[Feature]: Make Private Repository Optional for Out-of-Tree Deployments

[fabianokteixeira]

Suggestion Description

Problem Statement

The current AMD GPU Operator implementation requires a private image repository for Out-of-Tree driver deployments. This creates unnecessary overhead for internet-connected environments, as users must:

• Manually configure and maintain a private registry
• Manage registry credentials

Proposed Solution

Official AMD Public Repository for Out-of-Tree Deployments:

• Host pre-built AMD driver images in a public registry (e.g., docker.io/amd/rocm-driver)
• Automatically append version tags (e.g., 6.4.1) using existing Operator logic

Optional Custom Repositories:

• Example of new configuration in deviceconfig.yaml

spec:
driver:
image: "docker.io/amd/rocm-driver:version" # Default value latest
usePublicRepo: true # New flag

• Retain private repo support for air-gapped environments

Benefits

• Reduced configuration complexity
• Eliminates hours spent configuring private registries in connected environments.
• Flexibility to all customers, internet-connected and air-gapped.

Operating System

No response

GPU

No response

ROCm Component

No response

[yansun1996]

Hi @fabianokteixeira thanks for your suggestion. We will evaluate the gaps between what we have and your proposal.

[sscargal]

I have the same issue and requirement.