CrashLens: turn GPU job logs into a focused local investigation

ThermaCompute AI is focused on helping GPU teams understand how heat, power and workload behavior affect reliable, useful compute. Our current public tool is CrashLens: a free, local starting point for investigating saved GPU job logs.

A failed GPU job should give you a starting point

A training or inference job fails. You have pages of logs, a worker exception and perhaps a collective timeout. Which message deserves your attention first?

So we made CrashLens to turn those logs into a focused investigation. It scans for known signatures, attaches source line numbers and suggests practical checks. GPU allocation failures, host OOM events, collective timeouts, worker exits, NVIDIA Xid messages and explicit thermal/NVLink messages are kept distinct. A timeout is not proof that the reporting rank caused the failure.

CrashLens runs offline with Python's standard library. No account, API key, GPU or production SSH access is required. The CLI makes no network requests and changes no GPU settings. Your input logs and reports stay local. Review reports before sharing: secret masking is best effort and incomplete.

Try the synthetic demo

You need Python 3.10 or newer:

git clone https://github.com/rohitcn-hub/thermacompute-crashlens.git
cd thermacompute-crashlens
python crashlens.py examples/synthetic.log --out demo-report

Open demo-report/report.html locally. The directory also contains Markdown and JSON reports. Use a new output directory for each run.

Try CrashLens on GitHub

This is experimental version 0.2.0, tested on synthetic fixtures rather than validated across production fleets. It recognizes patterns, not confirmed root causes, and does not automatically fix jobs. No matches means insufficient evidence, not a healthy system.

Did it give you a useful next check? What was missing? Tell us in the comments or first-run feedback issue. Please use invented examples, not private production logs.

Optional: a deeper $80 thermal/GPU efficiency audit

CrashLens helps organize crash-log evidence. ThermaCompute's $80 audit separately reviews suitable temperature, power and reported-throttling telemetry, delivering supported findings, prioritized checks and a PDF with assumptions and limitations.

To upgrade, email [e-mail] with Thermal Audit in the subject, your GPU model/count, the problem you are investigating and your CSV headers. We confirm data suitability and scope before payment. The audit does not guarantee savings or include production changes; it can help define whether a larger engagement makes sense. CrashLens remains free, and the audit does not commit you to a subscription.

I'm building ThermaCompute and would value honest technical feedback.