Skip to content

vulkan: fix NaN issue in flash attention shader #12776

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 6, 2025
Merged

vulkan: fix NaN issue in flash attention shader #12776

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
7 changes: 5 additions & 2 deletions ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -227,8 +227,11 @@ void main() {

coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;

// Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M.
const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);

L = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(-1.0/0.0);
M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(NEG_FLT_MAX_OVER_2);

coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> slopeMat = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(1.0);

Expand Down Expand Up @@ -278,7 +281,7 @@ void main() {
uint R = ((i + 1) * Br > N) ? (N % Br) : Br;
uint C = ((j + 1) * Bc > KV) ? (KV % Bc) : Bc;

coopMatPerElementNV(S, S, replacePadding, ACC_TYPE(-1.0/0.0), R, C);
coopMatPerElementNV(S, S, replacePadding, ACC_TYPE(NEG_FLT_MAX_OVER_2), R, C);
}

coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> rowmax, P, rowsum, eM;
Expand Down
Loading