feat(gpu): consume the VHD-prebuilt CUDA kernel module (skip in-CSE recompile)

[ganeshkumarashok]

What

Splits the consume / skip-build half of #8661 into its own PR (#8661 stays open for reference). The bake + teardown half is #8786 and must land first to produce the marker.

Changes

• parts/linux/cloud-init/artifacts/cse_config.sh — when the aks-gpu prebake marker is present and its driver_kind matches this node, configGPUDrivers asks the aks-gpu container for install-skip-build (device-dependent steps only), skipping the ~80–150s in-CSE DKMS recompile. aks-gpu independently re-validates the marker (kernel + driver_version + driver_kind) and falls back to a full build on any mismatch.

Safety

• No marker (non-prebaked VHD) → install, i.e. today's behavior. Defense-in-depth: gated by marker presence and driver_kind match, with aks-gpu's own re-validation + full-build fallback underneath.

Validation

• shellspec 736/0, make generate-testdata clean, shellcheck clean.
• e2e on the real [TEST All VHDs] pipeline (bake + consume together): in progress.

Reference (full original, kept open): #8661.

[Copilot]

Pull request overview

This PR updates the Linux CSE GPU driver install flow to consume a VHD-prebaked NVIDIA CUDA DKMS module when a prebake marker is present, by invoking the aks-gpu container with install-skip-build (instead of install) to avoid the expensive in-CSE recompile. It fits into the AgentBaker provisioning scripts under parts/linux/cloud-init/artifacts/ and adds ShellSpec coverage to validate the marker-driven action selection (including the driver_kind guard).

Changes:

• Extend installGPUDriverImage to accept an install action (install vs install-skip-build) and pass the selected action from configGPUDrivers on Ubuntu.
• Add marker parsing and a driver_kind match guard to ensure skip-build only happens when the prebaked driver kind matches the node’s NVIDIA_GPU_DRIVER_TYPE.
• Add ShellSpec tests validating action selection: no marker → install, matching marker → install-skip-build, mismatched kind → install.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
parts/linux/cloud-init/artifacts/cse_config.sh	Select install vs install-skip-build based on prebake marker presence and driver_kind match, and plumb the action into the aks-gpu container invocation.
spec/parts/linux/cloud-init/artifacts/cse_config_spec.sh	Add ShellSpec coverage for marker-driven action selection, including the CUDA-vs-GRID guard behavior.

[ganeshkumarashok]

Good call — renamed this block to Describe 'configGPUDrivers prebake marker action selection' so it's distinct from the existing timing-event configGPUDrivers suite, keeping shellspec output unambiguous (commit 5b72619).

---

Response generated by GitHub Copilot.