropriate implementation? Choose ONE.
- Quality-first model selection
- Geographic cross-Region inference profile
- GPU memory headroom sizing
- Provisioned Throughput commitment sizing
Correct Answer: A
Correct Answer
Answer A is correct because Quality-first model selection: chooses the higher-quality model when task errors are materially more costly than added latency or inference expense. T06P2Q09 needs prioritize validated output quality over marginal serving cost or latency; it directly fits because it will select the model that best meets the required quality benchmark.
Incorrect Answers
Answer B is incorrect because Geographic cross-Region inference profile: allows Bedrock to route on-demand requests among supported Regions inside a defined geography. Best used when a workload can use multiple Regions but must keep processing within a permitted geography; T06P2Q09 instead needs prioritize validated output quality over marginal serving cost or latency, so it misses the deciding constraint.
Answer C is incorrect because GPU memory headroom sizing: selects accelerator memory with room for model weights plus runtime structures such as KV cache and framework overhead. Best used when token generation needs memory beyond the static size of the model artifact; T06P2Q09 instead needs prioritize validated output quality over marginal serving cost or latency, so it misses the deciding constraint.
Answer D is incorrect because Provisioned Throughput commitment sizing: aligns reserved capacity and commitment term with measured sustained demand so fixed capacity is not dramatically over- or under-sized. Best used when traffic is stable enough to forecast and the team is considering a paid commitment; T06P2Q09 instead needs prioritize validated output quality over marginal serving cost or latency, so it misses the deciding constraint.
Question 10
A commerce assistant sends every greeting and order-status lookup to the most expensive model even though simpler models handle them accurately. Which solution best fits this production requirement? Choose ONE.
- Cost-aware routing threshold
- Token-throughput capacity planning
- Scale on serving bottleneck metric
- Global cross-Region inference profile
Correct Answer: A
Correct Answer
Answer A is correct because Cost-aware routing threshold: uses measured task value and model cost to decide when an expensive model is justified. T06P2Q10 needs reduce spend by reserving premium inference for higher-value complexity; it directly fits because it will apply a routing policy that keeps routine requests on a cheaper validated model.
Incorrect Answers
Answer B is incorrect because Token-throughput capacity planning: sizes serving capacity using input and output token rates, concurrency, and latency rather than request count alone. Best used when requests vary greatly in context and generated length; T06P2Q10 instead needs reduce spend by reserving premium inference for higher-value complexity, so it misses the deciding constraint.
Answer C is incorrect because Scale on serving bottleneck metric: uses a metric tied to inference saturation or concurrency rather than an unrelated host signal when expanding serving capacity. Best used when autoscaling should react to the limiting resource that causes request queuing; T06P2Q10 instead needs reduce spend by reserving premium inference for higher-value complexity, so it misses the deciding constraint.
Answer D is incorrect because Global cross-Region inference profile: lets Bedrock route supported on-demand requests across commercial Regions for broader capacity access. Best used when data residency does not restrict processing and the workload values broad regional capacity and routing; T06P2Q10 instead needs reduce spend by reserving premium inference for higher-value complexity, so it misses the deciding constraint.
Question 11
A public assistant must remain usable during temporary primary-model capacity errors, and an alternate model has passed minimum quality checks. Which approach best satisfies the requirement? Choose ONE.
- Load-test maximum context behavior
- Container dependency pinning
- Fallback model for availability
- Application inference profile
Correct Answer: C
Correct Answer
Answer C is correct because Fallback model for availability: switches to an approved alternate model when the preferred model is unavailable or throttled, preserving a degraded but functional service. T06P2Q11 needs provide graceful degradation when the preferred model cannot serve traffic; it directly fits because it will configure an approved fallback model and bounded failover policy.
Incorrect Answers
Answer A is incorrect because Load-test maximum context behavior: tests long prompts and outputs because memory use and token latency can be far worse than short synthetic requests. Best used when production traffic includes large contexts that materially change capacity needs; T06P2Q11 instead needs provide graceful degradation when the preferred model cannot serve traffic, so it misses the deciding constraint.
Answer B is incorrect because Container dependency pinning: locks compatible serving-framework, CUDA, driver-facing, and model-library versions inside the deployment artifact. Best used when runtime drift can break model loading or GPU execution after an image rebuild; T06P2Q11 instead needs provide graceful degradation when the preferred model cannot serve traffic, so it misses the deciding constraint.
Answer D is incorrect because Application inference profile: provides a Bedrock invocation resource for tracking model usage and cost by application, optionally backed by cross-Region routing. Best used when teams need application-level usage attribution rather than new reserved capacity; T06P2Q11 instead needs provide graceful degradation when the preferred model cannot serve traffic, so it misses the deciding constraint.
Question 12
A team wants to replace a premium model with a cheaper small model, but only generic public benchmark scores are available. What should the developer implement? Choose ONE.
- SageMaker endpoint autoscaling
- Separate interactive and batch deployment paths
- Benchmark before model downsizing
- Model sharding across accelerators
Correct Answer: C
Correct Answer
Answer C is correct because Benchmark before model downsizing: compares candidate models against representative task data and acceptance criteria before replacing a larger model. T06P2Q12 needs prove that the smaller model meets this application’s real quality requirement; it directly fits because it will evaluate both models on representative production-like examples before switching.
Incorrect Answers
Answer A is incorrect because SageMaker endpoint autoscaling: adjusts hosted inference capacity in response to configured scaling metrics and policies. Best used when a managed endpoint has variable sustained load and should add or remove serving instances automatically; T06P2Q12 instead needs prove that the smaller model meets this application’s real quality requirement, so it misses the deciding constraint.
Answer B is incorrect because Separate interactive and batch deployment paths: keeps latency-sensitive online traffic from competing directly with throughput-oriented offline generation. Best used when online and offline workloads have different latency and scheduling objectives; T06P2Q12 instead needs prove that the smaller model meets this application’s real quality requirement, so it misses the deciding constraint.
Answer D is incorrect because Model sharding across accelerators: partitions large model state across accelerators so no single device must store every parameter. Best used when model size exceeds one accelerator’s memory even after reasonable optimization; T06P2Q12 instead needs prove that the smaller model meets this application’s real quality requirement, so it misses the deciding constraint.
Question 13
A product uses a managed Bedrock FM for summarization but must also serve its own domain model with a custom tokenizer and container. Which design most directly addresses the constraint? Choose ONE.
- GPU utilization monitoring
- Explicit model version pinning
- Lambda as thin on-demand invocation client
- Hybrid Bedrock and SageMaker deployment
Correct Answer: D
Correct Answer
Answer D is correct because Hybrid Bedrock and SageMaker deployment: combines managed Bedrock model APIs with SageMaker-hosted custom models when one hosting approach does not fit every model. T06P2Q13 needs use the appropriate managed deployment surface for each model type; it directly fits because it will keep the managed FM on Bedrock and deploy the proprietary model on SageMaker.
Incorrect Answers
Answer A is incorrect because GPU utilization monitoring: tracks accelerator use to reveal whether expensive GPU capacity is saturated, idle, or blocked by another bottleneck. Best used when operators need evidence before changing expensive inference instance capacity; T06P2Q13 instead needs use the appropriate managed deployment surface for each model type, so it misses the deciding constraint.
Answer B is incorrect because Explicit model version pinning: keeps production behavior tied to a validated model version until a new version passes evaluation and rollout controls. Best used when uncontrolled model changes could alter output quality or compatibility; T06P2Q13 instead needs use the appropriate managed deployment surface for each model type, so it misses the deciding constraint.
Answer C is incorrect because Lambda as thin on-demand invocation client: runs short-lived application logic that can call a managed FM API without hosting the model inside the function. Best used when bursty event-driven requests need lightweight orchestration around a managed FM; T06P2Q13 instead needs use the appropriate managed deployment surface for each model type, so it misses the deciding constraint.
Question 14
A regulated team must choose the exact container, instance family, VPC subnets, and security groups for its proprietary LLM. What is the most appropriate implementation? Choose ONE.
- Dedicated endpoint for regulated custom model
- SageMaker network-isolated endpoint container
- Bedrock Provisioned Throughput
- Separate model copies for concurrency
Correct Answer: A
Correct Answer
Answer A is correct because Dedicated endpoint for regulated custom model: keeps a proprietary model on controlled SageMaker hosting where the team selects compute, networking, and deployment configuration. T06P2Q14 needs retain infrastructure controls unavailable from a simple shared model API; it directly fits because it will host the custom model on a configured SageMaker endpoint.
Incorrect Answers
Answer B is incorrect because SageMaker network-isolated endpoint container: hosts inference code with controlled VPC placement and can restrict the model container’s network access. Best used when custom inference must run under tighter network controls than a generic public-service integration; T06P2Q14 instead needs retain infrastructure controls unavailable from a simple shared model API, so it misses the deciding constraint.
Answer C is incorrect because Bedrock Provisioned Throughput: reserves model invocation capacity at a defined throughput level and fixed hourly cost for supported Bedrock models. Best used when steady, business-critical traffic needs predictable capacity; T06P2Q14 instead needs retain infrastructure controls unavailable from a simple shared model API, so it misses the deciding constraint.
Answer D is incorrect because Separate model copies for concurrency: runs multiple serving replicas so independent requests can be processed concurrently when one model copy becomes a bottleneck. Best used when the model fits comfortably but a single replica cannot meet concurrent request demand; T06P2Q14 instead needs retain infrastructure controls unavailable from a simple shared model API, so it misses the deciding constraint.
Question 15
A design proposes putting Provisioned Throughput behind a cross-Region inference profile, but the profile does not support provisioned capacity. Which solution best fits this production requirement? Choose ONE.
- SageMaker endpoint startup timeout tuning
- Bedrock on-demand inference
- Warm capacity for strict latency SLO
- Inference profile versus Provisioned Throughput distinction
Correct Answer: D
Correct Answer
Answer D is correct because Inference profile versus Provisioned Throughput distinction: separates cross-Region on-demand routing from dedicated Bedrock provisioned capacity because the two mechanisms solve different capacity problems. T06P2Q15 needs avoid combining two incompatible Bedrock invocation modes; it directly fits because it will choose either supported cross-Region on-demand inference or Provisioned Throughput based on the requirement.
Incorrect Answers
Answer A is incorrect because SageMaker endpoint startup timeout tuning: extends model-download and container health-check windows so large model loading can complete before the platform declares failure. Best used when large model artifacts legitimately need more time to download and initialize; T06P2Q15 instead needs avoid combining two incompatible Bedrock invocation modes, so it misses the deciding constraint.
Answer B is incorrect because Bedrock on-demand inference: uses shared Bedrock capacity without a reserved throughput commitment and is billed according to actual model use. Best used when traffic is intermittent or uncertain and dedicated capacity would sit idle; T06P2Q15 instead needs avoid combining two incompatible Bedrock invocation modes, so it misses the deciding constraint.
Answer C is incorrect because Warm capacity for strict latency SLO: keeps serving resources initialized so production requests do not pay model-loading cold-start time. Best used when interactive requests have a tight latency objective incompatible with loading a large model on demand; T06P2Q15 instead needs avoid combining two incompatible Bedrock invocation modes, so it misses the deciding constraint.
Question 16
A production assistant has three months of token and concurrency measurements and finance wants predictable capacity without buying excessive headroom. Which approach best satisfies the requirement? Choose ONE.
- AWS Large Model Inference container
- Smaller specialist model
- Provisioned Throughput commitment sizing
- SageMaker real-time endpoint
Correct Answer: C
Correct Answer
Answer C is correct because Provisioned Throughput commitment sizing: aligns reserved capacity and commitment term with measured sustained demand so fixed capacity is not dramatically over- or under-sized. T06P2Q16 needs base the reserved-capacity purchase on measured sustained token demand and commitment needs; it directly fits because it will size the Provisioned Throughput from observed peak token load plus justified headroom.
Incorrect Answers
Answer A is incorrect because AWS Large Model Inference container: provides specialized serving components and libraries for deploying and optimizing LLM inference on SageMaker. Best used when large language model serving needs purpose-built loading and runtime optimizations; T06P2Q16 instead needs base the reserved-capacity purchase on measured sustained token demand and commitment needs, so it misses the deciding constraint.
Answer B is incorrect because Smaller specialist model: uses a lower-cost or lower-latency model that is sufficient for a narrow, well-bounded task. Best used when a simple task does not require the reasoning capability of the largest available FM; T06P2Q16 instead needs base the reserved-capacity purchase on measured sustained token demand and commitment needs, so it misses the deciding constraint.
Answer D is incorrect because SageMaker real-time endpoint: hosts a model behind a managed low-latency endpoint with configurable instances, containers, networking, and autoscaling. Best used when a team needs control of model artifacts or custom inference code; T06P2Q16 instead needs base the reserved-capacity purchase on measured sustained token demand and commitment needs, so it misses the deciding constraint.
Question 17
An LLM endpoint queues requests while average CPU remains low because GPU/token generation is the limiting stage. What should the developer implement? Choose ONE.
- Provisioned model ARN invocation
- Scale on serving bottleneck metric
- Tensor parallelism
- Confidence-based model cascade
Correct Answer: B
Correct Answer
Answer B is correct because Scale on serving bottleneck metric: uses a metric tied to inference saturation or concurrency rather than an unrelated host signal when expanding serving capacity. T06P2Q17 needs drive scaling from an inference-relevant utilization or backlog signal; it directly fits because it will configure scaling against a metric that reflects serving saturation rather than CPU alone.
Incorrect Answers
Answer A is incorrect because Provisioned model ARN invocation: identifies a purchased Bedrock Provisioned Throughput resource that can be supplied as the model identifier for inference. Best used when a provisioned resource is already InService and applications must use that reserved capacity; T06P2Q17 instead needs drive scaling from an inference-relevant utilization or backlog signal, so it misses the deciding constraint.
Answer C is incorrect because Tensor parallelism: splits model tensor computations across multiple accelerators so model execution is distributed beyond one GPU. Best used when a model is too large or compute-heavy for efficient execution on a single accelerator; T06P2Q17 instead needs drive scaling from an inference-relevant utilization or backlog signal, so it misses the deciding constraint.
Answer D is incorrect because Confidence-based model cascade: lets a cheaper model handle routine cases and escalates uncertain or difficult cases to a stronger model. Best used when most requests are simple but a minority need more capable reasoning; T06P2Q17 instead needs drive scaling from an inference-relevant utilization or backlog signal, so it misses the deciding constraint.
Question 18
An endpoint passes tests with 500-token prompts but fails during 30,000-token production conversations. Which design most directly addresses the constraint? Choose ONE.
- Provisioned Throughput readiness check
- Quantized model serving
- Task-based model routing
- Load-test maximum context behavior
Correct Answer: D
Correct Answer
Answer D is correct because Load-test maximum context behavior: tests long prompts and outputs because memory use and token latency can be far worse than short synthetic requests. T06P2Q18 needs validate deployment capacity using representative long-context requests; it directly fits because it will load-test the endpoint with realistic context lengths and generation limits.
Incorrect Answers
Answer A is incorrect because Provisioned Throughput readiness check: uses the provisioning status to distinguish a usable reserved-capacity resource from one still being created or failed. Best used when deployment automation must wait until reserved capacity is available before cutting traffic; T06P2Q18 instead needs validate deployment capacity using representative long-context requests, so it misses the deciding constraint.
Answer B is incorrect because Quantized model serving: reduces numerical precision of model weights or activations to lower memory use and often improve inference efficiency. Best used when the selected model nearly exceeds accelerator memory and an acceptable lower-precision variant is validated; T06P2Q18 instead needs validate deployment capacity using representative long-context requests, so it misses the deciding constraint.
Answer C is incorrect because Task-based model routing: selects different models according to request type so each workload uses an appropriate capability and cost profile. Best used when distinct task classes benefit from different model strengths; T06P2Q18 instead needs validate deployment capacity using representative long-context requests, so it misses the deciding constraint.
Question 19
A nightly document-generation job saturates the same serving pool used by an interactive assistant and causes daytime-like latency spikes overnight. What is the most appropriate implementation? Choose ONE.
- Geographic cross-Region inference profile
- Separate interactive and batch deployment paths
- Latency-first model selection
- Continuous batching
Correct Answer: B
Correct Answer
Answer B is correct because Separate interactive and batch deployment paths: keeps latency-sensitive online traffic from competing directly with throughput-oriented offline generation. T06P2Q19 needs protect interactive capacity from large offline jobs; it directly fits because it will use separate serving capacity or deployment paths for interactive and batch workloads.
Incorrect Answers
Answer A is incorrect because Geographic cross-Region inference profile: allows Bedrock to route on-demand requests among supported Regions inside a defined geography. Best used when a workload can use multiple Regions but must keep processing within a permitted geography; T06P2Q19 instead needs protect interactive capacity from large offline jobs, so it misses the deciding constraint.
Answer C is incorrect because Latency-first model selection: chooses a model whose response-time characteristics satisfy the interactive experience even if a slower model has marginally higher benchmark quality. Best used when user experience has a hard response-time target; T06P2Q19 instead needs protect interactive capacity from large offline jobs, so it misses the deciding constraint.
Answer D is incorrect because Continuous batching: dynamically groups token-generation work from concurrent requests to improve accelerator utilization during LLM decoding. Best used when many concurrent generation requests leave GPU execution inefficient when processed independently; T06P2Q19 instead needs protect interactive capacity from large offline jobs, so it misses the deciding constraint.
Question 20
A regulated application must reproduce prior behavior during audits and cannot accept silent model-version changes. Which solution best fits this production requirement? Choose ONE.
- Model preload before readiness
- Global cross-Region inference profile
- Explicit model version pinning
- Quality-first model selection
Correct Answer: C
Correct Answer
Answer C is correct because Explicit model version pinning: keeps production behavior tied to a validated model version until a new version passes evaluation and rollout controls. T06P2Q20 needs keep production inference on the specifically approved model revision; it directly fits because it will pin the validated model identifier/version and promote changes through controlled release.
Incorrect Answers
Answer A is incorrect because Model preload before readiness: loads required model weights and runtime state before the endpoint advertises itself as ready to receive production traffic. Best used when large model initialization would otherwise cause cold requests to fail or time out; T06P2Q20 instead needs keep production inference on the specifically approved model revision, so it misses the deciding constraint.
Answer B is incorrect because Global cross-Region inference profile: lets Bedrock route supported on-demand requests across commercial Regions for broader capacity access. Best used when data residency does not restrict processing and the workload values broad regional capacity and routing; T06P2Q20 instead needs keep production inference on the specifically approved model revision, so it misses the deciding constraint.
Answer D is incorrect because Quality-first model selection: chooses the higher-quality model when task errors are materially more costly than added latency or inference expense. Best used when the application can tolerate slower responses but not weaker reasoning on complex cases; T06P2Q20 instead needs keep production inference on the specifically approved model revision, so it misses the deciding constraint.
