GPU Management on Kubernetes
GPU management turns physical accelerators into discoverable, shareable and allocatable Kubernetes resources. Configuration defines the device class, sharing mode and resource profile, while device management components expose available capacity to the scheduler or device allocator.
Explore the interactive architecture →Device discovery and configuration
Cluster components detect accelerator hardware and expose device capabilities. Configuration determines which device classes are available and how the platform presents them to workloads.
Partitioning and sharing
Depending on hardware and platform policy, a physical GPU can be offered as a dedicated device, a hardware partition or a shared resource profile. GPU management advertises those resources to Kubernetes rather than deciding which model request runs next.
Resource allocation
The Kubernetes scheduler and device allocation mechanisms place a workload and assign an eligible device. Placement, topology and isolation belong to the infrastructure layer; token generation and KV-cache behavior belong to the model inference runtime above it.