Issue evidence · Launch · NVIDIA/gpu-operator

default-kind management.nvidia.com/gpu incompatible with non-cdi strategies, making GPU pods unresolvable

GitHub issue: https://github.com/NVIDIA/gpu-operator/issues/3077

Not run

With cdi.enabled and the volume-mounts device-list strategy, bare device names are qualified with a kind that cannot contain them, so GPU pods fail to start.

What Badgr ran

Nothing was run: it needs a Kubernetes cluster with the GPU Operator installed and a GPU node, and no cluster was available (kubectl had no context).

What came back

  • The report includes the pod manifest and the kubelet event for the failure.

Blocked by infrastructure, as the plan anticipated.

Checked 2026-10-14 in Badgr’s local development environment.

← All issue evidence