diff --git a/unsloth/kernels/cross_entropy_loss.py b/unsloth/kernels/cross_entropy_loss.py index 35df257329..4895d27310 100644 --- a/unsloth/kernels/cross_entropy_loss.py +++ b/unsloth/kernels/cross_entropy_loss.py @@ -373,9 +373,9 @@ def fast_cross_entropy_loss( logit_softcapping, logit_scaling, ) - print(n_items) if n_items is None: n_items = torch.count_nonzero(labels != -100) + print(n_items) return loss.sum() / n_items pass