PMLE Serving and Scaling Models Practice Test 3 — 15 Questions

Question 1

A company wants to deploy a TensorFlow model on edge devices for real-time inference without internet connectivity. Which Vertex AI service should they use to manage the deployment?

Accepted Answer

Vertex AI Edge Manager. Vertex AI Edge Manager allows you to deploy models to edge devices and manage them centrally.

Answer

TensorFlow Lite Converter

Answer

Vertex AI Endpoint

Answer

AI Platform Prediction (legacy)

Question 2

Which machine type is most suitable for a Vertex AI endpoint serving a GPU-accelerated model?

Accepted Answer

n1-standard-4 with attached GPU. Vertex AI endpoints require a machine type that supports GPU acceleration, and the n1-standard-4 with an attached GPU provides the necessary CPU-to-GPU balance for inference workloads. The n1 series supports GPU attachments via the `accelerator` configuration, enabling CUDA-based model serving, while the e2 series and n1-highmem-8 without GPU cannot leverage GPU acceleration.

Answer

e2-standard-4

Answer

n1-highmem-8

Answer

n1-standard-4

Question 3

A team wants to deploy a model on Vertex AI Edge Manager for offline inference on edge devices. Which three steps are required? (Choose 3)

Accepted Answer

Package the model and deploy using Vertex AI Edge Manager. Option B is correct because deploying a model for offline edge inference requires packaging the model artifacts and deploying them through Vertex AI Edge Manager, which handles the distribution of models to edge devices. Option C is correct because edge devices typically require optimized, lightweight model formats such as TensorFlow Lite or ONNX to perform efficient offline inference with limited compute and memory resources. Option D is correct because the model must first be uploaded to Vertex AI Model Registry, which serves as the central repository for managing model versions before they can be packaged and deployed to edge devices via Edge Manager. Option A is incorrect because Vertex AI Model Monitoring is used for detecting skew and drift in deployed models, not for preparing or deploying models to edge devices. Option E is incorrect because creating a Vertex AI Endpoint with an HTTP endpoint is for online prediction serving in the cloud, not for offline inference on edge devices.

Answer

Enable Vertex AI Model Monitoring

Answer

Create a Vertex AI Endpoint with HTTP endpoint

Question 4

A company uses Vertex AI for online predictions with a large ensemble model that requires GPU acceleration. They want to reduce inference latency by batching multiple requests into a single GPU inference call. What should they configure?

Accepted Answer

Deploy the model with NVIDIA Triton Inference Server configured for dynamic batching.. NVIDIA Triton Inference Server supports dynamic batching, which automatically groups multiple inference requests into a single GPU call. This reduces overhead and improves GPU utilization, directly addressing the need to lower latency for online predictions with a large ensemble model on Vertex AI.

Answer

Use Vertex AI Model Optimization for automatic compilation.

Answer

Increase the number of GPU replicas to handle higher concurrency.

Answer

Enable model quantization using TensorRT.

Question 5

A data scientist wants to deploy a model trained with PyTorch to a Vertex AI endpoint for online predictions. What is the recommended approach?

Accepted Answer

Package the model in a custom container with a web server (e.g., FastAPI) and deploy to Vertex AI.. Vertex AI requires a custom container for PyTorch models, as it does not provide a pre-built PyTorch serving container. The recommended approach is to package the trained PyTorch model with a web server like FastAPI (or Flask) that loads the model and exposes an HTTP endpoint for predictions. This container is then deployed to Vertex AI for online predictions, giving full control over the inference environment and dependencies.

Answer

Use Vertex AI's pre-built PyTorch container and upload the state dictionary.

Answer

Export the PyTorch model to a SavedModel format and deploy using Vertex AI's pre-built TensorFlow container.

Answer

Convert the model to TensorFlow.js and deploy to Cloud Functions.

Question 6

A team is deploying a model that has strict latency requirements: p99 response time under 100 ms. The model is CPU-only and will receive up to 1000 QPS. They want to minimize cost while meeting the SLO. Which machine type and scaling configuration is most appropriate?

Accepted Answer

n1-standard-4 with min_replicas=1 and max_replicas=5, CPU utilization target 60%. It uses a CPU-only machine (n1-standard-4) with autoscaling based on CPU utilization target of 60%, which balances cost and performance for a latency-sensitive, CPU-bound inference workload at 1000 QPS. The min_replicas=1 ensures a baseline capacity, while max_replicas=5 allows scaling to handle spikes without over-provisioning, keeping p99 under 100 ms.

Answer

GPU-enabled machine with min_replicas=1 and max_replicas=2

Answer

n1-standard-8 with min_replicas=3 and max_replicas=3 (fixed)

Answer

n1-highmem-2 with min_replicas=2 and max_replicas=10

Question 7

A company is deploying a complex model that requires GPU for inference. They want to use Vertex AI for serving. Which TWO steps are required to deploy the model with GPU support? (Choose 2)

Accepted Answer

Select a GPU-enabled machine type such as n1-standard-4 with 1 x NVIDIA Tesla T4.. Option A is correct because deploying a model on Vertex AI with GPU acceleration requires choosing a GPU-enabled machine type, such as n1-standard-4 with 1 x NVIDIA Tesla T4, in the DeployedModel's machineSpec (acceleratorType and acceleratorCount) so the serving node is provisioned with the requested GPU. Option C is correct because the model's serving container must include the GPU libraries needed to run inference on the hardware, i.e., a custom container with a compatible CUDA and cuDNN stack (matching the framework and driver requirements), since a standard CPU-only prebuilt container cannot use the GPU. Option B is not required: Vertex AI Model Optimization is a separate service for model compression/latency tuning and does not perform automatic GPU compilation for deployment. Option D is not required: minimum replica count is a scaling/availability setting, not a GPU enablement step, and GPU redundancy is not mandated for deployment. Option E is not required: gRPC can reduce latency for prediction traffic but is an optional protocol choice and has nothing to do with enabling GPU support.

Answer

Enable Vertex AI Model Optimization for automatic GPU compilation.

Answer

Increase the minimum replicas to at least 2 for GPU redundancy.

Answer

Use gRPC protocol for prediction requests to reduce latency.

Question 8

An ML engineer needs to deploy a model to Vertex AI for online predictions and enable autoscaling to zero when not in use. Which THREE conditions must be met? (Choose 3)

Accepted Answer

Enable serverless serving by selecting the appropriate serving mode.. Option A is correct because autoscaling to zero on Vertex AI requires serverless serving mode, which is selected when deploying the model to an endpoint and allows the endpoint to scale down to zero replicas when there is no traffic. Option D is correct because setting min_replica_count=0 in the endpoint deployment configuration explicitly permits the deployed model to scale down to zero replicas during idle periods. Option E is correct because max_replica_count must be greater than 0 to define the upper scaling bound and allow the endpoint to scale up from zero when requests arrive. Option B is not required because autoscaling to zero is a serving-mode and replica-count configuration concern, not a GPU requirement, and GPUs do not inherently enable scale-to-zero. Option C is not required because WebSockets support in a custom container is unrelated to Vertex AI's scale-to-zero behavior, which is controlled by serverless serving and replica settings.

Answer

Use a GPU-enabled machine type for faster scaling.

Answer

Deploy the model using a custom container with WebSockets support.

Question 9

Which TWO of the following are benefits of using Vertex AI Matching Engine (Vector Search) over a brute-force nearest neighbor search? (Choose 2)

Accepted Answer

Lower query latency for large datasets. Option C is correct because Vertex AI Matching Engine (Vector Search) uses scalable approximate nearest neighbor (ANN) indexes such as ScaNN/IVF, which avoid comparing a query against every vector and therefore deliver much lower query latency for large datasets than brute-force search. Option D is correct because brute-force search must keep all embeddings resident and scan them exhaustively, whereas Matching Engine uses optimized index structures and sharding that reduce the memory footprint required per query workload. Option A is not correct because brute-force search is generally simpler to implement, while Matching Engine requires index creation, deployment, and endpoint management. Option B is not correct because Matching Engine indexes typically require rebuilds or streaming updates with limitations, not fully real-time updates without any index rebuild. Option E is not correct because Matching Engine returns approximate nearest neighbors, not guaranteed exact nearest neighbor results.

Answer

Simpler to implement than brute-force

Answer

Supports real-time updates without index rebuild

Answer

Guaranteed exact nearest neighbor results

Question 10

A company runs batch predictions on Vertex AI every hour using a custom container. They want to reduce costs by minimizing idle time while ensuring the batch job completes within 10 minutes. Which endpoint configuration should they use?

Accepted Answer

Use Vertex AI batch prediction job with a custom service account and set machine_type to 'n1-standard-4' and batch_size to optimize throughput.. Vertex AI batch prediction jobs are designed for high-throughput, asynchronous processing of large datasets without maintaining persistent infrastructure. By tuning `machine_type` and `batch_size`, you can minimize idle time and ensure the job completes within the 10-minute window, as the job only runs while actively processing and scales resources as needed.

Answer

Use Vertex AI online prediction with a load balancer in front to distribute requests.

Answer

Create a Dataflow pipeline to read from BigQuery and write predictions to GCS, using the trained model as a side input.

Answer

Deploy the model to an endpoint with min_replicas=0 and max_replicas=10, then send batch requests to the endpoint.

Question 11

A recommendation model is deployed to a Vertex AI Endpoint. During a flash sale, request volume triples within two minutes. The endpoint scales out, but new replicas take 90 seconds to become ready because the model loads a large embedding table from Cloud Storage at startup. You must reduce the time to serve traffic for new replicas without changing the model. What should you do?

Accepted Answer

Package the model into a custom container that loads the embedding table during image build, and deploy with the container.. The startup delay is caused by loading a large embedding table from Cloud Storage when a replica initializes. A custom container that includes the table in the image eliminates that remote read, so replicas can serve traffic sooner. Other approaches either do not affect load time or only add capacity without reducing the per-replica cold-start duration.

Answer

Enable request-response logging on the endpoint and use the logs to warm the model.

Answer

Upload the model artifacts to a Cloud Storage bucket in the same region as the endpoint.

Answer

Configure the endpoint with a higher minReplicaCount and set up scheduled scaling based on historical traffic patterns.

Question 12

You have a Vertex AI endpoint serving a model that requires a GPU. During a traffic spike, new replicas take several minutes to become ready because the GPU nodes must be provisioned and the model loaded. You need to reduce the time it takes for new replicas to serve traffic. Which configuration should you use?

Accepted Answer

Configure the model container to use a warm-up routine that preloads the model and caches it in memory.. A warm-up routine in the model container loads the model into GPU memory and runs a dummy prediction before the container reports healthy. This ensures that the replica is fully initialized when it starts receiving traffic, and it reduces the observed time from replica creation to first successful request. It is the most direct way to shorten the readiness time of new replicas.

Answer

Use a Dedicated Endpoint with a higher number of pre-provisioned replicas.

Answer

Enable autoscaling with a lower target CPU utilization threshold.

Answer

Set the endpoint's minReplicaCount to 1 and maxReplicaCount to 10.

Question 13

A recommendation model on a Vertex AI endpoint is accessed by two clients: an internal dashboard that sends a handful of requests per hour, and a mobile app that occasionally bursts to several thousand requests per second. The dashboard must never experience cold-start latency, but the mobile bursts are short-lived and cost is a concern. The endpoint currently uses minReplicaCount=1 and maxReplicaCount=20. Which configuration should you apply?

Accepted Answer

Set minReplicaCount=2 and maxReplicaCount=20, and configure autoscaling metrics to include CPU utilization and request count per replica.. The dashboard demands zero cold-start latency, so the minimum replica count must be at least two. The mobile bursts are short and cost-sensitive, so the maximum should stay bounded and the autoscaler should watch both CPU and request count to react quickly. Combining a higher floor with dual metrics satisfies both requirements without overprovisioning.

Answer

Set minReplicaCount=1 and maxReplicaCount=40, and enable autoscaling based solely on CPU utilization.

Answer

Set minReplicaCount=2 and maxReplicaCount=40, and enable scale-to-zero during idle periods.

Answer

Set minReplicaCount=2 and maxReplicaCount=20, and enable request-response logging on the endpoint.

Question 14

You have deployed a Vertex AI model to an endpoint that serves online predictions with a strict P99 latency SLO of 200 ms. The model server logs show that most of the time is spent in pre-processing input features, which includes several string transformations and a lookup in a large embedding table. You want to reduce latency without retraining the model. What should you do?

Accepted Answer

Use Vertex AI Feature Store to precompute and serve the transformed features, and have the model server consume them directly.. The bottleneck is the preprocessing work performed on every online prediction request. Offloading that work to Vertex AI Feature Store allows the model server to retrieve already-transformed features, eliminating the string operations and embedding lookup from the serving path. This reduces per-request latency without changing the model or its accuracy, directly meeting the strict SLO.

Answer

Move the pre-processing logic into a custom prediction routine that runs inside the model server.

Answer

Enable request-response logging on the endpoint and analyze the logs to identify slow transformations.

Answer

Increase the number of replicas on the endpoint to distribute the pre-processing load across more machines.

Question 15

A media company is deploying a video recommendation model on Vertex AI. The model expects input features that include user watch history (up to 1000 recent video IDs) and video metadata. The features are stored in BigQuery and must be joined at serving time. They need to minimize prediction latency and ensure feature consistency between training and serving. What should they do?

Accepted Answer

Use Vertex AI Feature Store to ingest features from BigQuery, and configure the endpoint to fetch features online at serving time using entity IDs.. Vertex AI Feature Store ingests features from BigQuery and serves them online with low latency, ensuring training-serving consistency. It avoids custom code and provides a managed solution for real-time feature retrieval, which is ideal for dynamic features like watch history.

Answer

Precompute all possible feature combinations and store them in a Redis cache, then have the model container look up features by user ID and video ID.

Answer

Embed a BigQuery client in the model container and execute SQL queries at prediction time to fetch the required features.

Answer

Export features from BigQuery to Cloud Storage as CSV files, and have the model container read the CSV files at startup to build an in-memory lookup table.