feat(gpu): use aks-gpu-cuda-lts (R580 LTS) for the managed CUDA driver

[ganeshkumarashok]

What this PR does / why we need it:

Switches the managed CUDA GPU driver image from aks-gpu-cuda to the R580 LTS variant aks-gpu-cuda-lts (580.159.04-20260629214430).

Why: Enabling the CUDA driver prebake (bake #8786, flag #8803) requires an aks-gpu image that supports the build-only action (aks-gpu #162). The only build-only-capable aks-gpu-cuda builds are on the R595 line — and NVIDIA R595 drops Volta/V100 (R590+ deprecation). Fleet telemetry (AKSprod AgentPoolSnapshot) shows that a significant number of managed-GPU nodes, spanning many customer subscriptions, still run on V100 (NC*_v3, ND40rs_v2) — and nearly all of them rely on the managed driver (only a small fraction opt out via --gpu-driver None). So bumping to R595 (the approach in #8810) would break them.

aks-gpu already publishes exactly what we need: aks-gpu-cuda-lts = the NVIDIA R580 Long Term Support branch (driver_config.yml: "supported through Aug 2028"), which keeps V100 and whose post-#162 builds already have build-only. It also covers every other managed CUDA SKU (T4/A100/H100/H200). This keeps V100 working and unblocks the prebake — with no aks-gpu change. It's a move within the 580 line (580.126.09 → 580.159.04), not a driver-branch jump.

Wiring (traced end-to-end):

Area	Change
parts/common/components.json	aks-gpu-cuda → aks-gpu-cuda-lts (repo, renovateTag, version)
pkg/agent/datamodel/gpu_components.go	LoadConfig case → aks-gpu-cuda-lts (sets CUDA driver version/suffix for VHD build and runtime)
pkg/agent/baker.go GetGPUDriverType	modern CUDA SKUs → "cuda-lts" (selects aks-gpu-cuda-lts via cse_helpers.sh aks-gpu-${type}); legacy NCv1/K80 stays "cuda" with its pinned R470 driver
parts/.../cse_config.sh logGPUDriverPrebakeReadiness	map driver-type → aks-gpu marker driver_kind (cuda-lts→cuda, grid-v20→grid) so a CUDA-prebaked marker matches a cuda-lts node
vhdbuilder/packer/install-dependencies.sh	VHD-build prebake/caching image selection → aks-gpu-cuda-lts
ACL / Mariner cse_install_*.sh	comment accuracy; their grid-vs-non-grid sysext logic already handles cuda-lts

Scope note: ACL/AzureLinux install GPU drivers from OS sysext images, not the aks-gpu container, so this is effectively Ubuntu-scoped.

Validation: go test ✅ · make generate-testdata (no drift) ✅ · shellcheck (changed scripts clean) ✅ · shellspec 751/0 (incl. new cuda-lts→cuda match test) ✅ · make validate-components ✅

Relationship to other PRs:

• Supersedes feat(gpu): bump aks-gpu-cuda to 595.71.05 (build-only-capable; prereq for CUDA prebake) #8810 (bumped aks-gpu-cuda → R595; would have dropped V100). I'll close it.
• Complements feat(gpu): enable CUDA driver prebake on shared Ubuntu gen2 VHD builds #8803 (prebake enablement flag) — with this image, the flag's build-only prebake works while V100 stays supported.

Which issue(s) this PR fixes: Fixes #

Requirements

• Conventional-commit title · DCO signed-off · GPG-verified · branch on Azure/AgentBaker
• make generate-testdata clean; go test, shellspec, validate-components pass

[Copilot]

Pull request overview

This PR updates the managed NVIDIA CUDA driver container image selection so GPU nodes use the R580 LTS–based aks-gpu-cuda-lts image (to retain Volta/V100 support while enabling build-only prebake), and wires the new driver type through AgentBaker’s Go selection logic, VHD build caching, and CSE observability/tests.

Changes:

• Switch GPU container image metadata in parts/common/components.json from aks-gpu-cuda to aks-gpu-cuda-lts (including renovate tag + version).
• Update AgentBaker Go GPU component loading and driver-type selection to use cuda-lts for modern CUDA SKUs (while keeping legacy NCv1 on cuda).
• Update VHD build pre-pull logic and add ShellSpec coverage for the prebake marker kind mapping.

Package Update Analysis: aks-gpu-cuda-lts

Version change: 580.126.09-20260126030251 → 580.159.04-20260629214430 (minor update within the 580 driver branch)
OS variants affected: Ubuntu VHD build/runtime managed-driver path (container-based install)
OS variants NOT updated: AzureLinux/Mariner/ACL (they install via sysext/RPM paths, not the aks-gpu CUDA container)

Changes between versions: Upstream, version-specific release-note diffs for these exact driver point releases were not found in a reliably citable form during this review. Manual validation (e2e GPU scenarios + targeted V100/NCv3 coverage) is recommended before merge.

Overall Risk: Medium
Justification: Although the driver stays on the R580 line (not a major branch jump), the PR also changes the image repository and introduces a new driver-type string (cuda-lts) that must remain compatible with legacy selection paths (notably NCv1/K80).

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
parts/common/components.json	Switch CUDA GPU image metadata to aks-gpu-cuda-lts and bump version.
pkg/agent/datamodel/gpu_components.go	Load CUDA driver version/suffix from aks-gpu-cuda-lts repo entry.
pkg/agent/datamodel/gpu_components_test.go	Update repo parsing test to expect aks-gpu-cuda-lts.
pkg/agent/baker.go	Return cuda-lts for modern CUDA SKUs; keep legacy NCv1 on cuda.
pkg/agent/baker_test.go	Update expectations for cuda-lts and add explicit NCv1 legacy coverage.
vhdbuilder/packer/install-dependencies.sh	Pre-pull aks-gpu-cuda-lts during Ubuntu VHD build and update related error text.
parts/linux/cloud-init/artifacts/cse_config.sh	Map driver-type to marker driver_kind for prebake readiness logging (cuda-lts→cuda).
spec/parts/linux/cloud-init/artifacts/cse_config_spec.sh	Add ShellSpec test for cuda-lts marker-kind matching behavior.
parts/linux/cloud-init/artifacts/mariner/cse_install_mariner.sh	Update comments describing cuda-lts vs legacy cuda behavior (no logic change).
parts/linux/cloud-init/artifacts/acl/cse_install_acl.sh	Update comments describing cuda-lts vs legacy cuda behavior (no logic change).