thu-coai/CDial-GPT

启用半精度后，训练loss = nan

WuDiDaBinGe opened this issue · 0 comments

WuDiDaBinGe commented

Iter (loss= nan) lr=0.0001875: 0%| | 5/60000 [00:09<17:37:55, 1.06s/it]08/03/2022 14:51:06