New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

change resampler attention to scale_dot_product_attention #446

Open

Benkovichnikita wants to merge 1 commit into tencent-ailab:main from Benkovichnikita:feat/attention_optimization

Benkovichnikita commented Nov 19, 2024 •

edited

Loading

I noticed two issues:

PerceiverAttention consumes a huge amount of memory while calculating an attention map
There are two operations of math.srqt when scale calculating: scale = 1 / math.sqrt(math.sqrt(self.dim_head))

Both of them can be fixed using more effective implementation: F.scaled_dot_product_attention(q, k, v).

PS
If you need scale with two sqrt, it can be provided as an additional argument to F.scaled_dot_product_attention.


          change resampler attention to scale_dot_product_attention

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet