Skip to content

Fix CUDA dynamic mask attention scaling to match Python implementation #16

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 2 commits into from
May 22, 2025
Merged

Fix CUDA dynamic mask attention scaling to match Python implementation #16

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
55af05a
Initial plan for issue
Copilot May 22, 2025
882cf89
Fix attention score calculation in flash_attention_fwd_kernel.h
Copilot May 22, 2025
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
12 changes: 8 additions & 4 deletions csrc/src/flash_attention_fwd_kernel.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -454,8 +454,10 @@ inline __device__ void compute_attn_1rowblock(const Params &params, const int bi
auto mask_values_row = sDynamicMaskValues(m_idx, _);
auto predicate_k_row = sPredicate(m_idx, _);
if (predicate_k_row(k_idx)) {
// Scale the attention score before adding mask value, matching Python's behavior
acc_s(mma, mi, ki) = acc_s(mma, mi, ki) * params.scale_softmax + static_cast<ElementAccum>(mask_values_row(k_idx));
// First scale the attention score
ElementAccum scaled_score = acc_s(mma, mi, ki) * params.scale_softmax;
// Then add the mask value, matching Python's behavior
acc_s(mma, mi, ki) = scaled_score + static_cast<ElementAccum>(mask_values_row(k_idx));
} else {
// For positions where mask is 0, set attention score to -INFINITY so they don't contribute to softmax
acc_s(mma, mi, ki) = -INFINITY;
Expand Down Expand Up @@ -571,8 +573,10 @@ inline __device__ void compute_attn_1rowblock(const Params &params, const int bi
auto mask_values_row = sDynamicMaskValues(m_idx, _);
auto predicate_k_row = sPredicate(m_idx, _);
if (predicate_k_row(k_idx)) {
// Scale the attention score before adding mask value, matching Python's behavior
acc_s(mma, mi, ki) = acc_s(mma, mi, ki) * params.scale_softmax + static_cast<ElementAccum>(mask_values_row(k_idx));
// First scale the attention score
ElementAccum scaled_score = acc_s(mma, mi, ki) * params.scale_softmax;
// Then add the mask value, matching Python's behavior
acc_s(mma, mi, ki) = scaled_score + static_cast<ElementAccum>(mask_values_row(k_idx));
} else {
// For positions where mask is 0, set attention score to -INFINITY so they don't contribute to softmax
acc_s(mma, mi, ki) = -INFINITY;
Expand Down
Loading