Skip to content

fix: expose AMD GPUs as allocatable Kubernetes resources #586

Description

@castrojo

Problem

The Ghost lab node detects the AMD GPU capability but does not expose any schedulable GPU capacity to workloads.

Observed from the cluster:

Node: ghost
status.capacity.amd.com/gpu: 0
status.allocatable.amd.com/gpu: 0

The node has AMD GPU-related hardware labels, but a pod requesting an AMD GPU cannot be scheduled because the advertised resource count is zero.

Impact

GPU-accelerated workloads, including media-processing and AI inference jobs, must fall back to CPU despite the lab having GPU hardware.

Requested outcome

Configure and validate the AMD GPU device plugin/runtime on Ghost so amd.com/gpu is reported with the actual non-zero count in node capacity and allocatable resources. Document a minimal GPU pod validation command or manifest.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions