Question

Your team has a research prototype written in PyTorch that trains a custom convolutional network on a single GPU. Leadership wants to productionize it on Vertex AI with minimal rewrite, full support for distributed multi-worker training, and native model serving through Vertex AI Prediction with GPU acceleration. The lead researcher is comfortable keeping the code in PyTorch. What is the most appropriate approach for scaling this prototype into a production training and serving pipeline?

Accepted Answer

Keep the code in PyTorch, package it in a custom training container, use torch.distributed with Vertex AI's distributed training worker pools, and deploy with a prebuilt PyTorch serving container.

Answer

Rewrite the model in TensorFlow because Vertex AI only supports distributed training and GPU serving for TensorFlow models.

Answer

Convert the model to BigQuery ML so it can scale automatically without managing any training infrastructure.

Answer

Keep the model in PyTorch but train only on a single larger GPU (A100), since PyTorch cannot be distributed across multiple Vertex AI workers.