Choose a GPU scheduling policy

Set the annotation hami.io/gpu-scheduler-policy on a pod to choose how the scheduler picks GPUs on a node. The default comes from scheduler.defaultSchedulerPolicy.gpuSchedulerPolicy in the HAMi chart (spread). Several policies can be combined in one comma-separated value, for example binpack,mutex.

Policy	Effect on AMD GPUs
binpack	Fills one GPU before using the next.
spread	Spreads pods over different GPUs.
mutex	Uses only GPUs that no other pod holds. A pod that finds none stays Pending (ExclusiveDeviceAllocateConflict). Other pods can still share a GPU that a mutex pod holds.
topology-aware	For a request of two or more GPUs, picks the group with the most XGMI links, and uses a shared NUMA node as the tie-break. Requires a HAMi release newer than v2.10.0 (Project-HAMi/HAMi#3191).

apiVersion: v1
kind: Pod
metadata:
  name: amd-gpu-topology-aware
  annotations:
    hami.io/gpu-scheduler-policy: "topology-aware"
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: 2

To require a fully linked group instead of preferring one, use amd.com/link-bind. To require one NUMA node, use amd.com/numa-bind.