[dhhoang]

Motivation

Enable operator-managed AMD GPU driver installation on Ubuntu 26.04.

Close #631 — Cannot build image for Ubuntu 26.04.

Technical Details

• Ubuntu 26.04 builds now use the resolute repository and install linux-modules-${KERNEL_FULL_VERSION}. Earlier Ubuntu releases continue using linux-modules-extra.

• When DeviceConfig.spec.driver.version is omitted, the operator selects 31.50 for Ubuntu 26.04. Initial driver-version labels now resolve the default separately for each node and return an error if none exists. The Helm instructions explain how to override the chart's explicit driver version to use this default.

• The build also unpacks .ko.zst modules before running depmod, giving Kernel Module Management (KMM) the .ko files it expects for signing. Adding amddrm_exec.ko and amddrm_suballoc_helper.ko to the Ubuntu 26.04 signing list covers all nine AMD modules in the tested releases.

• I also fixed an issue in prepareimage: duplicate flag registrations.

Test Plan

OS	Kernel	GPU	AMD Driver Releases	Coverage
Ubuntu 26.04.1	7.0.0-43-generic	MI300X VF, MI350X VF	31.40.1, 31.50, 31.60	Driver loading, GPU allocation, computation, health metrics, reboot recovery
Ubuntu 26.04.1	7.0.0-42-generic	MI300X VF	31.50	Kernel change from 43 to 42 and back, computation, health metrics

Test Result

All configurations above passed. On both GPU models, matrix multiplication and neural-network inference matched CPU results before and after reboot. MI300X also recovered its GPU workload after automatic driver upgrades from the 31.50 default to 31.60 and back, and after changing kernels in both directions.

Unit tests, go vet, and Helm rendering checks passed. KMM signed all nine modules with driver 31.50 and kernel 7.0.0-42; independent verification accepted each signature and rejected altered module contents. The final container image contained no copy of the private signing key.

Secure Boot enforcement and host virtualization driver installation were not tested. Driver 31.60 logged RAS early init service failure! ret:-95 on MI300X; computation and health checks passed.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.