Why we can remove the softmax?

[HaoWuSR]

Hi,
Thanks for your excellent work!
I have read your paper carefully but still confused about the following issues:

1. I understand the normalization in Q and K for balancing the weights. But why we can remove the SOFTMAX, the description in the paper is: Given this simple normalization method, we remove the SOFTMAX layer, so the attention block can be ...... But what is the reason why we can remove the SOFTMAX?

2. The paper said the attention values are bounded between -D and D due to the Normalization. What is the difference between this format and the traditional format?

Thanks for your time and waiting for your reply.