Skip to content

Enhance CUDA flash attention kernel selection for DKQ=512 with low gq… #6

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

Open
Ooooze wants to merge 1 commit into
base: feature/turboquant-kv-cache
Choose a base branch
from
+8 −0
Open

Enhance CUDA flash attention kernel selection for DKQ=512 with low gq… #6

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
8 changes: 8 additions & 0 deletions ggml/src/ggml-cuda/fattn.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -519,6 +519,14 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
return BEST_FATTN_KERNEL_VEC;
}
}
// MMA template instances for (DKQ=512, DV=512) only exist with ncols2 in {4, 8}
// (see template-instances/generate_cu_files.py). When gqa_ratio < 3 the MMA
// dispatcher in switch_ncols2<512,512> falls through to GGML_ABORT (fattn.cu:109).
// fattn-tile has DKQ=DV=512 with ncols2 fallback to {2,1} (commit 425db5b),
// so route DKQ=512 + low gqa_ratio to TILE (e.g. Gemma 4 E4B with gqa_ratio=2).
if (Q->ne[0] == 512 && gqa_ratio < 3) {
return BEST_FATTN_KERNEL_TILE;
}
return BEST_FATTN_KERNEL_MMA_F16;
}

Expand Down
Loading