Skip to content

vulkan: use aligned loads for flash attention mask #12853

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 12, 2025
+7 −4
Merged

vulkan: use aligned loads for flash attention mask #12853

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
11 changes: 7 additions & 4 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 @@ -201,6 +201,11 @@ void main() {
uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
uint32_t k_stride = p.nb11;
uint32_t v_stride = p.nb21;
// When using grouped query attention, all rows use the same mask (stride 0).
// "p.gqa_ratio >> 16" is just a roundabout way of writing zero
// that prevents the compiler from folding the "&" through the select
// and breaking the alignment detection.
uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
// hint to the compiler that strides are aligned for the aligned variant of the shader
if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
{
Expand All @@ -209,6 +214,7 @@ void main() {
k_stride &= ~7;
v_stride &= ~7;
#endif
m_stride &= ~7;
}
tensorLayoutQ = setTensorLayoutStrideNV(tensorLayoutQ, q_stride, 1);
tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
Expand Down Expand Up @@ -261,10 +267,7 @@ void main() {
if (p.mask != 0) {
tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
// When using grouped query attention, all rows use the same mask.
if (p.gqa_ratio > 1) {
tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, 0, 1);
}
tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);

coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;

Expand Down
Loading