OMNI can provision GPU instances with MIG partitioning and optional DRA (Dynamic Resource Allocation) support. When you create an edge node, you can request GPUs with a MIG configuration and DRA enabled — see Create an edge. With DRA enabled, GPU resources are advertised through Kubernetes ResourceSlices and allocated via ResourceClaims, so you can create a ResourceClaim, assign it to a workload definition, and consume a single MIG slice.

📘

Note

The OMNI API applies one MIG configuration to the whole instance: the same partition sizes are applied to every GPU attached to the VM. You cannot define a non-homogeneous MIG layout — different profiles per GPU, or a mix of profiles on a single GPU — through the API alone.

This guide shows how to achieve custom, mixed MIG layouts on OMNI edge nodes by customizing the nvidia-gpu-operator add-on installed on the edge cluster.

Before you begin

• OMNI enabled on your cluster. See Getting started with OMNI.
• At least one edge location with the nvidia-gpu-operator add-on enabled.
• An edge instance with MIG-capable GPUs — for example, a GCP a2-highgpu-2g machine type with 2× A100-40GB GPUs. MIG is supported on NVIDIA GPUs from the Ampere architecture onward; see Multi-Instance GPU (MIG) and the NVIDIA MIG documentation for supported GPUs and partition sizes.

How MIG is managed on edge nodes

MIG management on an edge node is performed by the MIG Manager, a component of the NVIDIA GPU Operator installed through the nvidia-gpu-operator add-on. The MIG Manager:

1. Watches the nvidia.com/mig.config node label.
2. Looks up the label value in a mig-parted configuration — a ConfigMap on the edge cluster — to determine the MIG layout to apply.
3. Applies the layout using mig-parted.

By default, the add-on uses the default-mig-parted-config ConfigMap that ships with the GPU Operator. You can point the add-on at your own ConfigMap using the migManager.config.name Helm value. When a configuration mixes different MIG profiles (for example, 1g.5gb and 2g.10gb on the same GPU), the GPU Operator must run with the mixed MIG strategy (mig.strategy=mixed); the single strategy applies one uniform profile to all GPUs.

For details, see the NVIDIA GPU Operator documentation.

Step 1: Configure the gpu-operator add-on

curl --request PATCH \
  'https://api.cast.ai/omni-provisioner/v1beta/organizations/<ORG_ID>/clusters/<CLUSTER_ID>/edge-locations/<EDGE_LOCATION_ID>' \
  --header 'Content-Type: application/json' \
  --header 'Accept: application/json' \
  --header 'X-API-Key: <API_KEY>' \
  --data '{
    "edgeClusterSpec": {
      "addons": [
        {
          "name": "nvidia-gpu-operator",
          "values": {
            "mig": {
              "strategy": "mixed"
            },
            "migManager": {
              "config": {
                "name": "custom-mig-parted-config"
              }
            }
          }
        }
      ]
    }
  }'

The console and the API write the same edge location add-on values (edgeClusterSpec.addons) — the values are deep-merged over the add-on's default Helm values.

📘

Note

The addons list defines the complete set of optional add-ons installed on the edge cluster — when the list is set, only the listed add-ons are installed. If the edge location already has optional add-ons enabled (for example, nvidia-dra), include them in the list alongside nvidia-gpu-operator, otherwise the update removes them.

Step 2: Apply the custom MIG ConfigMap

Create the ConfigMap the add-on now references. The ConfigMap must contain a config.yaml key. In the examples below, 0x20B010DE is the device ID of the A100-SXM4-40GB GPU type.

Single GPU with mixed MIG profiles

apiVersion: v1
kind: ConfigMap
metadata:
  name: custom-mig-parted-config
  namespace: gpu-operator
data:
  config.yaml: |-
    version: v1
    mig-configs:
      custom-mixed:
        - device-filter: ["0x20B010DE"]
          devices: all
          mig-enabled: true
          mig-devices:
            "1g.5gb": 4
            "2g.10gb": 1

Two GPUs with different MIG profiles

apiVersion: v1
kind: ConfigMap
metadata:
  name: custom-mig-parted-config
  namespace: gpu-operator
data:
  config.yaml: |-
    version: v1
    mig-configs:
      custom-mixed:
        - device-filter: ["0x20B010DE"]
          devices: [0]
          mig-enabled: true
          mig-devices:
            "1g.5gb": 7
        - device-filter: ["0x20B010DE"]
          devices: [1]
          mig-enabled: true
          mig-devices:
            "2g.10gb": 3

Apply the manifest:

kubectl apply -f custom-mig-config.yaml

Step 3: Trigger the MIG reconfiguration

Label the edge node with the custom profile:

kubectl label nodes <NODE_NAME> nvidia.com/mig.config=custom-mixed --overwrite

The MIG Manager picks up the change and applies the layout. To watch it work:

kubectl logs -n gpu-operator -l app=nvidia-mig-manager -c nvidia-mig-manager -f

When the reconfiguration completes, confirm the node reports a successful state:

kubectl get node <NODE_NAME> -o jsonpath="{.metadata.labels['nvidia.com/mig.config.state']}"

The value should be success. To inspect the resulting MIG devices:

kubectl exec -it -n gpu-operator ds/nvidia-driver-daemonset -- nvidia-smi -L

⚠️

Warning

MIG Manager terminates workloads that use the GPUs before applying a new layout, and on some cloud providers the node must reboot to enable MIG mode. Reconfiguration is per node — repeat the labeling step on every edge node that should use the custom layout.

Related documentation

• Multi-Instance GPU (MIG)
• Manual edge provisioning
• Custom edge locations
• Create an edge location
• Create an edge
• Update edge location