CUDA: fix broken oob check for FA vec f32 kernel

ggerganov · Jun 12, 2024 · 60c373d · 60c373d
1 parent a9cae48
commit 60c373d
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/ggml-cuda/fattn-vec-f32.cuh b/ggml-cuda/fattn-vec-f32.cuh
@@ -149,7 +149,7 @@ static __global__ void flash_attn_vec_ext_f32(
             for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
                 const int i = i0 + threadIdx.x;
 
-                Q_f2[j][i0/WARP_SIZE]    = ncols <= 2 || ic0 + j ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
+                Q_f2[j][i0/WARP_SIZE]    = ncols <= 2 || ic0 + j < ne01 ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
                 Q_f2[j][i0/WARP_SIZE].x *= scale;
                 Q_f2[j][i0/WARP_SIZE].y *= scale;
             }