Diagnose ML training issues from symptom to root cause across GPU setups.
npx skills add https://github.com/Gonglitian/agent-skills --skill train-debug
Please help me install this Agent Skill. Skill: train-debug Source: https://github.com/Gonglitian/agent-skills/tree/main/skills/train-debug Command: npx skills add https://github.com/Gonglitian/agent-skills --skill train-debug
ML training often encounters crashes, non-converging losses, NaN/Inf gradients, OOM, and slowdowns that block progress. This skill provides a structured diagnostic workflow to identify root causes and guide efficient fixes.
Describe the observed issue, run triage steps to identify root cause, and apply a safe, incremental fix while verifying results.
High-intent search queries and answers about installing and using this skill.
Debug ML training crashes and NaN gradients by following a structured triage workflow that checks logs, assesses GPU state, and maps symptoms to root causes for safe, incremental fixes. This ensures minimal changes while verifying results.
Diagnose out of memory errors in multi-GPU training by assessing GPU state and applying category-specific diagnosis steps. This workflow identifies root causes for memory errors and provides concrete fixes with verification steps to restore training.
ML training slowdowns across multiple GPUs are diagnosed through a structured workflow that checks GPU state and logs to identify bottlenecks. This triage process maps the performance symptoms to root causes and recommends safe incremental optimizations.
Verify ML training fixes and ensure reproducible results by applying minimal changes and validating them in production-like pipelines. The diagnostic workflow enforces step-by-step verification to confirm the resolved issue produces stable outcomes.
Yes, you can use a diagnostic workflow to fix non-converging losses in ML training. It guides you through structured triage steps from symptom to root cause, enforcing safe log checks and incremental fixes to restore convergence.
JunMA98Detect CPU, memory, disk, and GPU metrics and generate a JSON report.
porcupine-mdDiagnose and fix flaky tests caused by timing dependencies, test pollution, and async leaks.
backengineeringInstrument binaries and trace runtime behavior using LLVM infrastructure.
UlanxxMonitor live URLs for console errors, performance regressions, and page failures.
