Deployment - FSDL 2022

Explore the process of transforming a promising machine learning model into a valuable ML-powered product in this comprehensive lecture. Learn about various deployment architectures, including model-in-server, model-in-database, and model-as-a-service. Discover how to create prototypes using tools like Gradio and Streamlit, implement REST APIs, manage dependencies, and containerize services with Docker. Delve into performance optimization techniques for both CPUs and GPUs, including distillation, quantization, caching, and batching. Examine horizontal scaling strategies, container orchestration with Kubernetes, and serverless options. Investigate rollout techniques such as shadow and canary deployments, and explore managed services like AWS SageMaker. Finally, gain insights into edge deployment, efficient model creation for edge devices, and key takeaways for successfully deploying ML models in various environments.

Overview
First, deploy a prototype with gradio or streamlit
Model-in-server architecture
Model-in-database architecture
Model-as-a-service architecture
REST APIs for model services
Dependency management for model services
Containerization for model services with Docker
Performance optimization: to GPU or not to GPU?
Optimization for CPUs: distillation, quantization, and caching
Optimization for GPUs: Batching and GPU sharing
Libraries for model serving on GPUs
Horizontal scaling
Horizontal scaling with container orchestration k8s
Horizontal scaling with serverless services
Rollouts: shadows and canaries
Managed options for model serving AWS Sagemaker
Takeaways on model services
Moving to edge
Frameworks for edge deployment
Making efficient models for the edge
Mindsets and takeaways for edge deployment
Takeways for deploying ML models