To add new GPU worker to existing CPU cluster

Hello All, is there any possibilities to add a new GPU worker node to existing ray cluster with CPU ? The current deployment is on Openshift.

Hi @Balaji_MP! Can you share how you’re setting up your cluster? I haven’t used Openshift before, am curious how you specify what resources you need.

Hello @cade, I used the recommended Helm chart: ray/deploy/charts/ray at master · ray-project/ray · GitHub for deployment and modified the values based on requirement. One key change, instead of latest image, used this one: ray:nightly-py39-cpu.

Hi @Balaji_MP!
The most critical details for setting up pods that use GPUs will come from the Openshift docs.
I think this page should be useful Installing the NVIDIA GPU Operator — NVIDIA Cloud Native Technologies documentation

Once the Openshift GPU setup is complete, you should be able to add a new worker type that uses GPU:

The key fields to set for the worker type would be GPU to indicate the number of GPUs,
possibly nodeSelector to make sure the pod is scheduled on the right node,
and possibly tolerations to tolerate a GPU taint.
nodeSelector and tolerations may or may not be necessary depending on the details of your GPU setup – but basically, you’d set whatever you need get a GPU-utilizing pod running in your environment.

1 Like

Thanks for your quick response. One question, is there any possibilities to provide different image for type2 worker ?

Please ignore the above question, I deployed based on your suggestion and it worked. Thank you :slight_smile:

The above question is a good one!

For simplicity, we don’t currently support setting images per worker type in the Helm chart. However, it’s not too hard to edit the chart to support this functionality, by slightly modifying the RayCluster template.