Skip to content

CUDA: fix FTZ in FA for Gemma 3 #13991

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
Jun 4, 2025
Merged

CUDA: fix FTZ in FA for Gemma 3 #13991

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
5 changes: 4 additions & 1 deletion ggml/src/ggml-cuda/fattn-mma-f16.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -652,9 +652,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
float KQ_max_scale[cols_per_thread];
#pragma unroll
for (int col = 0; col < cols_per_thread; ++col) {
KQ_max_scale[col] = expf(KQ_max[col] - KQ_max_new[col]);
const float KQ_max_diff = KQ_max[col] - KQ_max_new[col];
KQ_max_scale[col] = expf(KQ_max_diff);
KQ_max[col] = KQ_max_new[col];

*((uint32_t *) &KQ_max_scale[col]) *= KQ_max_diff >= SOFTMAX_FTZ_THRESHOLD;

// Scale previous KQ_rowsum to account for a potential increase in KQ_max:
KQ_rowsum[col] = KQ_max_scale[col]*KQ_rowsum[col] + KQ_rowsum_add[col];
}
Expand Down
Loading