🐛 Fix nightly E2E: update DEFAULT_MODEL_ID and add WVA_METRICS_SECURE

[clubanderson]

Summary

Two fixes for the nightly E2E tests on OpenShift:

1. Update DEFAULT_MODEL_ID from Qwen/Qwen3-32B to Qwen/Qwen3-0.6B

   • The llm-d repo changed its default model, but install.sh still had the old value
   • The yq model replacement silently failed — vLLM served Qwen/Qwen3-0.6B while WVA queried for unsloth/Meta-Llama-3.1-8B metrics
   • Result: MetricsAvailable=False, DesiredOptimizedAlloc empty, Scale-to-Zero tests fail

2. Add WVA_METRICS_SECURE env var (default: true) wired to wva.metrics.secure Helm value

   • When set to false, disables bearer token auth on the WVA /metrics endpoint
   • Needed for OpenShift user-workload-monitoring where the controller-manager SA token auth fails with "Token does not match server's copy"
   • Without this, Prometheus can't scrape wva_desired_replicas → external metrics API empty → ShareGPT test fails

Before fix: 1 Passed, 2 Failed, 22 Skipped
After model ID fix only: 3 Passed, 1 Failed, 21 Skipped
Expected after both fixes: 4+ Passed, 0 Failed

Companion PR: llm-d/llm-d-infra#20 (sets WVA_METRICS_SECURE=false in reusable workflow)

Note: Nightly workflow temporarily points to llm-d-infra@fix/nightly-metrics-secure. Will revert to @main after llm-d-infra#20 merges.

Test plan

• Trigger nightly from this branch — all core tests should pass
• Verify MetricsAvailable=True in metrics pipeline verification step
• Verify wva_desired_replicas is accessible via external metrics API

[github-actions]

GPU Pre-flight Check ✅

GPUs are available for e2e-openshift tests. Proceeding with deployment.

Resource	Total	Allocated	Available
GPUs	50	26	24

Cluster	Value
Nodes	16 (7 with GPUs)
Total CPU	993 cores
Total Memory	10383 Gi
GPUs required	4 (min) / 6 (recommended)

[Copilot]

Pull request overview

This PR fixes two issues affecting nightly E2E tests on OpenShift. The first fix updates the DEFAULT_MODEL_ID to match the llm-d repository's current default model, ensuring that the yq model replacement logic works correctly. The second fix adds a WVA_METRICS_SECURE environment variable to disable bearer token authentication on the WVA /metrics endpoint, which is needed for OpenShift's user-workload-monitoring to successfully scrape metrics. The PR also temporarily references a companion PR branch in the workflow configuration.

Changes:

• Update DEFAULT_MODEL_ID from "Qwen/Qwen3-32B" to "Qwen/Qwen3-0.6B" to align with llm-d repository defaults
• Add WVA_METRICS_SECURE environment variable (default: true) to control metrics endpoint authentication, wired through to Helm chart
• Temporarily reference llm-d-infra@fix/nightly-metrics-secure branch until companion PR merges

Reviewed changes

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

File	Description
deploy/install.sh	Updates DEFAULT_MODEL_ID constant and adds WVA_METRICS_SECURE environment variable with default value, wiring it to Helm chart via --set flag
.github/workflows/nightly-e2e-openshift.yaml	Temporarily references fix/nightly-metrics-secure branch of llm-d-infra reusable workflow instead of @main

[Copilot]

This temporary reference to the fix/nightly-metrics-secure branch deviates from the established pattern of referencing reusable workflows at @main. While documented in the PR description as temporary, this creates a merge dependency where this PR cannot be merged and tested properly until llm-d-infra#20 is merged first. Consider either: 1) merging llm-d-infra#20 first and updating this to @main, or 2) keeping this at @main and accepting that the nightly workflow may fail until llm-d-infra#20 is merged. The current approach blocks independent verification of these fixes.

Suggested change

	uses: llm-d/llm-d-infra/.github/workflows/reusable-nightly-e2e-openshift.yaml@fix/nightly-metrics-secure
	uses: llm-d/llm-d-infra/.github/workflows/reusable-nightly-e2e-openshift.yaml@main

[github-actions]

GPU Pre-flight Check ✅

GPUs are available for e2e-openshift tests. Proceeding with deployment.

Resource	Total	Allocated	Available
GPUs	50	26	24

Cluster	Value
Nodes	16 (7 with GPUs)
Total CPU	993 cores
Total Memory	10383 Gi
GPUs required	4 (min) / 6 (recommended)

[clubanderson]

Superseded by #721 which also includes the KEDA APIService conflict fix.