Run NVIDIA compute-sanitizer (memcheck / racecheck / initcheck / synccheck). FIRST action on `INCORRECT_NUMERICAL` or flaky output — rolling back destroys the evidence; sanitizer often points at the exact line. Also useful for diagnosing race conditions in multi-block kernels and uninitialized-memory reads.
Instalación
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Run NVIDIA compute-sanitizer (memcheck / racecheck / initcheck / synccheck). FIRST action on `INCORRECT_NUMERICAL` or flaky output — rolling back destroys the evidence; sanitizer often points at the exact line. Also useful for diagnosing race conditions in multi-block kernels and uninitialized-memory reads.
Sanitizer
Wrapper around compute-sanitizer. Command entry: bash scripts/sanitize.sh.
Run sanitizer FIRST on INCORRECT_NUMERICAL — before any rollback. Rolling back overwrites the failing kernel and destroys the only evidence of which line went wrong.
cuGetProcAddress_v2 noise inflates ERROR SUMMARY — these are PyTorch / triton / cupti probing optional CUDA driver entry points at import time. The wrapper's >>> NOISE FILTER <<< banner subtracts them; trust the "apparently-real kernel hits" count.
Real kernel errors always name a kernel / thread / block / address. If a hit only shows Python host frames, it's noise.
CUDA-graph-captured kernels show up but with sparse stack info; if traces are unhelpful, temporarily disable graph capture (env var or skip g.replay() path) for an eager-mode report, fix, then re-enable.
