Accelerate AI adoption
summary
Explains how a notebook connects to a GKE-managed TPU cluster to run large-model training end-to-end: create the GKE TPU cluster outside the notebook, start iPython kernels on all chips, install the iPython client, then control runs from the notebook. Shows Llama 2 70B on 64 TPU chips using AI Hypercomputer hardware and GKE orchestration, compares two variants of fully sharded data parallelization, and reports live outputs for choosing the best optimization.
AI-written summary · about
chapters (4)
related talks
38mAI in Kubernetes: How to Get Started?
46mWhen AI Meets High Performance Computing
27mHarnessing Kubernetes for Scalable AI/ML Workloads
12mAI-powered infrastructure:
34mGPU as a Service Over K8s: Drive Productivity and Increase Utilization
33mAI Workspaces for the Modern Era – Kubeflow Notebooks v2
54mEconomics and Best Practices of Running AI/ML Workloads on Kubernetes
36mAccelerating High-Performance Machine Learning at Scale i...