ai-rm protocol v0.1: AcceleratorProfile, airm_* metrics and placement in the console

CHANGELOG.md

@@ -4,6 +4,9 @@
 
 ## [Unreleased]
 
+### ai-rm protocol v0.1 (read-only, 2026-10-19)
+- Infra ▸ Devices shows an ai-rm panel when the cluster has any ai-rm source: `AcceleratorProfile` (`airm.ai-computing.io/v1alpha1`) vendor·model·memory·peak with its source (the selected device's profile is marked), `airm_*` Prometheus series with rm-core v0.1 names and units (vendor-derived `tok_s`/`req_s`/`power_w`/`accelerators` shown as `[vendor]` when no airm series exists), and the decisions in ConfigMap `airm-placement` (`LMD_AIRM_NS`, default `llm-d`). The device PWR column prefers `airm_power_w{scope="card"}`. Without the CRD, series and ConfigMap the view is unchanged and nothing is warned. Write actions are unchanged.
+
 ### Failing containers explain themselves (2026-10-19)
 - Containers that are crash-looping, failing to pull, OOMKilled, or restarted in the last 10 minutes are found cluster-wide (from the pod list the inventory already reads) and shown with the line that explains them: the waiting message, the termination message, or — for a few pods per cycle, cached per restart — `logs --previous`. Fatal/panic lines win over later noise; logfmt `msg=… error=…` is reduced to its message.
 - Known causes get a fix: image without a build for the node's architecture (`exec format error`), inotify instances exhausted, disk full, config apiVersion or CLI flag rejected by the image (noting a floating tag), the node unable to reach the API server, OOMKilled, missing tag / denied pull. Failures on a NotReady node blame the node and sort last.

README.md

@@ -235,6 +235,9 @@ python3 -m unittest discover -s tests -p 'test_*.py'    # 설치 스크립트 
 | `LMD_PROM` | Prometheus `host:port` (평문 HTTP, NodePort 또는 port-forward) |
 | `LMD_NS` | 모델 서버·EPP가 있는 네임스페이스 (기본 `llm-serving`) |
 | `LMD_GRAFANA` | Grafana 기본 URL (`:graf` 명령) |
+| `LMD_AIRM_NS` | ai-rm ConfigMap `airm-placement`의 네임스페이스 (기본 `llm-d`) |
+
+**ai-rm protocol v0.1 (읽기 전용).** Infra ▸ Devices 아래에 ai-rm 패널이 나온다: `AcceleratorProfile`(`airm.ai-computing.io/v1alpha1`)의 벤더·모델·메모리·peak(출처), Prometheus `airm_*` 지표(rm-core `metrics.py`의 이름·단위, 없으면 기존 vendor 값을 `[vendor]`로 표시), ConfigMap `airm-placement`의 배치 결정과 이유. 장치별 PWR 열은 `airm_power_w{scope="card"}`가 있으면 그 값을 쓴다. CRD·지표·ConfigMap이 모두 없으면 패널은 나오지 않는다. `airm-placement`의 각 data 키는 워크로드 이름, 값은 `{"accelerator_ids": [...], "kv_transport": "...", "reason": "..."}`(JSON 또는 YAML)이다.
 
 ## Contributing & license
 

ai-rm.yaml

@@ -1,4 +1,5 @@
 name: lmd-top
+protocol: "0.1"
 layer: l5-ops
 kind: tool
 summary: llm-d 클러스터의 가속기·서빙·EPP 라우팅·SLO를 한 화면에서 보고 조정하는 터미널 관측·운영 콘솔(ai-rm 관측 콘솔)
@@ -19,3 +20,5 @@ interfaces:
   provides: [Metrics, Placement]          # --json 상태·KPI 관측, heterogeneous-capacity EPP 정책(contrib/epp-score-observer)
   uses: [Metrics, Accelerator, Workload, Placement]   # Prometheus 지표, 가속기 팩·device plugin 리소스, serving/학습 Job, InferencePool/EPP
 kpi: [이종자원수, SLO, 전력효율]
+adapters: []                  # Rust; reads rm-core formats (AcceleratorProfile, airm_* series, airm-placement) without importing rm-core
+metrics: []                   # reads airm_* only, reports none

cluster.example.env

@@ -12,3 +12,6 @@ LMD_NS=llm-serving
 
 # Grafana base URL used for dashboard links
 LMD_GRAFANA=http://<grafana-host>:<port>
+
+# Namespace of the ai-rm ConfigMap airm-placement (read-only placement decisions)
+LMD_AIRM_NS=llm-d