Your team trains a transformer-based recommendation model written in JAX. Training runs involve very large matrix multiplications, use a fixed model architecture, and run for many hours at high utilization. You need to select the accelerator that minimizes cost per training run. What should you do?

You need to deploy a scikit-leam classification model to production. The model must be able to serve requests 24/7, and you expect millions of requests per second to the production application from 8 am to 7 pm. You need to minimize the cost of deployment. What should you do?

During batch training of a neural network, you notice that there is an oscillation in the loss. How should you adjust your model to ensure that it converges?

You are working on a prototype of a text classification model in a managed Vertex AI Workbench notebook. You want to quickly experiment with tokenizing text by using a Natural Language Toolkit (NLTK) library. How should you add the library to your Jupyter kernel?

Your production TensorFlow pipeline computes vocabulary mappings and z-score normalization statistics over the full training dataset. Engineers reimplemented the same transformations in the serving application, and subtle differences have caused prediction errors twice this year. You need to guarantee that serving transformations exactly match training transformations. What should you do?