What the docs say
The v26.7.0 Known Issues section in gpu-operator/release-notes.rst lists a single
issue (DRA passthrough allocation), and the GPUCluster Limitations in
gpu-operator/dra-intro-install.rst cover migration and GKE overrides. Neither
covers ownership of the ComputeDomain CRD when the DRA driver is installed from
its own Helm chart alongside the GPU Operator chart.
What appears to be true
Both charts ship the ComputeDomain CRD, so ownership of it conflicts when both
are installed. A plain helm install does not surface this, because Helm does not
record release ownership for CRDs installed from crds/ — but a GitOps installer
that does track ownership, such as Argo CD, fails on it. This is a common
deployment shape on GB200, where multi-node NVLink is exactly what ComputeDomains
are for, so the affected population overlaps heavily with the users of the
feature.
Suggested fix
A Known Issue entry under v26.7.0 with the workaround: skip CRDs in the GPU
Operator chart and apply the ClusterPolicy and NVIDIADriver CRDs separately.
Worth stating explicitly whether the same conflict applies when the Operator is
installed via OLM and the DRA driver via Helm.
What the docs say
The v26.7.0 Known Issues section in
gpu-operator/release-notes.rstlists a singleissue (DRA passthrough allocation), and the GPUCluster Limitations in
gpu-operator/dra-intro-install.rstcover migration and GKE overrides. Neithercovers ownership of the
ComputeDomainCRD when the DRA driver is installed fromits own Helm chart alongside the GPU Operator chart.
What appears to be true
Both charts ship the
ComputeDomainCRD, so ownership of it conflicts when bothare installed. A plain
helm installdoes not surface this, because Helm does notrecord release ownership for CRDs installed from
crds/— but a GitOps installerthat does track ownership, such as Argo CD, fails on it. This is a common
deployment shape on GB200, where multi-node NVLink is exactly what ComputeDomains
are for, so the affected population overlaps heavily with the users of the
feature.
Suggested fix
A Known Issue entry under v26.7.0 with the workaround: skip CRDs in the GPU
Operator chart and apply the
ClusterPolicyandNVIDIADriverCRDs separately.Worth stating explicitly whether the same conflict applies when the Operator is
installed via OLM and the DRA driver via Helm.