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

This commit is contained in:
Jeff Bolz
2026-06-26 16:03:57 +03:00
committed by Georgi Gerganov
parent 3fc2c57a63
commit 792da0ecdb
2 changed files with 2 additions and 0 deletions
@@ -463,6 +463,7 @@ void main() {
}
rowmaxf = max(rowmaxf, float(Sf[r][c]));
}
rowmaxf += FATTN_KQ_MAX_OFFSET;
float Moldf = Mf[r];
// M = max(rowmax, Mold)
@@ -352,6 +352,7 @@ void main() {
}
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];
// Compute max across the row