At KubeCon Europe, NVIDIA announced today that it is open sourcing KAI Scheduler, a GPU-centric Kubernetes scheduler that was originally developed by Run:ai, which NVIDIA acquired last year. Available under the Apache 2.0 license, KAI Scheduler helps its users optimize GPU resource allocations for AI and machine learning workloads in GPU clusters.
NVIDIA argues that traditional resource schedulers are ill-suited for managing AI workloads because GPU demand can fluctuate quite a lot, with bursty inference workloads and sustained model…