[haarchri]

Description of your changes

On a Civo cluster Modelplane provisions, every modelplane_gpu_* series and modelplane_energy_joules_total was empty. The hand-written Civo stack disabled the GPU Operator's DCGM exporter - not catched as other handwritten installing the GPU Operator as pre-req.
The telemetry collector discovers a GPU exporter by pod label, so with no exporter DaemonSet there was nothing to scrape.

This enables the exporter on Civo with the same dcgm/dcgmExporter values the generated clouds pin, so one review moves both halves; the metrics CSV covers every field metrics.py maps. The gpu-operator chart gains a kube-prometheus-stack edge: unlike the generated clouds, Civo's kube-prometheus-stack keeps its bundled CRDs enabled, so that release is what provides the ServiceMonitor CRD, and it now sets wait=True as a depended-on chart must.

On source: Existing clusters Modelplane doesn't install an exporter, on purpose: it can't know whether the cluster already runs one, and the collector scrapes every pod matching .*dcgm.*, so a second exporter would collect each GPU twice. The docs now say so. A new GPUs section in the telemetry page documents the discovery contract, the seven DCGM fields with their unit conversions, and how to check what an exporter publishes; the existing-cluster requirements gain a DCGM exporter bullet pointing at it.

Fixes #519.
Towards #524.

I have:

• Read and followed Modelplane's contribution process.
• Run nix flake check (or ./nix.sh flake check) and made sure it passes.
• Added or updated tests covering any composition function changes.
• Signed off every commit with git commit -s.

Validated

on a live Civo NYC1 cluster: all seven renamed series arrive at the collector's debug sink.

    kubectl logs deploy/modelplane-collector -n modelplane-system | grep -E 'modelplane_gpu_|modelplane_energy' | sort -u
    -> Name: modelplane_energy_joules_total
    -> Name: modelplane_gpu_compute_active_ratio
    -> Name: modelplane_gpu_memory_bandwidth_ratio
    -> Name: modelplane_gpu_memory_used_bytes
    -> Name: modelplane_gpu_power_watts
    -> Name: modelplane_gpu_temperature_celsius
    -> Name: modelplane_gpu_tensor_active_ratio

[github-actions]

Docs preview: https://modelplane-docs-pr-526.vercel.app (ready once the site's Content workflow finishes)

[dennis-upbound]

Agree on not installing an exporter on Existing. It's the same call as the GPU driver: a cluster running the GPU Operator already has one, and a second double-counts every GPU, which is worse than empty series. Documenting the contract plus the curl check is right.

On the Civo half:

1. Is the ServiceMonitor needed? The collector scrapes the exporter pods directly, and Civo's Prometheus only opens up PodMonitor selection, so as far as I can tell nothing reads it. With dcgmExporter.serviceMonitor.enabled: false, the new kube-prometheus-stack edge, its wait=True, and the Usage all go away, and the GPU Operator and DRA driver don't wait on a healthy Prometheus. The cost is Civo's values no longer matching the generated clouds verbatim. Your call on which matters more.
2. The CSV is a copy. It matches GKE/EKS/AKS today, but nothing enforces it. A test asserting Civo's CSV equals the generated one would make "one review moves both" true.
3. Nit: the E501 ignore covers all of civo.py; a # noqa: E501 on the CSV string would keep it scoped.
4. Nit: the new inference-cluster.md bullet repeats most of the telemetry GPUs section. A couple of lines and a link would do.

[haarchri]

updated the PR:

• ServiceMonitor: dropped. Verified your read
• CSV: went one further than the test, the aicr generator now emits a shared generated/aicr/dcgm.py (fails closed if the clouds CSVs ever diverge) and civo.py imports it. A test pins the generated inline copies to the shared module.
• E501: with the literal gone, civo.py has no long lines, ignore deleted entirely.
• Docs bullet: trimmed to the contract plus a link to the telemetry GPUs section.

[dennis-upbound]

Thanks, both points are handled, and the generated dcgm.py with its fail-closed checks is better than the test I suggested. Two nits before merge: the commit subjects carry typed prefixes (fix(servingstack):, feat(refactor):), which CONTRIBUTING.md asks us to drop, and the new paragraph in serving-stack-generation.md could lose its em dashes and split its long middle sentence.