[Feature] Should we remove other dependency of flashattention?

[sunpengsdu]

Describe the feature

Should we remove other dependency of flash-attention, and only keep the core attention related ops?

If possible, we can only use pip to install flash-attention, avoiding a lot of compiling operations.

To seek whether it is possible, we need to check whether it would reduce the training performance a lot.

Will you implement it?

• [ ] I would like to implement this feature and create a PR!