About FlashAttention

[ZekaiGalaxy]

Thank you for your great work :) !

Here is my question: I tried to follow the instructions but failed in the flash-attention related steps. According to this issue, V100 are not supported.

So I wonder the efficiency gain without the flash-attention module, or are there any methods to surpass the above issue and achieve comparable performance on V100s?

Thank you!

[oahzxl]

memory efficient attention from xformers may be a good choice

[bhack]

Can I ask you why you are still using the external flash attention?
torch.nn.functional.scaled_dot_product_attention has already a flashattention-2 implementation:
https://pytorch.org/docs/2.2/generated/torch.nn.functional.scaled_dot_product_attention.html

[oahzxl]

its implemtation sometimes is slower than flashattn on device < H100

[bhack]

Probably we could monitor this
pytorch/pytorch#120642

[oahzxl]

integrate into torch now!