How to Migrate NVIDIA GPU Scheduling from Device Plugins to Kubernetes DRA

This article provides a detailed walkthrough for migrating NVIDIA GPU scheduling from the traditional device plugin approach to Kubernetes Dynamic Resource Allocation (DRA). It explains the limitations of the device plugin model, such as the inability to specify GPU architecture, memory, or MIG profiles, and introduces DRA as a first-class API solution. The migration strategy emphasizes building a DRA-enabled node pool while keeping legacy nodes on the device plugin, avoiding dual allocation of the same GPU, and using Kubernetes 1.36's DRAExtendedResource bridge for compatibility. It recommends starting with full-GPU claims, validating scheduling evidence, and migrating workloads in waves, while noting that advanced features like dynamic MIG and MPS remain feature-gated.