Allocate XGMI-linked GPUs

Requires a HAMi release newer than v2.10.0 (Project-HAMi/HAMi#3191).

To run a multi-GPU pod only on GPUs that are all linked to each other over XGMI, set the annotation amd.com/link-bind to "true". amd-device-plugin publishes the XGMI peers of each GPU as xgmiPeers in the node registration. The scheduler needs every requested GPU to list every other requested GPU as a peer, and keeps the pod Pending when the node has no such group. A request for a single GPU is not affected.

apiVersion: v1
kind: Pod
metadata:
  name: amd-gpu-link-bind
  annotations:
    amd.com/link-bind: "true" # every requested GPU must be XGMI-linked to all the others
spec:
  containers:
    - name: rocm
      image: rocm/dev-ubuntu-24.04:7.2.4
      command: ["bash", "-c", "sleep infinity"]
      securityContext:
        # runs as the image's user (root) on purpose: GPU device access needs the node's video and render group ids, which differ per node. For non-root add runAsUser and supplementalGroups for those groups.
        allowPrivilegeEscalation: false
      resources:
        limits:
          amd.com/gpu: 4

Without amd.com/link-bind, the topology-aware policy prefers linked GPUs but still places a pod when no fully linked group exists. See Choose a GPU scheduling policy.