Loss = nan why?

[zhangqianYY]

hi, I run (image/vision_transformer/data_parallel/train_with_cifar10.py) colossalai run --nproc_per_node 1 train_with_cifar10.py --config config.py ,when BATCH_SIZE = 2048,gradient_accumulation=240 , lead to Loss = nan,How can it optimize?

[BoxiangW]

Hi, you could refer to discussion #1116.