Skip to content

Support group_size=64 for affine quantized SDPA #2

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

Merged
CC-Yeh merged 1 commit into from
Apr 18, 2026
Merged

Support group_size=64 for affine quantized SDPA #2

Show file tree
Hide file tree
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
3 changes: 3 additions & 0 deletions mlx/backend/metal/kernels/sdpa_vector.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -653,6 +653,9 @@ template <typename T, int D>
QUANT_SDPA_DISPATCH(Affine, 32, 4)
QUANT_SDPA_DISPATCH(Affine, 32, 6)
QUANT_SDPA_DISPATCH(Affine, 32, 8)
QUANT_SDPA_DISPATCH(Affine, 64, 4)
QUANT_SDPA_DISPATCH(Affine, 64, 6)
QUANT_SDPA_DISPATCH(Affine, 64, 8)
QUANT_SDPA_DISPATCH(Mxfp4, 32, 4)
QUANT_SDPA_DISPATCH(Nvfp4, 16, 4)
QUANT_SDPA_DISPATCH(Mxfp8, 32, 8)
Expand Down
4 changes: 2 additions & 2 deletions mlx/fast.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -901,9 +901,9 @@ array quantized_scaled_dot_product_attention(

// Validate mode-specific group_size and bits
if (is_affine) {
if (group_size != 32) {
if (group_size != 32 && group_size != 64) {
std::ostringstream msg;
msg << "[" << tag << "] Affine mode supports group_size 32 "
msg << "[" << tag << "] Affine mode supports group_size 32 or 64 "
<< "but received " << group_size << ".";
throw std::invalid_argument(msg.str());
}
Expand Down