Question 680A shopping application needs autoscaling tuned to preserve an interactive model inference SLO. Which property of the selected ML workflow accounts for this requirement? Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. A custom container controls dependencies and request handling but inherits security, compatibility and patching duties. Endpoint network reachability and cloud identity authorization are separate security boundaries. ExplanationCorrect Answer A:Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. This supports the specific implementation: Configure supported autoscaling and capacity limits for the chosen inference backend using measured request and utilization metrics. A suitable independent check is: Test load ramps and observe latency, replica counts and saturation before and after scaling configuration. Other Answers B:Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. The cause in the stated scenario is different: Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. C:A custom container controls dependencies and request handling but inherits security, compatibility and patching duties. The cause in the stated scenario is different: Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. D:Endpoint network reachability and cloud identity authorization are separate security boundaries. The cause in the stated scenario is different: Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes.
Question 681A shopping application needs autoscaling tuned to preserve an interactive model inference SLO. Which evidence would best demonstrate that this ML solution meets the requirement? Measure cold-start frequency, p95 first-token delay and monthly idle cost with and without warm capacity. Compare per-version request counts, quality, error rate and latency before raising candidate traffic share. Test load ramps and observe latency, replica counts and saturation before and after scaling configuration. Replay long-context workloads and measure accelerator memory, queueing and p95 response time. ExplanationCorrect Answer C:Test load ramps and observe latency, replica counts and saturation before and after scaling configuration. This tests the underlying failure: Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. Other Answers A:Warm replicas reduce first-request initialization latency but incur charges while receiving no traffic. That does not establish the required outcome; verify it by checking: Test load ramps and observe latency, replica counts and saturation before and after scaling configuration. B:Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. That does not establish the required outcome; verify it by checking: Test load ramps and observe latency, replica counts and saturation before and after scaling configuration. D:Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. That does not establish the required outcome; verify it by checking: Test load ramps and observe latency, replica counts and saturation before and after scaling configuration.
Question 682A shopping application needs autoscaling tuned to preserve an interactive model inference SLO. What preventive release checklist item best protects the intended outcome? Reserve quota headroom and revalidate autoscaling parameters during seasonal traffic planning. Predefine rollback thresholds and staged traffic percentages for model deployments. Review endpoint exposure and identity policies before enabling any new inference consumer. Set serving batch limits and monitor tail latency before automatically increasing batching thresholds. ExplanationCorrect Answer A:Reserve quota headroom and revalidate autoscaling parameters during seasonal traffic planning. This protects the selected approach because Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. The immediate solution is: Configure supported autoscaling and capacity limits for the chosen inference backend using measured request and utilization metrics. Other Answers B:Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. The continuing safeguard required here is: Reserve quota headroom and revalidate autoscaling parameters during seasonal traffic planning. C:Endpoint network reachability and cloud identity authorization are separate security boundaries. The continuing safeguard required here is: Reserve quota headroom and revalidate autoscaling parameters during seasonal traffic planning. D:Batching increases parallel accelerator use but adds queueing delay and may harm tail latency if configured poorly. The continuing safeguard required here is: Reserve quota headroom and revalidate autoscaling parameters during seasonal traffic planning.
Question 683A model endpoint has plenty of CPU capacity but repeatedly runs out of GPU memory for long prompts. What should the team change to address the specific technical requirement? Package and version inference preprocessing and postprocessing together with the approved model artifact. Profile per-request token generation, model size and decoding settings before scaling only the replica count. Choose serving accelerator memory and concurrency limits based on measured token or batch demand. Deploy a supported private endpoint or private networking controls with least-privilege service authentication. ExplanationCorrect Answer C:Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. To validate the correction: Replay long-context workloads and measure accelerator memory, queueing and p95 response time. Other Answers A:Training-serving feature skew can arise when model inputs are transformed inconsistently across environments. It does not resolve the specific failure; the targeted correction is: Choose serving accelerator memory and concurrency limits based on measured token or batch demand. B:Horizontal scaling raises concurrent capacity but does not necessarily accelerate computation within an individual inference request. It does not resolve the specific failure; the targeted correction is: Choose serving accelerator memory and concurrency limits based on measured token or batch demand. D:Endpoint network reachability and cloud identity authorization are separate security boundaries. It does not resolve the specific failure; the targeted correction is: Choose serving accelerator memory and concurrency limits based on measured token or batch demand.
Question 684A model endpoint has plenty of CPU capacity but repeatedly runs out of GPU memory for long prompts. What Google Cloud or model behavior explains the best solution? Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. Batching increases parallel accelerator use but adds queueing delay and may harm tail latency if configured poorly. Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. Model registry versioning connects a deployable model to metadata, approval and rollback history. ExplanationCorrect Answer C:Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. This supports the specific implementation: Choose serving accelerator memory and concurrency limits based on measured token or batch demand. A suitable independent check is: Replay long-context workloads and measure accelerator memory, queueing and p95 response time. Other Answers A:Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. The cause in the stated scenario is different: Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. B:Batching increases parallel accelerator use but adds queueing delay and may harm tail latency if configured poorly. The cause in the stated scenario is different: Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. D:Model registry versioning connects a deployable model to metadata, approval and rollback history. The cause in the stated scenario is different: Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use.
Question 685A model endpoint has plenty of CPU capacity but repeatedly runs out of GPU memory for long prompts. What is the most relevant acceptance check for the proposed ML change? Run smoke and load tests against the container image digest and confirm readiness and error handling. Replay long-context workloads and measure accelerator memory, queueing and p95 response time. Benchmark on-device accuracy, latency, memory and recovery after network disconnection. Measure cold-start frequency, p95 first-token delay and monthly idle cost with and without warm capacity. ExplanationCorrect Answer B:Replay long-context workloads and measure accelerator memory, queueing and p95 response time. This tests the underlying failure: Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. Other Answers A:A custom container controls dependencies and request handling but inherits security, compatibility and patching duties. That does not establish the required outcome; verify it by checking: Replay long-context workloads and measure accelerator memory, queueing and p95 response time. C:Edge inference trades centralized elasticity for local latency and offline availability with hardware constraints. That does not establish the required outcome; verify it by checking: Replay long-context workloads and measure accelerator memory, queueing and p95 response time. D:Warm replicas reduce first-request initialization latency but incur charges while receiving no traffic. That does not establish the required outcome; verify it by checking: Replay long-context workloads and measure accelerator memory, queueing and p95 response time.
Question 686A model endpoint has plenty of CPU capacity but repeatedly runs out of GPU memory for long prompts. Which repeatable process control should the ML engineering team adopt? Set serving batch limits and monitor tail latency before automatically increasing batching thresholds. Set tested context and concurrency limits and monitor memory saturation as traffic changes. Reserve quota headroom and revalidate autoscaling parameters during seasonal traffic planning. Make preprocessing parity and schema-contract checks mandatory in model rollout pipelines. ExplanationCorrect Answer B:Set tested context and concurrency limits and monitor memory saturation as traffic changes. This protects the selected approach because Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. The immediate solution is: Choose serving accelerator memory and concurrency limits based on measured token or batch demand. Other Answers A:Batching increases parallel accelerator use but adds queueing delay and may harm tail latency if configured poorly. The continuing safeguard required here is: Set tested context and concurrency limits and monitor memory saturation as traffic changes. C:Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. The continuing safeguard required here is: Set tested context and concurrency limits and monitor memory saturation as traffic changes. D:Training-serving feature skew can arise when model inputs are transformed inconsistently across environments. The continuing safeguard required here is: Set tested context and concurrency limits and monitor memory saturation as traffic changes.
Question 687A generative AI service's sequence lengths exceed the memory allocated to its serving accelerator. Which proposed approach best resolves the described model engineering problem? Register versioned model artifacts in Gemini Enterprise Agent Platform Model Registry and deploy approved versions by ID. Choose serving accelerator memory and concurrency limits based on measured token or batch demand. Profile per-request token generation, model size and decoding settings before scaling only the replica count. Configure bounded dynamic batching or compatible serving optimizations with a measured latency budget. ExplanationCorrect Answer B:Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. To validate the correction: Replay long-context workloads and measure accelerator memory, queueing and p95 response time. Other Answers A:Model registry versioning connects a deployable model to metadata, approval and rollback history. It does not resolve the specific failure; the targeted correction is: Choose serving accelerator memory and concurrency limits based on measured token or batch demand. C:Horizontal scaling raises concurrent capacity but does not necessarily accelerate computation within an individual inference request. It does not resolve the specific failure; the targeted correction is: Choose serving accelerator memory and concurrency limits based on measured token or batch demand. D:Batching increases parallel accelerator use but adds queueing delay and may harm tail latency if configured poorly. It does not resolve the specific failure; the targeted correction is: Choose serving accelerator memory and concurrency limits based on measured token or batch demand.
Question 688A generative AI service's sequence lengths exceed the memory allocated to its serving accelerator. What Google Cloud or model behavior explains the best solution? Edge inference trades centralized elasticity for local latency and offline availability with hardware constraints. Generative model outputs are probabilistic and require application validation before they can safely drive deterministic systems. Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. Hardware acceleration is beneficial only when supported operators and workload intensity justify the resource premium. ExplanationCorrect Answer C:Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. This supports the specific implementation: Choose serving accelerator memory and concurrency limits based on measured token or batch demand. A suitable independent check is: Replay long-context workloads and measure accelerator memory, queueing and p95 response time. Other Answers A:Edge inference trades centralized elasticity for local latency and offline availability with hardware constraints. The cause in the stated scenario is different: Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. B:Generative model outputs are probabilistic and require application validation before they can safely drive deterministic systems. The cause in the stated scenario is different: Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. D:Hardware acceleration is beneficial only when supported operators and workload intensity justify the resource premium. The cause in the stated scenario is different: Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use.
Question 689A generative AI service's sequence lengths exceed the memory allocated to its serving accelerator. Which evidence would best demonstrate that this ML solution meets the requirement? Compare per-request cost, p95 latency, memory saturation and measured throughput across hardware configurations. Load test p95 latency, error rate and scaling behavior under expected concurrent traffic. Replay long-context workloads and measure accelerator memory, queueing and p95 response time. Separate token throughput and queue time measurements and compare latency under the same per-request prompt. ExplanationCorrect Answer C:Replay long-context workloads and measure accelerator memory, queueing and p95 response time. This tests the underlying failure: Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. Other Answers A:Hardware acceleration is beneficial only when supported operators and workload intensity justify the resource premium. That does not establish the required outcome; verify it by checking: Replay long-context workloads and measure accelerator memory, queueing and p95 response time. B:Online prediction keeps inference capacity available for request-response workloads at the cost of capacity management. That does not establish the required outcome; verify it by checking: Replay long-context workloads and measure accelerator memory, queueing and p95 response time. D:Horizontal scaling raises concurrent capacity but does not necessarily accelerate computation within an individual inference request. That does not establish the required outcome; verify it by checking: Replay long-context workloads and measure accelerator memory, queueing and p95 response time.
Question 690A generative AI service's sequence lengths exceed the memory allocated to its serving accelerator. What preventive release checklist item best protects the intended outcome? Review endpoint exposure and identity policies before enabling any new inference consumer. Scan images, pin dependencies and test cold starts before releasing custom serving containers. Track serving performance by model and input length instead of only aggregate request rate. Set tested context and concurrency limits and monitor memory saturation as traffic changes. ExplanationCorrect Answer D:Set tested context and concurrency limits and monitor memory saturation as traffic changes. This protects the selected approach because Generative inference memory depends on model parameters, KV-cache and concurrent sequence lengths, not only average CPU use. The immediate solution is: Choose serving accelerator memory and concurrency limits based on measured token or batch demand. Other Answers A:Endpoint network reachability and cloud identity authorization are separate security boundaries. The continuing safeguard required here is: Set tested context and concurrency limits and monitor memory saturation as traffic changes. B:A custom container controls dependencies and request handling but inherits security, compatibility and patching duties. The continuing safeguard required here is: Set tested context and concurrency limits and monitor memory saturation as traffic changes. C:Horizontal scaling raises concurrent capacity but does not necessarily accelerate computation within an individual inference request. The continuing safeguard required here is: Set tested context and concurrency limits and monitor memory saturation as traffic changes.