feat(gpu): add NVIDIA GRID v20 driver support for RTX PRO 6000 BSE v6 SKUs

[ganeshkumarashok]

What

Adds NVIDIA GRID v20 (595.x) driver support, selecting the new aks-gpu-grid-v20 container image for RTX PRO 6000 Blackwell Server Edition v6 SKUs:

• Standard_NC128ds_xl_RTXPRO6000BSE_v6
• Standard_NC256ds_xl_RTXPRO6000BSE_v6
• Standard_NC320ds_xl_RTXPRO6000BSE_v6

All existing GRID SKUs keep using aks-gpu-grid (570.x); the CUDA path is untouched.

Changes

• parts/common/components.json — add aks-gpu-grid-v20 GPUContainerImages entry.
• pkg/agent/datamodel/gpu_components.go — parse it into NvidiaGridV20DriverVersion / AKSGPUGridV20VersionSuffix; refactor LoadConfig to match on the exact repo name (fixes a latent substring collision: aks-gpu-grid-v20 contains aks-gpu-grid); add RTXPro6000GPUDriverSizes.
• pkg/agent/baker.go — add useGridV20Drivers(); branch GetGPUDriverVersion / GetAKSGPUImageSHA / GetGPUDriverType on it (checked before grid); driver type string "grid-v20".
• .github/renovate.json — add aks/aks-gpu-grid-v20 package rule.
• Unit tests for the new selection paths.

Design notes

On Ubuntu the driver image repo is built as mcr.microsoft.com/aks/aks-gpu-${GPU_DRIVER_TYPE} (cse_helpers.sh), so setting the driver type to grid-v20 resolves the new repo automatically.

Scope is Ubuntu-only by design. RTX PRO 6000 BSE v6 runs on Ubuntu GPU nodes. The non-Ubuntu install paths (Mariner RPM / ACL sysext) do not use the container image and have no v20 packages, so those CSE checks are deliberately left unchanged.

The new image comes from aks-gpu PR #158 (merged).

⚠️ Do not merge yet

aks-gpu-grid-v20 is not yet published to MCR (onboarding tracked separately). The version tag suffix in components.json (595.58.03-20260101000000) is a placeholder and must be replaced with the real published tag before merge. Until then nodes would attempt to pull a nonexistent tag.

make generate produces no testdata/manifest diff (no existing scenario uses these SKUs), so the placeholder does not leak into generated snapshots.

Testing

• go build ./pkg/agent/...
• go test ./pkg/agent ./pkg/agent/datamodel — pass
• make validate-components — pass

[Copilot]

Pull request overview

Adds a new NVIDIA GRID v20 (595.x) driver selection path for RTX PRO 6000 Blackwell Server Edition v6 NC SKUs by introducing a new GPU driver container image (aks-gpu-grid-v20) and ensuring config parsing doesn’t confuse it with the existing aks-gpu-grid image.

Changes:

• Add aks-gpu-grid-v20 to GPUContainerImages and parse it into new datamodel globals (version + suffix), using exact repo-name matching to avoid substring collisions.
• Add SKU-based routing so RTX PRO 6000 BSE v6 sizes use GRID v20 for GetGPUDriverVersion, GetAKSGPUImageSHA, and GetGPUDriverType (new type: grid-v20).
• Extend Renovate rules and unit tests to cover the new config fields and selection paths.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
parts/common/components.json	Adds new GPU container image entry for aks-gpu-grid-v20 (tag currently a placeholder per PR description).
pkg/agent/datamodel/gpu_components.go	Adds v20 config globals and exact repo-name parsing; introduces RTX PRO 6000 BSE v6 SKU map.
pkg/agent/datamodel/gpu_components_test.go	Validates v20 config values are populated and correctly formatted.
pkg/agent/baker.go	Routes RTX PRO 6000 BSE v6 SKUs to GRID v20 driver/version/type before standard GRID selection.
pkg/agent/baker_test.go	Adds unit tests for GRID v20 selection in version/type/image suffix.
.github/renovate.json	Adds Renovate package rule for aks/aks-gpu-grid-v20.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.