[gspark-etri]

Roadmap W2 (maturity): helm install reproduces the control path, and a sim e2e runs on every PR.

What

• charts/ai-rm (umbrella)
  • L1: rm-accel's rm-abstract-node chart as a subchart (alias node, file:// from the components/l1-accel/rm-accel submodule; helm dependency update charts/ai-rm)
  • crds/: AcceleratorProfile CRD copied verbatim from rm-core v0.1.1 k8s/
  • L4 rm-controller: Deployment (single writer, non-root, read-only root), ConfigMap with its spec built from controller.pools + placementNamespace, ClusterRole (profiles, nodes: read) and a Role limited to airm-placement
  • airm-placement: empty v0.1 document, rendered only when the ConfigMap does not exist (lookup) and helm.sh/resource-policy: keep, so rm-controller owns it after install and upgrades don't fight it. Coexists with hetero-llm-d's airm-install.sh.
  • Example site values deploy/clusters/example/ai-rm.values.yaml; deploy/clusters/example/rm-abstract-node.values.yaml no longer carries the init-container workaround.
• rm-controller image: packages/rm-controller/Dockerfile (python:3.11-slim, rm-core tarball, no git needed, uid 65534) + .github/workflows/rm-controller-image.yml (build+smoke on PR; push :master, and :X.Y.Z from tags rm-controller-vX.Y.Z; GITHUB_TOKEN, packages: write).
• Simulated cluster: tools/fake_openai.py (stdlib OpenAI-compatible server: /v1/models, streaming/plain chat, completions with pd timing), tools/airm_sim.py behind tools/airm.py up-sim | e2e-sim [--no-cluster] | down-sim and make chart | up-sim | e2e-sim | down-sim.
  • up-sim: kind cluster airm-sim (control-plane + worker), builds airm/rm-accel:sim (rm-accel Dockerfile) and airm/rm-controller:sim, loads them, installs the chart with ci/sim-values.yaml (provider sim, no host access), waits for 2 profiles per node.
  • e2e-sim: kubectl exec … rm-controller --once (exit 0), reads airm-placement, asserts each pool has prefill+decode with the requested card kinds on one node, no card reused, the two pools on different nodes; runs rm-bench e2e --scenario pd-smoke against the fake pool and asserts exit 0 and slo_attainment_pct = 100 (0 errors).
  • .github/workflows/sim-e2e.yml: chart job (helm lint/template: default, example site, sim) and e2e job (unit tests, kind, make up-sim, make e2e-sim, cluster dump on failure, results artifact).
• Docs: deploy/README.md + .ko.md (umbrella chart, simulated cluster), docs/getting-started.md + ko (section "simulated cluster"), rm-controller README/manifest.

Depends on

• Node daemon image ghcr.io/ai-computing/rm-accel; chart 0.5.0 uses it rm-accel#8 (node image, chart 0.5.0 with hostBin). The submodule here points at that PR's head (fba24e3); after it merges, make bump COMPONENT=rm-accel (or --ref v0.4.2) to the merge commit / tag.
• After merge: make the new ghcr packages rm-controller / rm-accel public and tag rm-controller-v0.1.0 so the chart's default controller.image.tag: 0.1.0 exists (:master is published by the merge).

Verified

• helm lint / helm template (v3.16.4, Windows binary): default, example site and sim values pass; rendered YAML has no duplicate keys (12 objects for sim).
• python -m pytest -q tools/tests: 17 passed (fake server endpoints/streaming, placement and bench checks incl. 8 failure cases, e2e-sim --no-cluster).
• python tools/airm.py e2e-sim --no-cluster: 2 pools on 2 sim nodes (RNGD prefill + ATOM decode each), rm-bench pd-smoke sent=32 failed=0, ttft_p90 ≈ 20 ms, PASS.
• rm-controller / rm-bench package tests: 12 / 26 passed. The Dockerfile install sequence was replayed in a clean venv (rm-controller --help OK).
• tools/check_docs.py PASS, tools/check_sensitive.py . clean.
• Docker/kind are not available on the dev host (Windows, no Docker in WSL): the kind run is verified by this PR's sim-e2e job.

🤖 Generated with Claude Code