torch:clip-grad-norm

(torch:clip-grad-norm params max-norm)

Gradient-norm clipping (PyTorch's torch.nn.utils.clip_grad_norm_) over params: a module, an optimizer's parameter list, or a plain list of tensors.

Returns the TOTAL L2 norm of every gradient, taken over all of them at once as if they were one long vector -- the norm as MEASURED, before any clipping, so a training loop can log it. When that norm exceeds max-norm, every gradient is scaled IN PLACE by max-norm / (norm + 1e-6), PyTorch's denominator; otherwise nothing is touched. A parameter no gradient reached is skipped.

Call it between torch:backward and torch:step: it rewrites the gradients the optimizer is about to read, and touches no tape.