This tutorial is tailored for Kubernetes and Devops engineers looking to deepen their understanding of machine learning operations with a focus on Kubernetes and containers.
To put this in perspective, there are 3 different personas.
- Application developers using APIs (eg. Openai api) to build applications. This is not relevant to you.
- Data Scientists building and finetuning models as a product. This is not relevant to you.
- Engineers looking to build/operate/learn about ML infrastructure and processes. This is for you.
We'll use MicroK8s on Ubuntu 22.04, running on Paperspace cloud by DigitalOcean, giving you real-world experience in managing ML workflows in Kubernetes.
Why MicroK8s?
- Thin footprint and low operational overhead to run on a single VM. Hence useful for learning.
- Highly available and production-ready
- 1. Basics of GPU
- 2. MLOps and Developer Experience
- 3. GPU + Container
- 4. GPU + Kubernetes
- 5. Multi-instance GPU
- 6. Ansible Setup
Learn the fundamentals of GPU technology, with a focus on NVIDIA GPUs. We'll cover hardware checks, driver installation, and dive into GPU compute, memory, and scheduling, including hands-on command-line examples.
Compare between DevOps and MLOps for Machine learning. While MLOps tooling can be very broad, start with developer view and add the tool only when there is a need.
Set up container environments and deploying GPU-based applications.
Explore how GPUs integrate with Kubernetes, using Microk8s. We self-host Microk8s/Ubuntu on Paperspace.
Optimize Kubernetes for complex scenarios like multi-instance GPUs, enhancing resource utilization and ML workload performance.
Set up a GPU-ready, multi-node Microk8s cluster with shared storage on Paperspace using Ansible with opinionated settings.