Skip to content

Free VS Code extension · local-first

Diagnose GPU training and inference failures in VS Code.

Diagnose GPU errors: CUDA out of memory, PyTorch crashes, NCCL timeouts, Xid errors, Slurm and vLLM. Free offline log analysis. No account required.

Install from inside VS Code

Open Extensions (Ctrl+Shift+X on Windows/Linux or Cmd+Shift+X on macOS), search for Denpex and choose Install. Or use the Marketplace extension ID from your terminal:

code --install-extension denpex.denpex-diagnostics

This command uses the Marketplace listing without pinning a version.

  • Unlimited local use
  • No account required
  • Remote-SSH ready
Denpex Console showing a local vLLM diagnosis with likely cause, evidence and suggested checks.
The actual Console renderer, captured with a bundled sample log.

See Console, Fleet and Topology

See the same extension screens shown in the Marketplace overview. Local diagnosis is free; connected Fleet and Topology require a reporting account and an eligible plan.

Inside the extension

Captured extension screens with sample logs and telemetry. Choose a view to keep it on screen.

Denpex Console showing a local vLLM diagnosis with likely cause, evidence and suggested checks.
Inspect the likely cause and the log lines supporting it, without leaving VS Code.

Diagnose failed training and inference from the evidence you have

GPU cloud operators, platform and infrastructure teams, SREs, ML engineers and researchers can use Denpex to investigate saved PyTorch, CUDA, NCCL, scheduler and inference logs. Start with the earliest supported failure, its rank and the lines that support it. A watchdog timeout on another rank may be a downstream symptom. Missing rank logs, unsynchronized clocks or absent host events can leave the initiating cause unresolved.

Choose the right diagnostic evidence

Complementary tools for saved logs, collective traces, memory and host faults
EvidenceUse it forImportant limit
Saved logs with DenpexInspect covered error signatures, supporting lines and rank-failure cascades locally in your editor.Cannot reconstruct events that were never recorded or prove recovery from a proposed action.
PyTorch Flight Recorder dumpsCompare per-rank collective records when investigating NCCL hangs and mismatches.Requires relevant tracing to have been enabled and dumps to survive the incident.
PyTorch memory snapshotsInvestigate recorded allocations and CUDA out-of-memory events.Allocation history must be recorded; memory outside the PyTorch allocator may be missing.
Kernel Xid events and NVIDIA diagnosticsInvestigate device, driver and host evidence alongside application logs.An Xid number alone does not establish the defective component. Coordinate disruptive tests with the host owner.

See the PyTorch Flight Recorder tutorial, PyTorch memory snapshot documentation and NVIDIA GPU node triage guidance.

Evaluate on an incident you already understand

  1. Preserve all available rank logs, launcher output and host events before a retry overwrites them. Review sensitive information before any optional upload.
  2. Run a local diagnosis. Compare its proposed initiating failure with your confirmed incident and inspect the original supporting lines.
  3. Check whether later NCCL timeouts are separated from an earlier OOM, process exit or device event. If ownership is uncertain, keep it unresolved.
  4. Record what was correct, what was missing and the result of your controlled verification. A sample result is not an accuracy benchmark or evidence of production recovery.

The same views as the Marketplace overview

These captures use the actual extension renderer with synthetic sample records. Connected accounts see their reported inventory; accounts without telemetry see setup instructions.

Console

Denpex Console showing a local vLLM diagnosis with likely cause, evidence and suggested checks.
Inspect the likely cause and the log lines supporting it, without leaving VS Code.

Fleet

Denpex Fleet incident queue with a selected incident and its recorded guidance.
Search incidents and review the selected report beside the queue.

Topology

Denpex Topology showing three sample clusters, 384 GPUs, a red fault and an orange inspection signal.
Find the affected host in your reported fleet and inspect its GPUs.

Device inspector

Selected GPU showing its UUID, PCI address, temperature and reason for inspection.
Keep exact device identity, measurements and the next evidence step together.

Marketplace categories

Machine Learning · Debuggers · Data Science

Tools and workloads

  • gpu
  • cuda
  • pytorch
  • torchrun
  • cuda-oom
  • cuda-error
  • nccl
  • xid
  • xid-errors
  • slurm
  • vllm
  • deepspeed
  • sglang
  • megatron
  • infiniband
  • nvidia-smi
  • hpc
  • h100
  • a100
  • b200
  • multi-gpu
  • distributed-training
  • gpu-memory
  • gpu-debugging
  • log-analyzer
  • nccl-timeout
  • fine-tuning
  • fsdp
  • ddp
  • nvlink

Visible wherever GPU failures happen

Selected traceback

Right-click selected error text and choose Diagnose GPU/ML Error.

Terminal failure

Use Diagnose Last Failure from the terminal context menu.

Log editor

Open a .log or .out file, or traceback content, and use the editor-title action.

Denpex sidebar

Current logs, recent local diagnoses, samples, and privacy status stay one click away.

Local/offline privacy

The free engine is the product, not a metered preview

The deterministic engine and pattern database are bundled in the extension. Local diagnosis is free and unlimited. No API key, sign-in or network is required.

  • Logs remain on the extension host during local diagnosis.
  • Remote-SSH diagnosis runs beside the remote logs.
  • Cloud escalation is visible, contextual, and optional.

Five one-click incidents included

Rank-local OOM

Find rank 3 before the NCCL watchdog errors on every other rank.

Xid 79

Preserve device, PCIe and driver evidence before choosing platform recovery or vendor escalation.

NCCL interface mismatch

Catch peers selecting ib0 and eth0 before changing collective timeouts.

Kubernetes stale GPU health

Separate recovered hardware from stale device-plugin registration.

vLLM KV-cache exhaustion

Fit context length and concurrency to the cache blocks that actually exist.

From first answer to team value

Prove the fix on your own incident before you buy

  1. 1. Find the initiating failure

    Run a bundled sample, then diagnose your own CUDA, NCCL, or Xid log locally. See the first failed rank, evidence, fix, and verification step.

  2. 2. Test the harder case

    Run one immediate cloud deep analysis without an account. Further anonymous analyses may need a browser check. Compare the answer with the cause and action your team confirmed.

  3. 3. Evaluate with your team

    Start a free 30-day Team trial with your work email. Paste the code we email you into VS Code: your account and trial are set up and the editor is connected, with no browser step. Pending review does not activate a trial; once approved, run Denpex: Finish Trial with Emailed Code again. Established company domains can activate automatically across mail providers after verification; uncertain domains go to support review. No API key copying and no automatic log upload. The trial includes up to 50 cloud diagnoses a day and saved history.

Request the 30-day Datacenter trial

No card. No automatic subscription. Local diagnosis stays free.

Purpose-built diagnosis, without giving up your other tools

These tools answer different questions and work well together.

Denpex compared with general AI chat and NVIDIA Nsight
QuestionDenpex extensionGeneral AI chatNVIDIA Nsight
Which rank initiated this cascade?Cross-rank causal orderingDepends on pasted contextProfiling and timeline evidence
Can it run without uploading logs?Yes, bundled offline engineDepends on the deploymentYes, local developer tooling
Does it return an exact operational fix?Fix ladder plus verificationPrompt-dependentEvidence for manual debugging
Best useIncident root cause and remediationExploration and novel reasoningKernel, systems, and performance analysis

Looking for the Open VSX extension release?

Denpex is available on the independent Open VSX registry for VSCodium and other compatible editors. View our dedicated Open VSX installation page.

VS Code extension FAQ

Does the Denpex VS Code extension upload my logs?

Local diagnosis runs in the engine bundled with the extension and makes no network request. Logs are sent only when you explicitly run cloud deep reasoning or enable the opt-in auto-escalation setting.

Does it work over Remote-SSH or without Node.js installed?

Yes. The engine uses the runtime shipped with VS Code, so a separate system Node.js installation is not required. When the extension runs remotely, diagnosis runs where the extension host and logs are located.

What can the free extension diagnose?

The bundled deterministic engine covers CUDA, NCCL, PyTorch DDP and FSDP, Xid and NVLink faults, DeepSpeed, Slurm, Kubernetes GPU components, vLLM, Triton, JAX, InfiniBand and RoCE. Local use is unlimited.

When should I use cloud deep reasoning?

Use it after a low-confidence or novel local result, when multiple causes remain plausible, or when you need saved incident history and team or fleet workflows. The local answer remains available either way.

Diagnose a real failure before deciding what else you need

Install the free extension, run one sample, then try your own logs. Cloud and fleet features stay optional.