Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention - View it on GitHub