Is there any comparison of the effects related to token attention? For example, compare with page attention

[skykiseki]

If there's a paper or other proof, that would be even better.

[skykiseki]

Not only the performance of throughput or the speed of single inference, what I'm referring to is the content of the model inference effect, such as accuracy and other metrics.

[hiworldwzj]

@skykiseki The paper is being written. and we have some kernel write by cuda c that is faster than page attention in some cases.