RuntimeError: Magnitude of gradient is bad: inf

This happens occasionally (once every 20~ runs), stopping my training process:

> RuntimeError: Magnitude of gradient is bad: inf

I have not yet managed to replicate this reliably.

Start by investigating the reported “Magnitude of gradient is bad: inf” RuntimeError in the training path. The report provides no file, test, or reliable reproduction, so first try to capture the failure and identify the relevant gradient-validation entry point. Done would mean reproducing the failure reliably and establishing a verified fix or clear diagnosis.

Written by the indexing model from the issue text.