Skip to content

metal : fix floating-point range of attention scores in FA kernels #13090

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
merged 1 commit into from
Apr 24, 2025
Merged

metal : fix floating-point range of attention scores in FA kernels #13090

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: 3 additions & 3 deletions ggml/src/ggml-metal/ggml-metal.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3192,7 +3192,7 @@ kernel void kernel_flash_attn_ext(

{
float S[Q] = { [0 ... Q-1] = 0.0f };
float M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 };
float M[Q] = { [0 ... Q-1] = -__FLT_MAX__/2 };

// thread indices inside the simdgroup
// TODO: see if we can utilize quad-group functions for better performance
Expand Down Expand Up @@ -3452,7 +3452,7 @@ kernel void kernel_flash_attn_ext(
// reduce the warps sequentially
for (ushort sg = 1; sg < nsg; ++sg) {
float S = { 0.0f };
float M = { -__FLT16_MAX__/2 };
float M = { -__FLT_MAX__/2 };

threadgroup_barrier(mem_flags::mem_threadgroup);

Expand Down Expand Up @@ -3699,7 +3699,7 @@ kernel void kernel_flash_attn_ext_vec(

{
float S = 0.0f;
float M = -__FLT16_MAX__/2;
float M = -__FLT_MAX__/2;

// thread indices inside the simdgroup
const short tx = tiisg%NL;
Expand Down
Loading