Skip to content

Inference Setup

Inference Jobs are an integral part of Run:AI and do not require setting up per se. However, Running multiple production-grade processes on a single GPU is best performed with an NVIDIA technology called Multi-Process Service or MPS

By default, MPS is not enabled on GPU nodes.

Enable MPS

To enable the MPS server on all nodes, you must edit the cluster installation values file:

  • When installing the Run:AI cluster, edit the values file.
  • On an existing installation, use the upgrade cluster instructions to modify the values file.

Use:

runai-operator:
  config:
    mps-server:
      enabled: true

Wait for the MPS server to start running:

 kubectl get pods -n runai

When the MPS server pod has started to run, restart the nvidia-device-plugin pods:

kubectl delete pods -n runai --selector=name=nvidia-device-plugin-ds

To enable the MPS server on selected nodes, please contact Run:AI customer support.

Verify MPS is Enabled

Run:

kubectl get pods -n runai --selector=app=runai-mps-server -o wide
  • Verify that all mps-server pods are in Running state.

  • Submit a workload with MPS enabled using the --mps flag. Then run:

runai list
  • Identify the node on which the workload is running. In the get pods command above find the pod running on the same node and then run:
kubectl logs -n runai runai-mps-server-<name> -f

You should see activity in the log


Last update: November 8, 2021