PMLE Scaling Prototypes into ML Models — Timed 10 Questions

Question 1

You are deploying a deep learning model on edge devices with limited computational resources. The model must run inference in <10 ms and the model size must be under 50 MB. Currently, your trained model is 200 MB and runs in 50 ms. Which combination of model compression techniques should you apply?

10 minute time limit — choose an answer to begin.

Accepted Answer

Apply post-training quantization (INT8) and pruning. Post-training quantization (INT8) reduces model size ~4x and speeds inference by using 8-bit integers instead of 32-bit floats, while pruning removes redundant weights to further shrink the model and reduce compute. Together they can bring a 200 MB / 50 ms model under 50 MB and under 10 ms with minimal accuracy loss, making them the most practical combination for edge deployment. Quantization-aware training and knowledge distillation are heavier lifts that may not be necessary to hit these targets.

Answer

Only apply weight pruning

Answer

Apply quantization-aware training and knowledge distillation

Answer

Use knowledge distillation to train a smaller student model

Question 2

You are running a Vertex AI custom training job with pre-built TensorFlow container. You want to use TPU v3 pods for faster training. Which configuration is required?

Accepted Answer

Specify machine type as tpu-v3-8 in the worker pool spec and use tf.distribute.TPUStrategy. To use TPU v3 pods in a Vertex AI custom training job with a pre-built TensorFlow container, you specify the TPU machine type (e.g., tpu-v3-8) in the worker pool spec and use tf.distribute.TPUStrategy in your training code. The pre-built container already includes the TPU libraries, so no custom container is needed, and the distribution strategy is required for the code to address the TPU cores correctly.

Answer

Set the worker pool machine type to n1-standard-8 and add --accelerator type=tpu

Answer

Use a custom container with TPU libraries and set --tpu-topology=v3-8

Answer

Enable TPU by setting the environment variable TPU_NAME

Question 3

You are fine-tuning a large language model using Vertex AI Training with spot VMs to reduce cost. Your training job keeps getting preempted, causing delays. Which THREE strategies can help mitigate the impact of preemption?

Accepted Answer

Enable checkpointing and restore from the latest checkpoint after preemption. Option A is correct because periodic checkpointing to Cloud Storage lets the job resume from the last saved state after a spot VM is preempted, so no training progress is lost and the restart cost is bounded. Option B is correct because adding worker replicas (distributed data-parallel training) increases throughput, so the job completes more work per unit of wall-clock time and is more likely to finish before the next preemption event. Option E is correct because a smaller, more efficient architecture requires fewer compute cycles and less memory, shortening total training time and thus reducing exposure to preemption windows. Option C is not right because reducing batch size mainly lowers memory pressure and can even slow convergence, which does not directly mitigate preemption impact. Option D is not right because switching to dedicated, non-preemptible VMs eliminates preemption entirely rather than mitigating its impact, and it forfeits the cost savings that motivated using spot VMs.

Answer

Reduce batch size to lower memory usage

Answer

Switch to dedicated (non-preemptible) VMs

Question 4

You are building a machine learning pipeline on Google Cloud. You need to perform feature engineering on large datasets stored in BigQuery and store the resulting features in Vertex AI Feature Store for both online and offline use. Which TWO Google Cloud services should you use?

Accepted Answer

Dataflow. Dataflow can read from BigQuery, compute features via Apache Beam, and write to Feature Store. Alternatively, Dataproc can also do this but Dataflow is more serverless. Cloud Functions is not suitable for large-scale. Cloud Build is for CI/CD. BigQuery ML is for in-database ML.

Answer

Cloud Functions

Answer

BigQuery ML

Answer

Cloud Build

Question 5

A data scientist wants to train a TensorFlow model on Vertex AI using a pre-built container. Which of the following pre-built containers is NOT available for custom training in Vertex AI?

Accepted Answer

Apache Spark. Vertex AI provides pre-built containers for TensorFlow, PyTorch, scikit-learn, and XGBoost. There is no pre-built container for Apache Spark in the Vertex AI training service; Spark jobs are typically run on Dataproc.

Answer

TensorFlow

Answer

PyTorch

Answer

scikit-learn

Question 6

You are performing hyperparameter tuning on Vertex AI with Vizier. You want to maximize the accuracy of your model, and you have a budget of 50 trials. Which algorithm should you choose to best explore the search space?

Accepted Answer

Bayesian optimization. Vertex AI Vizier's default and recommended algorithm for hyperparameter tuning is Bayesian optimization, which builds a probabilistic surrogate model of the objective function and uses an acquisition function to pick the next trial most likely to improve accuracy. With a 50-trial budget, Bayesian optimization converges faster than random or grid search by exploiting information from prior trials. This makes it the best choice for maximizing accuracy within a fixed budget.

Answer

No algorithm; use default Vertex AI tuning

Answer

Grid search

Answer

Random search

Question 7

You are fine-tuning a pre-trained BERT model from Hugging Face for a sentiment analysis task using Vertex AI training. The dataset has 100k examples. To avoid catastrophic forgetting, which layer freezing strategy should you apply?

Accepted Answer

Unfreeze all layers and fine-tune the entire model. The option marked [CORRECT] is A because full fine-tuning of all layers is the standard and most effective approach for adapting a pre-trained BERT model to a downstream sentiment analysis task with a sufficiently large dataset (100k examples). Catastrophic forgetting is primarily a concern when the target dataset is small or when the new task is very different from the pre-training objective; with 100k examples, the model has enough signal to adjust all weights without losing general language understanding. Freezing layers would unnecessarily restrict the model's capacity and typically lead to lower accuracy. Thus, unfreezing all layers and fine-tuning the entire model is the recommended strategy.

Answer

Freeze the first 6 layers, fine-tune the last 6 layers

Answer

Freeze all layers except the classification head

Answer

Use transfer learning only on the embeddings layer

Question 8

Which Vertex AI service allows you to discover, fine-tune, and deploy foundation models with a few clicks, including models like Llama and Gemma?

Accepted Answer

Vertex AI JumpStart. Vertex AI JumpStart is the service within Vertex AI Model Garden that provides a curated collection of foundation models (including Llama, Gemma, and others) with one-click deployment, fine-tuning, and notebook samples. It simplifies the process of discovering and using open-source and third-party models. Model Garden is the broader catalog, but JumpStart is the specific feature for quick deployment and fine-tuning.

Answer

Vertex AI Prediction

Answer

Vertex AI Vizier

Answer

Vertex AI Model Garden

Question 9

You want to deploy a trained scikit-learn model to Vertex AI for online predictions. The model file is 2 GB. Which option should you use?

Accepted Answer

Use a custom container with the scikit-learn model. The correct option is B: use a custom container with the scikit-learn model. Vertex AI's pre-built scikit-learn containers have a model size limit (typically around 500 MB for the default serving container), so a 2 GB model file exceeds what the pre-built container can handle, and a custom container lets you package the model and any custom inference code without that constraint. Option D fails for the same size reason, since the pre-built scikit-learn container cannot accommodate a 2 GB artifact. Option A is not a real Vertex AI deployment mechanism (there is no 'large model VM' option for this purpose), and Option C is unsuitable because Cloud Functions has deployment package and memory limits far below 2 GB and is not designed for hosting ML models for online prediction.

Answer

Use AI Platform (now Vertex AI) with a large model VM

Answer

Use Cloud Functions to host the model

Answer

Upload the model to Vertex AI Model Registry and deploy using pre-built scikit-learn container

Question 10

You need to reduce the cost of training a large model on Vertex AI while maintaining fault tolerance. Which THREE actions should you take? (Choose 3)

Accepted Answer

Use spot VMs. Option A (Use spot VMs) is correct because spot (preemptible) VMs on Vertex AI cost significantly less than standard VMs, and fault tolerance is preserved when combined with checkpointing so training can resume after preemption. Option D (Enable checkpointing and save to Cloud Storage) is correct because periodic checkpoints written to durable Cloud Storage (GCS) allow the job to restart from the last saved state after a spot VM is preempted, which is the mechanism that makes cheap preemptible capacity safe to use. Option E (Use a single worker with multiple GPUs instead of multiple workers) is correct because consolidating onto one worker with multiple GPUs avoids the inter-worker communication overhead and synchronization complexity of distributed multi-worker training, reducing cost and failure surface while still using multiple accelerators. Option B (MultiWorkerMirroredStrategy) is not appropriate here because it is a distributed training strategy that increases coordination overhead and fault-tolerance complexity rather than reducing cost. Option C (Use TPUs instead of GPUs) is not a guaranteed cost-reduction action for this scenario, since TPU pricing and availability vary and switching hardware does not by itself provide fault tolerance.

Answer

Use MultiWorkerMirroredStrategy

Answer

Use TPUs instead of GPUs