dynamo-interconnect-check

par nvidia

Valider que le déploiement Dynamo est prêt pour l'interconnexion NIXL/UCX/NCCL en vue d'un service désagrégé sur RDMA/NVLink. Utiliser après que recipe-runner a mis en place un déploiement…

npx skills add https://github.com/nvidia/skills --skill dynamo-interconnect-check

Dynamo Interconnect Check

Purpose

Confirm that the transport disaggregated serving depends on actually works. A deployment can pass an endpoint smoke test while disagg is silently wrong: if NIXL/UCX cannot reach the peer worker over RDMA or NVLink, KV transfer falls back to a slow or broken path. Catch that with read-only checks before trusting a disagg deployment or its benchmark numbers.

This skill is read-only. It never mutates the cluster and never prints secrets.

Prerequisites

  • Python 3.10+ on the operator machine.
  • kubectl exec access to a worker pod in the target Dynamo deployment.
  • Read access to the recipe directory (recipes/<model>/<framework>/<mode>).
  • For node-capability checks: tools like ibstat, nvidia-smi, lsmod available in the worker pod image (missing tools are reported as skipped, not failures).

When To Use

  • After dynamo-recipe-runner deploys a disagg or multi-node recipe.
  • Before reporting disagg throughput/latency, so numbers reflect the real transport.
  • When agg works but disagg is slow, hangs, or returns wrong output and you suspect the fabric rather than the model.

For diagnosing pods that are already crashing or unschedulable, use dynamo-troubleshoot first.

Instructions

1. Check Transport Env Vars On The Recipe

python3 scripts/check_interconnect.py env recipes/<model>/<framework>/<mode>

Reports which NIXL/UCX/NCCL transport variables are set and flags disagg-critical ones (e.g. UCX_TLS, UCX_NET_DEVICES, NCCL_IB_HCA) that are absent. Missing here is only a warning — they may be baked into the image — so confirm with the node and NIXL checks. See references/interconnect-env-vars.md for what each variable does.

2. Check Node Capabilities

Locally on a GPU node, or inside a running worker pod:

python3 scripts/check_interconnect.py node \
  --namespace "${NAMESPACE}" --pod <worker-pod>

Probes (read-only) for: InfiniBand devices and Active links, GPUDirect RDMA (nvidia_peermem), GDRCopy, and NVLink in the GPU topology. Missing tools are reported as skipped, not failures.

3. Validate NIXL Reachability

python3 scripts/check_interconnect.py nixl \
  --namespace "${NAMESPACE}" --pod <worker-pod>

Looks for NIXL test tooling in the pod and surfaces the exact next step to run a pairwise prefill↔decode transfer test. A full cross-pod transfer test requires two scheduled GPU pods on the fabric.

Available Scripts

ScriptPurposeArguments
scripts/check_interconnect.py envInspect NIXL/UCX/NCCL env vars on a recipepositional recipe path
scripts/check_interconnect.py nodeProbe InfiniBand, GPUDirect RDMA, GDRCopy, NVLink on a node or pod--namespace, --pod
scripts/check_interconnect.py nixlSurface NIXL transfer-test readiness for a pod--namespace, --pod

Invoke via the agentskills.io run_script() protocol:

run_script("scripts/check_interconnect.py", args=["env", "recipes/qwen3-coder-480b/sglang/disagg"])
run_script("scripts/check_interconnect.py", args=["node", "--namespace", "dynamo-demo", "--pod", "qwen-worker-0"])

Examples

Verify a disagg recipe's transport env shape before deploy:

python3 scripts/check_interconnect.py env recipes/qwen3-coder-480b/sglang/disagg

After deploy, validate a worker pod's fabric:

python3 scripts/check_interconnect.py node \
  --namespace dynamo-demo --pod qwen-worker-0
python3 scripts/check_interconnect.py nixl \
  --namespace dynamo-demo --pod qwen-worker-0

Equivalent through the agent protocol:

run_script("scripts/check_interconnect.py", args=["nixl", "--namespace", "dynamo-demo", "--pod", "qwen-worker-0"])

Output Contract

Each check returns ok / warn / fail / skipped with a one-line detail, plus a rolled-up verdict on disagg transport readiness. Report:

  • transport env vars present vs. disagg-critical ones missing
  • RDMA / GPUDirect / NVLink capability status
  • whether NIXL reachability was validated, and the next command if not
  • a clear statement of whether disagg can be trusted, or what to fix first

Limitations

  • Read-only fabric probe; does not run a full pairwise NIXL transfer (requires two scheduled GPU pods and the in-pod NIXL test tools).
  • skipped results for missing tools (ibstat, nvidia-smi, lsmod) are inconclusive, not a pass.
  • Env-var check inspects the recipe text; values injected at runtime via initContainers or operator-applied envs are not detected.
  • Single-node agg deployments do not exercise the transport — this skill is for disagg / multi-node validation.

Troubleshooting

SymptomLikely causeNext step
env reports all critical vars missingVars baked into image or injected by operatorRun the node check inside the worker pod to verify actual env
node reports no Active IB linkFabric down or HCA not provisioned to the nodeContact cluster admin; verify kubectl describe node shows nvidia.com/gpu and IB labels
nvidia_peermem missingGPUDirect RDMA module not loadedAsk cluster admin to load nvidia-peermem; without it, NIXL falls back to staged copies
nixl finds no test toolsWorker image lacks NIXL test harnessUse a NIXL-enabled image or run the standalone transfer test from a debug pod

Benchmark

See BENCHMARK.md for the NVCARPS-EVAL performance report (auto-generated by the NVSkills CI pipeline). To refresh, re-run /nvskills-ci on an upstream PR touching this skill.

References

  • references/interconnect-env-vars.md — NIXL/UCX/NCCL env var catalog and IB capability checklist.
  • Use scripts/check_interconnect.py for all read-only checks.

Plus de skills de nvidia

compileiq-debug
nvidia
Utilisez quand quelque chose ne va pas : Search() bloque, toutes les évaluations retournent INVALID_SCORE, les scores ne s'améliorent pas, chaque configuration retourne le même nombre, erreurs ptxas…
create-github-pr
nvidia
Créer des pull requests GitHub en utilisant l'interface en ligne de commande gh. Utiliser lorsque l'utilisateur souhaite créer une nouvelle PR, soumettre du code pour révision, ou ouvrir une pull request. Mots-clés de déclenchement -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Analyse les autres problèmes ouverts pour trouver ceux qu’une PR donnée pourrait également corriger ou casser accidentellement. Génère des opportunités de correctifs adjacents et des risques de contradiction avec fichier:ligne…
fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
maintain-dynamic-plugins
nvidia
Maintenir les chargeurs de plugins dynamiques NeMo Relay, les manifestes, les SDK natifs Rust, le protocole worker gRPC, le SDK worker Python, la documentation, les tests et la couverture du workflow de publication
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…