What problem does it solve?
This Skill helps diagnose difficult CUDA crashes in SGLang by capturing the tensor inputs, metadata, and execution context at kernel API boundaries before failures prevent normal debugging.
Core Features & Use Cases
- Crash-Safe Kernel Logging: Configure SGLang kernel API logging from function names through detailed tensor statistics and pre-execution dumps.
- LLM and Diffusion Coverage: Investigate custom operations, attention, linear, quantization, rotary, and model-specific kernel paths across SGLang workloads.
- Failure Isolation: Correlate shapes, dtypes, devices, strides, NaN or Inf values, multi-process logs, and sanitizer output to identify illegal memory access, device-side assert, out-of-memory, and numerical failures.
- Use Case: When a multi-GPU SGLang inference job crashes with an illegal memory access, capture per-process level-10 inputs, identify the final failing API boundary, and compare its tensor metadata with compute-sanitizer results.
Quick Start
Use the debug-cuda-crash skill to reproduce the SGLang CUDA failure with level-3 logging first, then escalate to level-10 crash-safe input dumps if the failing kernel remains unclear.
