Track standalone GFD and CUDA validator E2E enablement

[giuliocalzo]

Summary

The legacy nvml-mock-e2e.yaml workflow keeps the standalone GFD and CUDA validator steps disabled by default with if: false because the standalone images are pulled from nvcr.io and may require NGC authentication.

The Go/Ginkgo E2E harness has parity scaffolding for these steps, but the specs should remain skipped by default until the image/auth story is resolved.

Legacy reference

Disabled legacy workflow section pinned to the current default-branch commit:

k8s-test-infra/.github/workflows/nvml-mock-e2e.yaml

Lines 221 to 268 in 4299fcb

	# GFD and CUDA validator standalone images from nvcr.io may require NGC auth.
	# The GPU Operator path (e2e-gpu-operator job) bundles these publicly and works
	# without credentials. These standalone steps are disabled until verified public.
	# See tests/e2e/README.md for local testing instructions.
	- name: Deploy GPU Feature Discovery
	if: false
	run: |
	docker pull nvcr.io/nvidia/gpu-feature-discovery:v0.17.0
	kind load docker-image nvcr.io/nvidia/gpu-feature-discovery:v0.17.0 --name nvml-mock-e2e-${{ matrix.gpu-profile }}
	kubectl apply -f tests/e2e/gfd-mock.yaml
	kubectl -n kube-system wait --for=condition=ready \
	pod -l name=nvidia-gfd-mock --timeout=120s
	- name: Verify GFD node labels
	if: false
	run: |
	NODE=$(kubectl get nodes -o jsonpath='{.items[0].metadata.name}')
	for i in $(seq 1 30); do
	PRODUCT=$(kubectl get node "$NODE" -o jsonpath='{.metadata.labels.nvidia\.com/gpu\.product}' 2>/dev/null || true)
	if [ -n "$PRODUCT" ]; then break; fi
	echo "Attempt $i: waiting for GFD labels..."
	sleep 2
	done
	echo "Checking GFD labels on node $NODE..."
	for LABEL in nvidia.com/gpu.product nvidia.com/gpu.memory nvidia.com/gpu.compute.major; do
	VALUE=$(kubectl get node "$NODE" -o json | jq -r ".metadata.labels[\"$LABEL\"] // empty")
	if [ -z "$VALUE" ]; then
	echo "FAIL: GFD label $LABEL not set"
	kubectl get node "$NODE" -o json | jq '.metadata.labels | to_entries[] | select(.key | startswith("nvidia.com"))' || true
	exit 1
	fi
	echo " $LABEL=$VALUE"
	done
	echo "PASS: GFD labels verified"
	- name: Run CUDA validator job
	if: false
	run: |
	docker pull nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda12.5.0
	kind load docker-image nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda12.5.0 --name nvml-mock-e2e-${{ matrix.gpu-profile }}
	kubectl apply -f tests/e2e/validator-mock.yaml
	kubectl wait --for=condition=complete job/gpu-validator-mock --timeout=120s || {
	echo "Validator job did not complete. Checking status..."
	kubectl describe job/gpu-validator-mock
	kubectl logs -l name=gpu-validator-mock --tail=50 || true
	exit 1
	}
	echo "PASS: CUDA validator job completed successfully"

Disabled legacy steps

• Deploy GPU Feature Discovery
• Verify GFD node labels
• Run CUDA validator job

Acceptance criteria

• Decide how CI should access the standalone nvcr.io images, or replace them with images that are public/fork-safe.
• Enable the Go validator/GFD specs in CI only when credentials or public images are available.
• Keep fork PR behavior safe and deterministic.
• Update tests/e2e/README.md and .github/workflows/nvml-mock-e2e-go.yaml with the final enablement path.