Scaling prototypes into ML models

Free Google Cloud Professional Machine Learning Engineer practice — 6 questions on Scaling prototypes into ML models, with explanations. No sign-up. Full 12-question mixed test →

Question 1 of 6 · Scaling prototypes into ML models

A team trains a large deep learning model using multi-node, multi-GPU training on Vertex AI Custom Training with 4 A100 GPUs per node across 8 nodes. Profiling shows the all-reduce step consumes most of the step time due to network bandwidth saturation between nodes. Which change should they make to reduce this overhead with minimal code changes?

Reduction Server is a Vertex AI feature that adds lightweight reducer VMs specifically to offload the all-reduce gradient aggregation from GPU workers, cutting network communication time for synchronous distributed training with minimal code changes (just a strategy/config change).

Question 2 of 6 · Scaling prototypes into ML models

A prototype PyTorch model uses a custom CUDA kernel to implement a specialized loss function that has no equivalent XLA op. The team must scale training on Vertex AI from days to hours while keeping this custom kernel. Which hardware choice is correct?

Custom CUDA kernels only run on NVIDIA GPUs; TPUs require ops to be expressible in XLA, so any custom CUDA/C++ kernel is unsupported there. GPUs with data-parallel distribution meet both the compatibility requirement and the throughput requirement.

Question 3 of 6 · Scaling prototypes into ML models

You are configuring a Vertex AI hyperparameter tuning job. The 'optimizer' parameter is categorical (adam, sgd), and the 'momentum' parameter is only meaningful when optimizer=sgd. You want Vizier to avoid wasting trials searching momentum values when optimizer=adam. What should you configure?

Vertex AI Vizier supports conditional (parent-child) parameter specifications, letting you declare that a parameter like momentum is only sampled when its parent parameter (optimizer) takes a specific value, preventing wasted trials on invalid combinations.

Question 4 of 6 · Scaling prototypes into ML models

A team fine-tunes a pretrained ResNet-50 (ImageNet) on 500 labeled medical images by unfreezing all layers and training with the same high learning rate originally used for pretraining. Validation accuracy ends up worse than a baseline that only trained a new classification head on frozen features. What is the most likely cause and best fix?

Fully unfreezing a pretrained network and training with the original (typically large) pretraining learning rate is the classic recipe for catastrophic forgetting — large gradient updates overwrite the useful pretrained representations. The standard fix is a small fine-tuning learning rate (often 10-100x smaller) and/or gradual unfreezing.

Question 5 of 6 · Scaling prototypes into ML models

During training, both training loss and validation loss plateau at high values and neither improves with more epochs, even after increasing model capacity (adding layers). The team then removes dropout (originally set to 0.5 on every layer) and adds data augmentation; both training and validation loss drop substantially and the model performs well. What was the underlying issue?

Both training and validation loss being high and plateaued together (not diverging) is the signature of underfitting, not overfitting. Excessive dropout across every layer over-regularizes the network, capping its effective capacity; reducing it lets the model fit the training data, which is why both curves improved together.

Question 6 of 6 · Scaling prototypes into ML models

A data scientist has a Jupyter notebook prototype that uses scikit-learn for feature engineering and XGBoost with a custom objective function not supported by any prebuilt Vertex AI training container. To scale this into a reproducible, production-grade Vertex AI Custom Training workload with full control over the runtime environment and pinned dependency versions, what should be done first?

A custom objective function unsupported by prebuilt containers requires full control over the runtime, which Vertex AI achieves via custom containers: a Dockerfile pins exact dependency versions and packages the custom logic, and the image is pushed to Artifact Registry for use in a Vertex AI custom training job — the standard, reproducible path for scaling beyond prebuilt container capabilities.

Ready for the real thing?

The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.

$129.99 $34.99 with code FREETEST33 — valid through September 7.

Get my $34.99 deal →