Skip to content

docs(gpu-operator): State that the DRA Driver is opt-in - #535

Open
mikemckiernan wants to merge 1 commit into
mainfrom
devin/1794335069-dra-opt-in-install
Open

mikemckiernan wants to merge 1 commit into
mainfrom
devin/1794335069-dra-opt-in-install

Conversation

@mikemckiernan

Copy link
Copy Markdown
Member

Summary

The DRA Driver page never said whether the GPU Operator installs the DRA Driver by default, or which value turns it on or off. Readers also had to work out what the gpuCluster.deployCR, clusterPolicy.deployCR, and driver.nvidiaDriverCRD.enabled values do from the Helm commands alone. The "DRA driver operands" wording was jargon, and a link sent readers back to the page they were already on.

dra-intro-install.rst

  • Puts the answer near the top of the page. By default, the GPU Operator uses the NVIDIA Kubernetes Device Plugin and does not install the DRA Driver.
    • Opt in: set gpuCluster.deployCR=true and clusterPolicy.deployCR=false. If the Operator also installs the GPU driver, set driver.nvidiaDriverCRD.enabled=true too.
    • Opt out: omit gpuCluster.deployCR or set it to false.
    • Remove: uninstall the GPU Operator. A Helm upgrade with gpuCluster.deployCR=false doesn't remove the DRA Driver.
  • The Install section now lists what each --set value does. It also says that driver.nvidiaDriverCRD.enabled controls the NVIDIA GPU driver, not the DRA Driver.
  • Adds the dra-driver-install and dra-driver-uninstall anchors for the new cross-references.

getting-started.rst: rewrites the clusterPolicy.deployCR, gpuCluster.deployCR, and driver.nvidiaDriverCRD.enabled chart option rows in reader terms, and links them to the install anchor instead of the page.

I checked each behavior claim against the GPU Operator v26.7.2 chart:

  • Defaults (values.yaml): clusterPolicy.deployCR: true and gpuCluster.deployCR: false.
  • Validation (templates/validations.yaml): the chart fails when both deployCR values are true. It also fails when gpuCluster.deployCR=true, driver.enabled=true, and driver.nvidiaDriverCRD.enabled=false.
  • Retention (templates/gpucluster.yaml): the GPUCluster resource has helm.sh/resource-policy: keep, and a pre-delete cleanup hook deletes it when you uninstall.

Review notes

  • clusterPolicy.deployCR=false wording. The Install list says this value "disables the NVIDIA Kubernetes Device Plugin". The value actually skips the whole ClusterPolicy resource. Should the wording say that instead?
  • Removing only the DRA Driver. Is kubectl delete gpucluster gpu-cluster a supported way to remove the DRA Driver and keep the GPU Operator? The page doesn't mention it.
  • "Uninstall the DRA Driver" link text. The link goes to a section that uninstalls the whole GPU Operator.
  • Overview sentence. "The Helm chart creates this resource when you set gpuCluster.deployCR=true" now repeats the new opening answer. Keep it or cut it?
  • Overlap with gpu-operator: document DRA validator and DCGM Exporter limitation with MIG mode #526. That PR also edits dra-intro-install.rst. A test merge of the two branches had no conflicts.

Validation

  • Docs build: ./repo docs -p gpu-operator in ghcr.io/nvidia/cloud-native-docs:0.5.1 gives the same 41 warnings and errors before and after this change. They include the existing tab-set errors and unreachable external inventories.
  • Vale: the NVIDIA Vale style finds nothing on the changed lines.
  • Pre-commit: not run, because the hooks couldn't install in my environment.
  • Runtime: not tested on a cluster.

I drafted this change with Devin, an AI assistant.

Link to Devin session: https://nvidia-cloud.devinenterprise.com/sessions/c60207a0e98342e98cefb31c56508ff6
Open in Devin Desktop: https://nvidia-cloud.devinenterprise.com/desktop/session/c60207a0e98342e98cefb31c56508ff6?variant=devin
Requested by: @mikemckiernan

Lead the DRA Driver page with the default behavior and the opt-in and
opt-out Helm values. Explain each --set flag in the Install section and
separate the NVIDIA GPU driver from the DRA Driver. Rewrite the
clusterPolicy.deployCR, gpuCluster.deployCR, and
driver.nvidiaDriverCRD.enabled chart option rows in reader terms and link
to the install procedure anchor.

Signed-off-by: Mike McKiernan <mmckiernan@nvidia.com>
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@devin-ai-integration

Copy link
Copy Markdown
Contributor

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

  • Disable automatic comment, CI, and merge conflict monitoring

@github-actions

Copy link
Copy Markdown

Documentation preview

https://nvidia.github.io/cloud-native-docs/review/pr-535

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant