MLOps and LLMOps BootCamp

Subtopics:

• DevOps vs. MLOps
• The MLOps Lifecycle Stages
• Intro to GCP for ML & Vertex AI Platform
• Core Components: Projects, IAM, Cloud Storage, Vertex AI Workbench

Learning Outcomes:

• Grasp the purpose of MLOps and become proficient in setting up core MLOps infrastructure on Google Cloud.

Hands-on/Lab:

• Set up a GCP project, enable Vertex AI APIs, and launch a managed notebook in Vertex AI Workbench.

Subtopics:

• Orchestration with Vertex AI Pipelines
• Building with the Kubeflow Pipelines (KFP) SDK
• Using pre-built Google Cloud Pipeline Components
• Designing reusable & parameterized pipelines

Learning Outcome:

• Build, execute, and manage automated, multi-step ML workflows specifically on Vertex AI.

Hands-on/Lab:

• Construct a complete training pipeline using the KFP SDK and run it on Vertex AI Pipelines.

Subtopics:

• Tracking with Vertex AI Experiments
• Hyperparameter Tuning with Vertex AI Vizier
• Versioning & Staging in the Vertex AI Model Registry
• Integrating MLflow with Vertex AI (optional)

Learning Outcome:

• Systematically track ML experiments for reproducibility and manage the model lifecycle using Vertex AI’s native tools.

Hands-on/Lab:

• Run an HPT job using Vertex AI Vizier, log results in Vertex AI Experiments, and register the best model in the Vertex AI Model Registry.

Subtopics:

• CI/CD Principles for ML
• Git with Cloud Source Repositories or GitHub
• Automating pipelines with Cloud Build triggers
• Packaging code and dependencies for CI/CD

Learning Outcome:

• Implement robust MLOps automation using Google Cloud’s native CI/CD services to link Git commits to model production.

Hands-on/Lab:

• Create a CI/CD workflow using Cloud Build that automatically triggers a Vertex AI Pipeline run when code is pushed to a repository.

Subtopics:

• Deployment Strategies: Real-time vs. Batch
• Deploying to Vertex AI Endpoints
• Running Vertex AI Batch Prediction jobs
• Containerizing models with Artifact Registry

Learning Outcome:

• Deploy trained models as scalable, secure prediction services for both online and offline use cases on Vertex AI.

Hands-on/Lab:

• Deploy a model from the Vertex AI Model Registry to a live Vertex AI Endpoint. Separately, run a large-scale Batch Prediction job.

Subtopics:

• Detecting Training-Serving Skew & Prediction Drift
• Monitoring model performance with Vertex AI Model Monitoring
• Alerting with Cloud Monitoring
• Data Validation with Great Expectations

Learning Outcome:

• Implement an integrated monitoring solution to detect model degradation and automatically trigger alerts.

Hands-on/Lab:

• Configure Vertex AI Model Monitoring for a deployed endpoint. Simulate data drift and demonstrate a triggered alert in Cloud Monitoring.

Subtopics:

• Security with GCP IAM & Secret Manager
• Fairness & Bias with Vertex Explainable AI
• Unit & Integration testing for pipelines
• Scaling with custom jobs on Google Kubernetes Engine (GKE)

Learning Outcome:

• Apply principles for building secure, governable, explainable, and scalable ML systems on Google Cloud.

Hands-on/Lab:

• Secure a Vertex AI Endpoint using IAM roles. Use Vertex Explainable AI to get feature attributions for a model.

Subtopics:

• Need for a Feature Store
• Vertex AI Feature Store for online/offline consistency
• Data Versioning with DVC & Cloud Storage

Learning Outcome:

• Manage ML features centrally and version datasets to ensure full pipeline reproducibility on GCP.

Hands-on/Lab:

• Use DVC with Google Cloud Storage to version a dataset. Set up and use the Vertex AI Feature Store for training and serving.