ml-debug

What problem does it solve?

This Skill systematically diagnoses and resolves failures in ML/AI workflows, such as Out-of-Memory (OOM) errors, NaN values, divergence, crashes, poor throughput, incorrect outputs, and dependency conflicts, by leveraging framework-specific knowledge and grounding in documentation.

Core Features & Use Cases

• Root Cause Analysis: Identifies the underlying cause of ML failures through systematic diagnosis.
• Framework-Specific Debugging: Utilizes knowledge bases and web fetching to provide accurate, context-aware solutions for various ML frameworks (PyTorch, DeepSpeed, vLLM, Hugging Face Transformers, etc.).
• Guided Fixes: Provides step-by-step instructions, including specific configuration changes, code patches, and verification scripts, to resolve identified issues.
• Prevention Strategies: Offers actionable advice and runnable guardrails to prevent similar issues in the future.
• Use Case: When a distributed training job fails with an OOM error on a specific GPU, this Skill can pinpoint whether it's due to activation memory, optimizer states, or KV cache, and provide a precise configuration adjustment to fix it.

Quick Start

Use the ml-debug skill to diagnose and fix an OOM error encountered during LLM fine-tuning.