Global gradients clipping should be used in stead of clipping each matrix separately.

I found your code measuring norms and clipping gradients for each parameter separately. But in the paper, the authors said "the l2 norm of the whole gradient of all parameters..." and your method was used in QA tasks.

Locate the code that measures gradient norms and clips gradients for each parameter separately. Compare that behavior with the paper's requirement to use the L2 norm across all parameters, then verify that the implementation applies global clipping and preserves the expected QA-task behavior.

Written by the indexing model from the issue text.