v1.0.0
First official release of the AMD GPU Dynamic Resource Allocation (DRA) driver for Kubernetes.
Highlights
- DRA driver for AMD GPUs — kubelet plugin that publishes a
gpu.amd.comDeviceClass and servesResourceSlice/ResourceClaimallocations usingresource.k8s.io/v1. - GPU discovery & attributes — per-device attributes including PCI device ID,
pciBusID, NUMA node, and compute partition profile, published via ResourceSlices. - Compute partitioning — partition-aware allocation; non-partitionable GPUs (e.g. MI210) are handled correctly and do not advertise a
partitionProfile. - Helm chart packaged and auto-published to OCI and
gh-pages:driver-initcontainer waits for theamdgpukernel driver before the plugin starts.- Configurable CDI host path (
cdi.dynamicPath). - Selectable DRA API version at install time (
v1/v1beta2/v1beta1). deviceClass.createflag to disable DeviceClass creation when co-installed withgpu-operator.
- RHEL UBI-based driver image; libdrm/CGo dependencies removed for a smaller, simpler build.
- CI/CD — automated build, test, lint, and release workflows.
- Docs — Installation & Developer Guide, Demo guide, Driver Attributes reference, CLI options reference, and a Troubleshooting & Known Limitations guide. GitHub issue templates included.
Requirements
- Kubernetes 1.32+ (DRA beta). Examples use
resource.k8s.io/v1(K8s 1.34+); adjustapiVersionfor older clusters. amdgpukernel driver loaded on each GPU node.- Helm v3.7.0+.
Quick Start
# Install Helm
curl -fsSL -o get_helm.sh https://raw.githubusercontent.com/helm/helm/main/scripts/get-helm-3
chmod 700 get_helm.sh
./get_helm.sh
# Add the Helm repository
helm repo add rocm-k8s-gpu-dra-driver https://rocm.github.io/k8s-gpu-dra-driver
helm repo update
# Install the DRA driver
helm install k8s-gpu-dra-driver rocm-k8s-gpu-dra-driver/k8s-gpu-dra-driver \
--namespace kube-amd-gpu \
--create-namespaceStatus & Known Limitations
- The alpha/experimental flag has been removed for this release, but operators should review the Troubleshooting & Known Limitations guide before production rollout.
- DRA itself is still beta upstream in some early version Kubernetes; cluster admins must ensure the DRA feature gate and APIs are enabled.