Skip to content

Fit FP32 wide-head attention tiles within Metal shared memory #17

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

Draft
Gajesh2007 wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

Fit FP32 wide-head attention tiles within Metal shared memory #17

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions mlx/backend/metal/scaled_dot_product_attention.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -239,6 +239,12 @@ void sdpa_full_self_attention_metal(
int bd = q.shape(-1);
int bq = 32;
int bk = bd < 128 ? 32 : 16;
// Keep FP32 wide-head tiles below Metal's 32 KiB threadgroup limit.
if (q.dtype() == float32 && bd >= 192) {
bq = 16;
bk = bd == 256 ? 8 : 16;
wm = 2;
}

const bool align_Q = (qL % bq) == 0;
const bool align_K = (kL % bk) == 0;
Expand Down