kubernetes-sigs/dra-driver-nvidia-gpu v0.5.0 on GitHub

This release expands GPU and ComputeDomain deployment support and improves overall reliability, feature coverage, and platform compatibility.

New Features

Support deployment with the NVIDIA GPU Operator v26.7.0.
Host-managed IMEX for ComputeDomains
Passthrough support on NVIDIA Grace and Blackwell systems, including VFIO module variants.
Fabric Manager partition activation and deactivation for supported multi-GPU passthrough configurations
Multi-user MPS support
Optional management of the nvidia.com/gpu.clique node label to be compatible with GPU Feature Discovery (GFD).
Standard NUMA node attributes for GPUs
System mediated GPU sharing across namespaces using Consumable Capacity feature in DRA.
Enhanced documentation using Hugo and Docsy.

Feature gate changes

Feature gate	Stage	Default	Description
`HostManagedIMEXDaemon`	Alpha	`false`	Allows `resources.computeDomains.imex.mode=hostManaged`, where you manage the host `nvidia-imex` service instead of the DRA Driver creating per-ComputeDomain daemon DaemonSets. This gate only unlocks the mode. Driver-managed DNS daemon naming and `ComputeDomainClique` tracking are not used in this mode. See Validate host-managed IMEX.
`FabricManagerPartitioning`	Alpha	`false`	Enables Fabric Manager partition discovery and lifecycle management for full GPUs and VFIO devices on supported HGX and single-node NVL systems. VFIO devices require `PassthroughSupport`. Requires Fabric Manager with `FABRIC_MODE=1`. Each full-GPU or VFIO claim on a participating node must exactly match one published partition. See Fabric Manager partitioning.
`ConsumableShares`	Alpha	`false`	Implements Kubernetes DRA consumable capacity for full GPUs and MIG devices by publishing them as multi-allocatable and defining their capacity request policies. Also set the `consumableShares` Helm value to select an accounting mode. See Consumable capacity.
`DRAListTypeAttributes`	Alpha	`false`	Publishes list-valued DRA device attributes, including `resource.kubernetes.io/numaNode` as a one-element list. See NUMA locality.

Bug Fixes

Fix upgrade issues around compute domain by @shengnuo in #1129
Remove hardcoded DeviceTaintEffectNone by @guptaNswati in #1131
Adding imagepullsecret and pullpolicy for MPS deployment by @visheshtanksale in #1175
Stop injecting implicit TimeSlicing and reset on unprepare by @guptaNswati in #1183
Fix: Use kubelet-plugin service account for MPS control daemon by @shivamerla in #1213
Skip non-regular files in findFile library/binary discovery by @arcusbuilds in #1275
Validate CD domain ID before generating the IMEX configuration files by @shivamerla and @thc1006 in #1368
Disallow Sharing Configuration for AdminAccess Claims by @visheshtanksale and @arcusbuilds in #1280
Make VFIO device preparation robust by @varunrsekar in #1295
Don't fail on incomplete fabric registration on non-MNNVL node by @varunrsekar in #1289

Known Limitations

Host-managed IMEX supports domain isolation only. All workloads in the same IMEX domain receive channel 0, channel isolation is not supported.
Changing resources.computeDomains.imex.mode or resources.computeDomains.imex.isolation while ComputeDomain workloads are active is not supported. Drain the workloads and remove existing ComputeDomains before changing either value.

Installation

Helm install

helm install dra-driver-nvidia-gpu oci://registry.k8s.io/dra-driver-nvidia/charts/dra-driver-nvidia-gpu \
  --version 0.5.0 \
  --namespace dra-driver-nvidia-gpu \
  --create-namespace \
  --set gpuResourcesEnabledOverride=true

Upgrade

Upgrade from v0.4.1:

helm upgrade -i nvidia-dra-driver-gpu oci://registry.k8s.io/dra-driver-nvidia/charts/dra-driver-nvidia-gpu \
  --version 0.5.0 \
  --namespace nvidia-dra-driver-gpu \
  --set gpuResourcesEnabledOverride=true \
  --set nameOverride=nvidia-dra-driver-gpu

--set nameOverride=nvidia-dra-driver-gpu is only required if it has not already been applied during an earlier upgrade to v0.4.0 or later. Refer to the upgrade documentation for more details.

Container images

registry.k8s.io/dra-driver-nvidia/dra-driver-nvidia-gpu:v0.5.0

New Contributors

@runzhen made their first contribution in #1172
@Phillezi made their first contribution in #1153
@Miaoxiang-philips made their first contribution in #1186
@mesutoezdil made their first contribution in #1178
@arpitjain099 made their first contribution in #1126
@Sreeja1725 made their first contribution in #1218
@vishalanarase made their first contribution in #1237
@enoodle made their first contribution in #1165
@troychiu made their first contribution in #1257
@arcusbuilds made their first contribution in #1275
@killianmuldoon made their first contribution in #1261
@Smallfu666 made their first contribution in #1294
@thc1006 made their first contribution in #1368
@mikemckiernan made their first contribution in #1281

Full Changelog: v0.4.0...v0.5.0