Why does the 'loss=nan' appear when finetuning model using fp16 or bf16 mixed precision in 'fine_tune.py' ?

Same as title. This problem has existed since the v9 version. It is related to the GPU type. It appears in 3090, A40, but doesn't appear in A100.

Start with fine_tune.py and reproduce the reported loss=nan behavior using fp16 and bf16 mixed precision on a 3090 or A40, then compare it with an A100. Determine the conditions that distinguish the affected GPU types and document a clear fix or limitation, with a reproducible test case.

Written by the indexing model from the issue text.