AutoGineer
Generated banner for Dynamo Troubleshoot
Generated logo for Dynamo Troubleshoot

Dynamo Troubleshoot

NVIDIAOfficialCreated by Dan GilDistributed by NVIDIA

  • Skill
  • Review and audit
  • Cloud and DevOps
  • Apache-2.0
  • Medium risk
  • Has scripts
Download ZIP

From the author

Diagnose failed or unhealthy Dynamo deployments. Use when pods, model-cache jobs, PVCs, workers, frontend/router health, endpoints, or benchmark jobs fail; use recipe-runner/router-starter before this for normal bring-up.
— the description by Dan Gil

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

Dynamo Troubleshoot provides a structured, top-down methodology for diagnosing issues within Dynamo deployments. It utilizes a read-only debug bundle collection script to gather cluster state, events, and resource status, allowing operators to classify failures into specific buckets such as storage, GPU scheduling, or operator reconciliation. The skill emphasizes a systematic approach to troubleshooting, ensuring that users identify the root cause before applying minimal, reversible changes to their cluster configuration.

What it does

  • Collects a read-only debug bundle containing pods, events, jobs, and PVC status.
  • Provides a decision tree to classify failures into specific operational buckets.
  • Guides users through a top-down diagnostic order from infrastructure to application endpoints.
  • Suggests specific remediation steps for common failure modes like PVC binding or GPU scheduling.
  • Ensures safe troubleshooting by never executing mutations or collecting sensitive secrets.

When to use it

  • When a Dynamo deployment fails to reach a healthy state.
  • When pods, model-cache jobs, or benchmark jobs enter a failed or pending status.
  • When frontend or router services are unresponsive or failing to register workers.
  • When you need to audit the status of a DynamoGraphDeployment resource.

Advantages

  • Standardizes the diagnostic process for Dynamo environments.
  • Minimizes risk by focusing on read-only data collection and manual remediation.
  • Reduces mean time to resolution by categorizing failures into actionable buckets.
  • Prevents accidental exposure of sensitive information like Hugging Face tokens.

Usage examples

“Collect debug information for the dynamo-demo namespace.”
The agent runs the collection script and returns a bundle of cluster state, events, and resource status for the specified namespace.
“Troubleshoot a failing deployment named qwen-vllm-disagg in the production namespace.”
The agent collects scoped debug data for that specific deployment and provides a classification of the failure based on the decision tree.

Install

  1. Download the skill as a ZIP:

    https://autogineer.com/agent-skills/nvidia/dynamo-troubleshoot/download

  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip dynamo-troubleshoot.zip -d .agents/skills

Documentation by Dan Gil

Dynamo Troubleshoot

<!-- SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: CC-BY-4.0 -->

Purpose

Turn a Dynamo failure into a clear problem class, strongest signal, and next action. Start with read-only evidence, avoid secrets, and fix one layer at a time.

Prerequisites

  • Python 3.10+ on the operator machine.

  • kubectl configured with read access to the target namespace.

  • Permission to read pods, events, jobs, PVCs, and DynamoGraphDeployment resources (NOT secrets).

  • Network reachability to the cluster API server.

Instructions

1. Collect A Read-Only Bundle

Run:

python3 scripts/collect_dynamo_debug_bundle.py \
  --namespace "${NAMESPACE}"

If the user names a deployment, include it:

python3 scripts/collect_dynamo_debug_bundle.py \
  --namespace "${NAMESPACE}" \
  --deployment-name <deployment-name>

Do not collect Kubernetes secrets. Do not print Hugging Face tokens.

2. Classify The Failure

Use references/failure-decision-tree.md and classify into one primary bucket:

  • cluster/platform

  • namespace/secret

  • model cache/PVC/download

  • image pull/runtime image

  • GPU scheduling/resources

  • operator/DynamoGraphDeployment reconciliation

  • frontend/router

  • worker/backend

  • endpoint/API

  • benchmark/perf job

3. Debug Top Down

Check in this order:

  1. namespace, storage class, GPU nodes, and HF secret existence

  2. PVC and model-download job

  3. DynamoGraphDeployment status and events

  4. pod status, describe pod, and container logs

  5. frontend service and port-forward

  6. /v1/models

  7. /v1/chat/completions

  8. benchmark job only after endpoint smoke test passes

4. Fix One Layer At A Time

Prefer the smallest reversible change:

  • create missing namespace or HF secret

  • patch storageClassName

  • patch image tag or image pull secret

  • reduce GPU request only if the recipe can still be valid

  • switch KV router to approximate mode only if workers do not publish events

  • restart failed jobs after fixing the underlying config

After each fix, rerun the relevant readiness check before moving deeper.

Available Scripts

ScriptPurposeArguments
scripts/collect_dynamo_debug_bundle.pyCollect a read-only debug bundle (pods, events, jobs, PVCs, CR status)--namespace, --deployment-name, --output-dir

Invoke via the agentskills.io run_script() protocol:

run_script("scripts/collect_dynamo_debug_bundle.py", args=["--namespace", "dynamo-demo"])

Examples

Collect everything in a namespace for triage:

python3 scripts/collect_dynamo_debug_bundle.py --namespace dynamo-demo

Scope to a single failing deployment:

python3 scripts/collect_dynamo_debug_bundle.py \
  --namespace dynamo-demo \
  --deployment-name qwen-vllm-disagg

Equivalent through the agent protocol:

run_script("scripts/collect_dynamo_debug_bundle.py", args=["--namespace", "dynamo-demo", "--deployment-name", "qwen-vllm-disagg"])

Output Contract

Return:

  • problem class

  • evidence checked

  • strongest signal

  • likely cause

  • exact next command or patch

  • what was ruled out

  • whether it is safe to continue deployment or benchmarking

Limitations

  • Read-only. Never mutates the cluster; remediation commands are returned, not executed.

  • Will not collect secrets or print Hugging Face tokens; some failure modes (auth) may need user-side inspection.

  • Bundle size grows with deployment size; on very large namespaces, scope with --deployment-name.

  • Does not validate disagg transport — use dynamo-interconnect-check for that.

Troubleshooting

SymptomLikely causeNext step
kubectl returns Forbidden on events/podsService account lacks read RBACAsk operator for read-only role binding on the namespace
Bundle missing DynamoGraphDeployment statusOperator not installed or different namespaceVerify dynamo-platform operator is installed and watching the namespace
Model-download job in PendingPVC unbound or HF secret missingFix PVC binding or create the named HF secret, then rerun the job
Worker pods CrashLoopBackOffImage/runtime mismatch or GPU not availableInspect container logs; check nvidia.com/gpu allocatable on nodes

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

  • Read references/failure-decision-tree.md for bucket-specific checks.

  • Use scripts/collect_dynamo_debug_bundle.py for read-only bundle collection.

Files

references/failure-decision-tree.md

Dynamo Failure Decision Tree

<!-- SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: CC-BY-4.0 -->

Cluster Or Namespace

Signals:

  • kubectl config current-context fails

  • namespace does not exist

  • no GPU nodes are visible

Checks:

kubectl config current-context
kubectl get namespace "${NAMESPACE}"
kubectl get nodes -o wide
kubectl describe nodes | grep -A5 -E "nvidia.com/gpu|Capacity|Allocatable"

Next action: create namespace, switch context, or use a GPU-capable cluster.

Secret Or Model Access

Signals:

  • model-download job exits quickly

  • logs show authentication, 401, 403, gated model, or missing token

  • manifest references HF_TOKEN but secret/key does not exist

Checks:

kubectl get secret hf-token-secret -n "${NAMESPACE}"
kubectl logs job/model-download -n "${NAMESPACE}" --tail=100

Next action: create or fix the HF secret. Never paste the token into manifests.

PVC Or Storage Class

Signals:

  • PVC is Pending

  • pod waits on volume mount

  • model-download pod cannot write model cache

Checks:

kubectl get storageclass
kubectl get pvc -n "${NAMESPACE}"
kubectl describe pvc -n "${NAMESPACE}"

Next action: patch storageClassName in the recipe model-cache YAML and recreate the PVC/job if needed.

Image Pull Or Runtime Image

Signals:

  • ImagePullBackOff

  • ErrImagePull

  • auth errors against a private registry

  • backend binary missing at container start

Checks:

kubectl describe pod <pod> -n "${NAMESPACE}"
kubectl get events -n "${NAMESPACE}" --sort-by=.lastTimestamp | tail -50

Next action: patch image tag, add image pull secret, or choose a recipe image that contains the requested backend.

GPU Scheduling

Signals:

  • pod is Pending

  • events mention insufficient nvidia.com/gpu

  • wrong node selector, taint, or toleration

Checks:

kubectl describe pod <pod> -n "${NAMESPACE}"
kubectl describe nodes | grep -A8 -E "nvidia.com/gpu|Taints|Allocatable"

Next action: use the correct recipe for the available GPU SKU/count or adjust scheduling constraints only if the recipe remains valid.

DynamoGraphDeployment Or Operator

Signals:

  • manifest applied but no pods appear

  • DynamoGraphDeployment has reconcile errors

  • CRD is missing

Checks:

kubectl get dynamographdeployment -n "${NAMESPACE}"
kubectl describe dynamographdeployment <name> -n "${NAMESPACE}"
kubectl get crd | grep -i dynamo

Next action: install/fix Dynamo Kubernetes Platform or repair invalid DGD YAML.

Frontend Or Router

Signals:

  • frontend pod ready but /v1/models empty or 503

  • logs show no registered workers

  • KV mode enabled but workers do not publish events

Checks:

kubectl logs <frontend-pod> -n "${NAMESPACE}" --tail=200
kubectl get svc -n "${NAMESPACE}" | grep frontend

Next action: verify workers are ready and registered. For KV smoke tests without worker KV events, set DYN_ROUTER_USE_KV_EVENTS=false.

Endpoint/API

Signals:

  • port-forward succeeds but /v1/models fails

  • chat completion fails while models list works

Checks:

curl http://127.0.0.1:8000/v1/models
curl http://127.0.0.1:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"<model>","messages":[{"role":"user","content":"hello"}],"max_tokens":16}'

Next action: check frontend and worker logs for request-time errors.

Benchmark/Perf Job

Signals:

  • endpoint smoke test passes but perf.yaml job fails

  • benchmark cannot reach service

  • benchmark uses wrong model name or URL

Checks:

kubectl get jobs -n "${NAMESPACE}"
kubectl logs job/<benchmark-job> -n "${NAMESPACE}" --tail=200

Next action: fix benchmark URL/model/concurrency only after the endpoint smoke test passes.

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs Python

Security

What our scanner noticed· Medium risk

  • Large base64-encoded blobs
  • Instructions that use the network
  • Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by
Dan Gil
Publisher
NVIDIA (Official)
Distributed by
NVIDIA
Source
NVIDIA/skills · skills/dynamo-troubleshoot · commit fce3cca
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE-APACHE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by Dan Gil or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download dynamo-troubleshoot.zip