GPU Doctor

Get your GPU working.

Check your setup, understand a failure, and know what to try next.

Check a workload

Workload

What are you checking? Before launch Running GPU / failed job

VRAM per GPU (GB)Estimated peak VRAM needed (GB)

Use peak workload memory, including batches and context—not just model weights. Multi-GPU memory is not automatically pooled.

Disk space needed (GB)Persistent output storage

Check a running GPU

In your GPU's SSH terminal, download and inspect the probe, then run it using your workload's Python environment from its data directory.

Download diagnostic probe

python doctor_probe.py

The probe checks packages, free disk space, GPU access, a tiny GPU operation, and default localhost ports. It installs nothing and sends no data. Port checks do not verify application health.

Paste the JSON report

Explain an error log

Relevant error lines

When you check, the report and error lines are sent to TrackMyGPU for analysis. Doctor doesn't save them. Known credential patterns are removed, but review your input for other sensitive data.

Your diagnosis

＋

A clearer next step.

Enter what you know. Doctor will separate verified checks, likely issues, and information still needed.

Doctor suggests fixes. It never installs packages, changes files, restarts a GPU, or launches compute.