[Feature]: Streamline AMD GPU Operator Deployment into a Single-Step Workflow

[fabianokteixeira]

Suggestion Description

The current AMD GPU Operator deployment process involves multiple manual steps—installing cert-manager via Helm, deploying the GPU Operator via Helm, and then applying the DeviceConfig with kubectl. This multi-step approach increases operational complexity and the chance of misconfiguration, especially since the Operator depends on the DeviceConfig to function properly.

Simplifying the deployment into a single streamlined workflow would significantly reduce overhead and improve the user experience. By consolidating these steps into one cohesive process, the deployment becomes more efficient, less error-prone, and easier to maintain. This would enable customers to get the GPU Operator up and running faster, with minimal manual intervention, ultimately enhancing productivity and satisfaction.

Operating System

No response

GPU

No response

ROCm Component

No response

[yansun1996]

Hi @fabianokteixeira thanks for the suggestion.

As you pointed out the steps to deploy the AMD GPU Operator requires the cert-manager installation, GPU Operator installation and configuration.

1. The dependent Cert-Manager installation, we used to consider put it as a subchart in our Helm Chart. However, by doing that the cert-manager won't be able to be upgraded separately, currently helm doesn't have function to upgrade the subchart independently. Consider that cert-manager already has been adopted widely and the GPU Operator just use its CustomResource and didn't configure it, that's the thinking why we didn't put it into helm chart's subchart.
   Let me know what you think about this point.

2. The GPU Operator's installation and configuration, we have developed the feature to do combine them into one-single-step installation + configuration via Helm and it will be available in v1.3.0, which will be released soon. You don't have to do install + configure separately anymore.

[fabianokteixeira]

Understood. I believe it should be fine having cert-manager as a separate step. The most important thing is to have GPU Operator and Device Config in one single step. Thanks for your response.

[fabianokteixeira]

Just a quick clarification. You mentioned that installation + configuration will be available in v1.3.0. From what I can see, v1.3.0 has been already released, but v1.3.0 still requires a separate step for DeviceConfig/Custom Resource. Is this correct or you meant a different Operator version?

[yansun1996]

@fabianokteixeira we're still closing the v1.3.0 release. Hopefully it will be released soon. In v1.3.0 it will support installation + configuration in one step.

[fabianokteixeira]

AMD Documentation shows v1.3.0 as released (upper left corner) - https://instinct.docs.amd.com/projects/gpu-operator/en/latest/index.html

[yansun1996]

@fabianokteixeira yep, let me clarify:

1. The docs are already out but the formal release, we're still preparing the rest of the stuff like more docs update, formal announcement on Release page and so on. You can see the release branch is still not created yet and code opensource for v1.3.0 is still in-progress, etc.

2. You can go ahead try the latest helm chart of v1.3.0, In the Kubernetes (Helm) section we described the instruction of configuring the default DeviceConfig while running helm install. The full default DeviceConfig values.yaml is in

   gpu-operator/helm-charts-k8s/values.yaml

   Line 34 in 17c8db6

   deviceConfig:

   .

[fabianokteixeira]

Got it. Thanks!