Deploy and manage a Vultr GPU Fabric Cluster with RoCE or InfiniBand fabric networking using the API.
A GPU Fabric Cluster connects GPU instances via an automated, high-performance fabric network (RoCE or InfiniBand) for AI/ML training, distributed inference, and HPC workloads. Unlike a Compute Cluster or Bare Metal Cluster, a GPU Fabric Cluster does not automatically provision instances to match desired_pool_count. You deploy each instance yourself and attach it to the cluster.
This guide explains how to deploy and manage a GPU Fabric Cluster using the Vultr API.
Raising desired_pool_count on a GPU Fabric Cluster does not provision any instances. The field still tracks your intended pool size, but reaching it is a manual process on this cluster type.
Send a POST request to the Create Cluster endpoint with gpu_fabric set to true to create the cluster. Replace REGION with the region identifier, PLAN with a GPU-capable bare metal plan, CLUSTER_LABEL with a descriptive name, OS_ID with the operating system ID, and HOSTNAME with a hostname for the cluster's instances. See How to Create a Cluster for the full set of create options.
A successful request returns an HTTP 200 OK response. The cluster's state reads empty until you attach an instance, even if desired_pool_count is 1 or higher.
Send a POST request to the Create Bare Metal Instance endpoint to deploy an instance to attach. Use the same plan and region as the cluster. Replace OS_ID with the operating system ID and INSTANCE_LABEL with a descriptive label.
Wait for the instance's status to read active before attaching it. Note the instance's id.
Send a POST request to the Add Single Instance endpoint to attach the instance to the cluster as a compute node. Replace {cluster-id} with the cluster's ID and {instance-id} with the instance's ID. To attach it as the cluster's head node instead, see Attach a Head Node below.
A successful request returns an HTTP 202 Accepted response. The fabric interconnect, PKey partition, and VLAN assignment configure automatically on attach. No manual network setup is required.
Repeat steps 2 and 3 for each additional instance you want in the cluster, since the cluster does not provision them for you.
A GPU Fabric Cluster can designate one attached instance as its head node, separate from the compute node pool. This action is only available on GPU Fabric Clusters.
Deploy an instance as in step 2 above, using the same plan and region as the cluster.
Send a POST request to the Add Single Instance endpoint with attach_head_node as the action, to designate the instance as the cluster's head node. Replace {cluster-id} with the cluster's ID and {instance-id} with the instance's ID.
A successful request returns an HTTP 202 Accepted response.
A cluster can only have one head node. Attaching a second returns a 400 error reporting that the cluster already has one. Sending this action to a non-GPU cluster also returns a 400 error, since only GPU clusters support head nodes.
Send a GET request to the Get Cluster Metrics endpoint to retrieve GPU and fabric metrics for each attached instance, including per-GPU temperature and power usage, and per-NIC network utilization and error counts. See How to Retrieve Cluster Metrics for the request format.
0 Comments
Be the first to comment and share your perspective with the community.