chunked/content_aware_chunking/_deepspeed/chunk_118.txt

If your model doesn't work well with mixed precision, for example if it wasn't pretrained in mixed precision, you may encounter overflow or underflow issues which can cause NaN loss.