[DinoChou-amd]

Summary

SSH-preinstalled client nodes previously installed AMD DKMS unconditionally, including Ryzen APUs using Ubuntu OEM kernels. This change uses Ubuntu's inbox amdgpu driver for Ryzen APU and Radeon dGPU clients, with ROCm user-space supplied by workload containers.

• Remove AMD installer, DKMS, kernel headers, and build-helper installation.
• Validate the inbox module, loaded driver, and KFD/render devices before provisioning.
• Reject existing AMD/ROCK DKMS driver and firmware packages with migration guidance.
• Ensure Ubuntu firmware is installed and retain the existing GPU access configuration.
• Document kernel prerequisites, migration steps, and reboot requirements.

Existing DKMS installations are not automatically removed. Nodes must have a working Ubuntu GPU stack before running the playbook.

Validation

• Four new preflight tests and ten existing GPU access contract tests passed.
• Ansible syntax check passed.
• Applied HX370 with kernel 6.14.0-1020-oem.
• Confirmed inbox amdgpu, KFD/render device permissions, and active k3s-agent.
• ROCm 7.13 container compute passed on Radeon 890M (gfx1150):
  rocminfo, GPU matrix multiplication against a CPU reference, and autograd.
• Container tests ran using PyTorch 2.9.1+rocm7.13.0 and GPU allocation through the Kubernetes device plugin.

Radeon dGPU hardware and migration from an existing DKMS installation have not yet been tested.