After six hours training a ResNet on a medical‑imaging dataset, loss dropped to NaN without crashing. The author enabled torch.autograd.set_detect_anomaly, which throttled the run tenfold and only highlighted a downstream symptom. Later analysis traced the issue to a learning‑rate scheduler clashing with a custom normalization layer, causing gradient explosion. A lightweight forward‑hook detector now pinpoints the exact layer and batch where NaNs emerge in production effectively.

The detector registers a forward hook on each nn.Module, runs torch.isnan() and torch.isinf() on the output tensor, and logs a NaNEvent with batch index, layer name, and stats. Overhead averages 3‑4 ms per pass, about five times faster than set_detect_anomaly. Hooks lock only when mutating shared state, and an overhead buffer capped at 1,000 entries prevents memory bloat.

A secondary guard inspects gradient norms after loss.backward(); if a norm exceeds a threshold it records a GradEvent, catching exploding gradients one step before NaNs appear. The check runs before the next forward pass, allowing a quick learning‑rate tweak. The open‑source implementation is thread‑safe, caps memory, and lives on GitHub, giving engineers a fast way to debug silent NaN failures.