Skip to content

cuda: disable sparse V skip (warp divergence regression) #105

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
TheTom merged 1 commit into from
Apr 24, 2026
Merged

cuda: disable sparse V skip (warp divergence regression) #105

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: 10 additions & 0 deletions ggml/src/ggml-cuda/fattn-vec.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -333,6 +333,10 @@ static __global__ void flash_attn_ext_vec(
}

// Sparse V: skip V dequant if all attention weights for this position are negligible
// Disabled — per-lane branching causes warp divergence that costs more than the
// skipped dequants save (-0.3% to -2.8% on RTX 3090/4090).
// TODO: revisit with warp-level ballot skip.
#if 0
{
bool dominated = true;
#pragma unroll
Expand All @@ -341,6 +345,7 @@ static __global__ void flash_attn_ext_vec(
}
if (dominated) { continue; }
}
#endif

#pragma unroll
for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V*V_rows_per_thread/2) {
Expand Down Expand Up @@ -373,6 +378,10 @@ static __global__ void flash_attn_ext_vec(
}

// Sparse V: skip V dequant if all attention weights for this position are negligible
// Disabled — per-lane branching causes warp divergence that costs more than the
// skipped dequants save (-0.3% to -2.8% on RTX 3090/4090).
// TODO: revisit with warp-level ballot skip.
#if 0
{
bool dominated = true;
#pragma unroll
Expand All @@ -381,6 +390,7 @@ static __global__ void flash_attn_ext_vec(
}
if (dominated) { continue; }
}
#endif

#pragma unroll
for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V*V_rows_per_thread/2) {
Expand Down
Loading