Skip to main content

NaN Forensics

A single NaN doesn't stay put — it spreads to every number it touches.

Key Insight​

A NaN ("Not a Number") poisons every later calculation, so by the time training visibly breaks the real cause is many steps behind. Turning on anomaly detection makes autograd stop at the exact operation that first produced the NaN.

Why This Matters​

Chasing the symptom — a loss that suddenly reads NaN — wastes hours. Finding the first bad op tells you which layer and which math (a log(0), a divide-by-zero, an exploding gradient) to fix.