jdelony/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-06-27 23:50:20 -05:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Apply bias before softmax in FA, to avoid overflow (#24909)

Browse Source
This commit is contained in:
Jeff Bolz 2026-06-23 22:34:00 -05:00 committed by GitHub
parent be4a6a63eb
commit ac4105d68b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 2 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
View File

@ -463,6 +463,7 @@ void main() {
} }
rowmaxf = max(rowmaxf, float(Sf[r][c])); rowmaxf = max(rowmaxf, float(Sf[r][c]));
} }
rowmaxf += FATTN_KQ_MAX_OFFSET;
float Moldf = Mf[r]; float Moldf = Mf[r];
// M = max(rowmax, Mold) // M = max(rowmax, Mold)

1
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
View File

@ -352,6 +352,7 @@ void main() {
} }
rowmaxf = max(rowmaxf, float(sfsh[r_vec + (c * cols_per_iter + col_tid) * sfshstride][r_comp])); rowmaxf = max(rowmaxf, float(sfsh[r_vec + (c * cols_per_iter + col_tid) * sfshstride][r_comp]));
} }
rowmaxf += FATTN_KQ_MAX_OFFSET;
float Moldf = Mf[r]; float Moldf = Mf[r];
// Compute max across the row // Compute max across the row