Skip to content

[None][feat] Fix attention sink load in xqa #8836

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
Nov 3, 2025
+7 −3
Merged

[None][feat] Fix attention sink load in xqa #8836

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
10 changes: 7 additions & 3 deletions cpp/kernels/xqa/mha_sm90.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2078,9 +2078,13 @@ __device__ inline RegColWiseVec loadGmemColWiseVecWithDup(ShmQWiseVec const& gme
for (uint32_t i = 0; i < exactDiv(ShmQWiseVec::size, gmma::instNBase); i++)
{
static_assert(nbThrdsPerInstNBase * RegColWiseVec::size == exactDiv(ShmQWiseVec::size, GmmaAccCoreMat::cols));
ret[i] = reinterpret_cast<
Vec<Vec<float, GmmaAccCoreMat::cols>, exactDiv(ShmQWiseVec::size, GmmaAccCoreMat::cols)> const&>(
gmemVec)[mha::min(i * nbThrdsPerInstNBase + idx, bound)];
uint32_t const clampedIdx = mha::min(i * nbThrdsPerInstNBase + idx, bound);
uint32_t const baseOffset = clampedIdx * GmmaAccCoreMat::cols;
#pragma unroll
for (uint32_t j = 0; j < GmmaAccCoreMat::cols; j++)
{
ret[i][j] = gmemVec[baseOffset + j];
}
}
return ret;
}
Expand Down
Loading