Vertex AI ML Workflows

• Vertex AI is the unified platform for the whole custom-ML lifecycle
• Use Vertex AI Pipelines to automate, reproduce, and audit the ML lifecycle
• Vertex AI Pipelines tracks artifact lineage in Vertex ML Metadata
• Pipeline execution caching skips steps whose inputs have not changed
• Pick TPUs for matrix-bound, weeks-long training with no custom ops
• Pick GPUs when the model needs custom ops or framework flexibility
• AI Hypercomputer is an integrated stack, not a single product
• Commit to reservations or CUDs for steady large training at guaranteed capacity
• Use Dynamic Workload Scheduler for short accelerator runs that tolerate a wait
• Run interruption-tolerant training on Spot for the deepest discount
• Choose online Endpoints when a caller waits on each prediction
• Choose batch prediction for bulk scoring with no waiting caller
• Centralize features in Vertex AI Feature Store to share and avoid skew
• Managed datasets centralize training data and apply consistent splits
• AutoML trains a custom model with minimal code; custom training gives full control
• Register a trained model in the Model Registry before deploying it
• Distributed training scales a job across multiple workers with accelerators
• Use BigQuery ML for SQL-shaped modelling, not a separate Vertex training stack
• Grant the Notebooks Service Agent (not the user) compute.networkUser on the host project for Workbench in a Shared VPC
• Cross-project custom training images need Artifact Registry Reader on the Vertex AI Service Agent
• Custom containers report hyperparameter-tuning metrics with the cloudml-hypertune library
• Build GPU custom training containers on the nvidia/cuda base image

Unlock with Premium — includes all practice exams and the complete study guide.