Unify ARO and ROSA Helm charts under single charts - #15
Merged
Conversation
Update gitops.basePath to charts/ and add gitops.cloud (aro|rosa) parameter in both Terraform bootstraps. Update Makefile CHARTS_DIR.
AGENTS.md reflects the new charts/ layout and six-wave structure. deploy_existing_openshift/README.md drops stale PCA_Deployment_ROSA/charts path references. Add unify-aro-rosa-plan.md to local-docs.
Single chart set replaces the two parallel chart trees. ROSA charts serve as the base; cloud-specific config lives in values-aro.yaml values-rosa.yaml per chart. ARO gains observability, guardrails, MCP, and full IDE coverage (Roo Code + Continue + Cline). ROSA gains EPP, TLS cert job, benchmarks, and DevSpaces dashboard samples.
…pipelines - Delete standalone vscode-extensions-config.yaml (already created in devspaces-global-config.yaml with full 4-extension list) - Rename datasciencepipelines → aipipelines in DataScienceCluster template (RHOAI 3.4 renamed the component; fixes ArgoCD ComparisonError on ARO) Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Sprig's default function treats boolean false as empty, so `false | default true` returns true. Use `not (eq false ...)` instead, which correctly respects an explicit false value. Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Replace lookup+randAlphaNum with sha256sum(ns/release/pca-grafana). lookup and randAlphaNum produce different values on each ArgoCD sync, causing constant drift. Deterministic hash survives restarts and syncs. Add checksum annotation to Deployment to roll on password change. Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
…spaces fields - tlsJob.enabled: false (RHOAI auto-creates qwen36-vllm-kserve-self-signed-certs) - aiGateway.tls.secretName updated to match RHOAI-created cert - observability.enabled: true with Grafana + Langfuse + OTel for ARO - pca-observability.prometheus.accessMode: namespace (existing OCP) - devspaces list uses correct field names (name/user not workspaceName/username)
- Remove continue-config from devspaces-global-config.yaml; DevSpaces operator propagates global CMs into user namespaces (subpath mount, read-only) and would clobber per-ns API key/headers. Continue config lives only in per-namespace continue-configmaps.yaml. - DSC: modelsAsService → Removed, remove modelmeshserving (RHOAI 3.4) - NFD image: v4.17 → v4.21 Co-Authored-By: Itay Katav <itay1551@users.noreply.github.com> Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Add vllm.useCustomRuntime flag for ARO where RHOAI bundled vLLM does not support Qwen3.5-MoE (Qwen3.6-35B requires vLLM >= 0.18). When enabled, renders ServingRuntime + InferenceService instead of LLMInferenceService. - servingruntime.yaml: new, conditional on vllm.useCustomRuntime - inferenceservice.yaml: new, conditional on vllm.useCustomRuntime - llminferenceservice.yaml: gated on not .Values.vllm.useCustomRuntime - llm-d-gateway-httproute.yaml: InferencePool rules only when epp.enabled; fallback service uses configurable model.workloadServiceSuffix (default kserve-workload-svc for ROSA, predictor for ARO) - tls-secret-job.yaml: fix CN > 64 char limit; use short CN=llm-d-gateway, full FQDN in SAN only (modern TLS ignores CN, checks SAN) - values.yaml: add vllm.useCustomRuntime: false default - values-aro.yaml: useCustomRuntime: true, tlsJob: true, predictor suffix, full observability enabled, CUDA compat env vars removed (driver 580+ supports vLLM 0.19.0 natively without compat shims) Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
- llm-d-gateway-httproute: fallback service uses port 80 (KServe service exposes 80→8000; port 8000 is pod-direct only for InferencePool/EPP) - pca-observability/values.yaml: update adminPassword comment to reflect deterministic hash approach (aligned with upstream PR rh-ai-quickstart#14) - values-aro.yaml (operators): nfd.enabled: true — ARO requires NFD to label GPU nodes with nvidia.com/gpu.deploy.* so driver DaemonSet schedules Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
kserve-workload-svc (ROSA/LLMInferenceService) exposes port 8000. InferenceService predictor (ARO/useCustomRuntime) exposes port 80→8000. Add model.workloadServicePort (default 8000) and override to 80 in values-aro.yaml so the llm-d gateway HTTPRoute uses the correct port for each cloud. Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
- pca-observability.namespace: hacohen-ai-coder → ai-serving (test artifact)
- add {{- fail }} guard for empty gitops.cloud in app-of-apps template
- pca-benchmarks/values.yaml: enabled: false (opt-in, not opt-out)
- hardware-profiles.yaml: add | int cast for cpu counts (HardwareProfile CRD expects integer)
- subscriptions.yaml: double-chain serviceMesh/serverless guards to match NFD pattern
- datasciencecluster.yaml: add modelmeshserving: Removed explicitly
- values-aro.yaml: epp.enabled: false (InferencePool/InferenceModel CRDs absent on RHOAI 3.3.1)
Hadar301
marked this pull request as ready for review
July 23, 2026 09:12
Collaborator
|
Thank you for the huge PR @Hadar301 I see that you wrote:
The 18 fails on ROSA deployment, wasn't pre-existing. Regarding the existing OpenShift cluster, we had 0 failures during the full deployment. We need to merge when we don't have failed tests. |
itay1551
approved these changes
Jul 23, 2026
- model.name: qwen36-vllm → qwen3-coder in values-aro.yaml so service names and pod labels match ROSA smoke test expectations - model.servedName: new field lets vLLM serve a model under a different API name (ARO: serves Qwen3.6-35B as Qwen3-Coder-30B for test parity) - vllm.enableThinking: false adds --chat-template-kwargs to ServingRuntime args to disable Qwen3 thinking mode per cloud (default true = enabled) - epp.enabled: false in values-aro.yaml (InferencePool CRDs absent RHOAI 3.3.1) - pca-observability.namespace: ai-serving (was hacohen-ai-coder test artifact) - ARO HELM_ARGS documented in deploy-existing-openshift Cursor skill Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Collaborator
Author
ARO31 passed / 17 failed / 6 skipped — full stack deployed: Qwen3.6-35B on H100 NVL via vLLM v0.19.0, Grafana, Langfuse, OTel, RHCL gateway, Continue/Roo Code/Cline. Additional ARO-specific failures beyond the ROSA infrastructure issues above:
Proposed test fixes (separate PR)
Full per-test table: |
4 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
PCA_Deployment_ARO/charts/andPCA_Deployment_ROSA/charts/) into a singlecharts/directory at the repo rootvalues-aro.yamlandvalues-rosa.yamloverlays;pca-app-of-appsselects the overlay viagitops.cloud: aro|rosainjected from Terraformpca-benchmarkschart (ported from ARO, disabled on ROSA by default)oc runstdout mixing), not regressionsKey changes
Chart unification
charts/pca-app-of-apps: addsgitops.cloudparam and Wave 5 forpca-benchmarkscharts/pca-operators: gates NFD / ServiceMesh / Serverless on feature flags; channels parameterized per cloudcharts/pca-platform-config:datasciencepipelines→aipipelines(RHOAI 3.4); NFD image v4.21charts/pca-ai-serving: model/hardware/storage/vLLM fully parameterized; addsvllm.useCustomRuntimeflag for ARO (usesServingRuntime+InferenceServiceinstead ofLLMInferenceServicebecause RHOAI 3.3.x bundles vLLM ~0.13 which does not support Qwen3.5-MoE); EPP/TLS job/InferenceRouting conditional; TLS CN length fix (≤64 chars); fallback service port configurable per cloudcharts/pca-devspaces: removes globalcontinue-config(DevSpaces operator subpath propagation overwrites per-ns keys); onlyvscode-extensions-configremains global; dashboard samples and OpenCode build ported from AROcharts/pca-benchmarks: new chart from ARO, parameterized,enabled: falseby default on ROSAcharts/pca-ai-serving/charts/pca-observability: Grafana password uses deterministic sha256 hash (ArgoCD-safe;randAlphaNum/lookupdrift on every sync)Infrastructure
gitops-bootstrap.tf:gitops.basePath→charts,gitops.cloud: aro|rosaaddedMakefile:CHARTS_DIR := chartsPCA_Deployment_ARO/charts/andPCA_Deployment_ROSA/charts/deletedDocumentation
AGENTS.md: updated directory structure and 6-chart wave tableREADME.md: deployment guide extracted to newDEPLOYMENT.mdlocal-docs/aro-rosa-gap-analysis.md: documents all ARO/ROSA gaps found during testing (vLLM compat, GuardrailsOrchestrator controller removed in RHOAI 3.4, Sprig boolean bug, ArgoCD cascade deletion risk).cursor/skills/deploy-existing-openshift/SKILL.md: kuadrant shared-cluster workaround, guardrails teardown finalizer fixTest plan
helm templatedry-run all 6 charts × 2 cloud values — no errorsmake smoke AI_NAMESPACE=hacohen-ai-coder DEV_NAMESPACE=hacohen-dev1-devspaces— ROSA: 34 pass / 14 fail (pre-existing) / 6 skipmake smoke AI_NAMESPACE=hacohen-ai-coder DEV_NAMESPACE=hacohen-dev1-devspaces— ARO: 30 pass / 18 fail (pre-existing + model name mismatch) / 6 skip