Manual edge provisioning

By default, edge nodes are created automatically: the Autoscaler provisions them whenever a node template with edge locations enabled needs capacity. Manual edge provisioning lets you create a single edge node directly, with an exact instance type, GPU sharing configuration, and edge configuration of your choosing — outside of the Autoscaler's normal decision-making.

When to use manual provisioning

Manual provisioning is useful when you want a specific outcome rather than the most cost-effective one, for example:

• Validating a new edge location before adding it to a node template, without needing a pending pod to trigger provisioning.
• Reserving a specific instance type or GPU (e.g. a specific MIG partition layout) for a known workload.
• Testing GPU sharing configurations (time-slicing or MIG) on a given instance type before rolling them out via node templates.
• One-off or ad hoc capacity that shouldn't be subject to Autoscaler bin-packing or downscaling decisions.

📘

Note

Manually provisioned edge nodes still appear as standard edge nodes in your cluster (see How edge nodes appear in your cluster) and are subject to the same Evictor behavior as Autoscaler-provisioned ones.

Prerequisites

• A Phase 2 cluster with OMNI enabled.
• At least one edge location that has completed onboarding (status Incomplete setup or In use — not Pending or Failed).

Provision an edge node using the console

1. In the Cast AI console, navigate to Autoscaler → Edge locations.

2. Select the edge location you want to provision the node in, and open Provision an edge node.

3. Configure the node:

   • Edge location — the edge location to provision into. Pre-filled if you started from a specific edge location.
   • Instance type — the cloud instance type for the node.
   • Availability zone — filtered to zones available for the selected instance type within this edge location. Select the instance type first.
   • Edge configuration — the edge configuration to use for this node (networking, defaults, etc.). If none is selected, the edge location's default configuration is used.
   • Resource offering — On-Demand or Spot.

4. (Optional, GPU instance types only) Configure GPU sharing via the standard NVIDIA device plugin:

   • GPU time-slicing — set Replicas to time-slice a GPU into that many virtual replicas for shared access.
   • Multi-instance GPU (MIG) — select Partition sizes to split the GPU into isolated slices with dedicated memory. Available for A100/H100 instances only.

   These two settings can be combined: on a MIG-capable instance type, you can partition the GPU into MIG slices and additionally time-slice each MIG partition, exposing more schedulable GPU replicas from the same physical device. This is standard NVIDIA device plugin behavior and is separate from Dynamic Resource Allocation (DRA).

5. (Optional) Configure Kubernetes orchestration:

   • Taint nodes — applies a taint so pods without a matching toleration won't be scheduled on this node.
   • Use custom labels — apply additional Kubernetes labels to the node.

6. (Optional) Under Edge configuration overrides → Storage, set a Boot disk size in GiB. Leave blank to inherit the size from the linked edge configuration.

7. Click Provision to submit the request.

The node is created asynchronously; it appears in kubectl get nodes once provisioning completes (see Edge node provisioning time).

Provision an edge node using the API

You can provision an edge node programmatically with the Create an edge endpoint.

POST https://api.cast.ai/omni-provisioner/v2beta/organizations/{organizationId}/clusters/{clusterId}/edge-locations/{edgeLocationId}/edges

Path parameters

Body parameters

Example request

curl -X POST \
  "https://api.cast.ai/omni-provisioner/v2beta/organizations/${ORGANIZATION_ID}/clusters/${CLUSTER_ID}/edge-locations/${EDGE_LOCATION_ID}/edges" \
  -H "X-API-Key: ${CASTAI_API_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "gpu-edge-node-01",
    "instanceType": "a2-highgpu-1g",
    "schedulingType": "ON_DEMAND",
    "zone": "us-central1-a",
    "nodeArchitecture": "X86_64",
    "bootDiskGib": 200,
    "kubernetesLabels": {
      "workload-type": "gpu-inference"
    }
  }'

A successful request returns 200 OK with the created edge object. See the API reference for the full response schema and error format.

Verifying the node

kubectl get nodes

A manually provisioned node shows up the same way as an Autoscaler-provisioned edge node — as a virtual node with role agent, tainted virtual-node.omni.cast.ai/not-allowed=true:NoExecute by default, and labeled with omni.cast.ai/edge-location-name, omni.cast.ai/edge-id, and omni.cast.ai/csp. See How edge nodes appear in your cluster for the full list of labels and taints.

To schedule workloads on it, follow the same steps as for any edge node — see Scheduling workloads on edge nodes.

Troubleshooting

FAQ

Does a manually provisioned edge node count toward node template constraints?
Yes. It's a regular node in the cluster and is counted like any other node when Autoscaler and Evictor evaluate constraints and bin-packing.

Can I combine GPU time-slicing and MIG on the same node?
Yes, on MIG-capable instance types. MIG partitions the physical GPU into isolated slices, and time-slicing can then be applied on top of each MIG partition to expose additional virtual replicas.

Will the Autoscaler remove a manually provisioned node?
The Evictor can evict workloads off it like any edge node, but it doesn't delete manually provisioned nodes outright — the same downscaling behavior that applies to other nodes in your cluster applies here.