diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal
index aba0b9a03..79a6ba756 100644
--- a/ggml/src/ggml-metal.metal
+++ b/ggml/src/ggml-metal.metal
@@ -2144,19 +2144,21 @@ kernel void kernel_flash_attn_ext_f16(
                     const short tx = tiisg%4;
                     const short ty = tiisg/4;
 
-                    // mqk = mqk*scale
-                    ss[8*cc + ty*TF + 2*tx + 0] *= scale;
-                    ss[8*cc + ty*TF + 2*tx + 1] *= scale;
+                    if (iq1 + ty < ne01) {
+                        // mqk = mqk*scale
+                        ss[8*cc + ty*TF + 2*tx + 0] *= scale;
+                        ss[8*cc + ty*TF + 2*tx + 1] *= scale;
 
-                    if (logit_softcap != 0.0f) {
-                        ss[8*cc + ty*TF + 2*tx + 0] = logit_softcap*precise::tanh(ss[8*cc + ty*TF + 2*tx + 0]);
-                        ss[8*cc + ty*TF + 2*tx + 1] = logit_softcap*precise::tanh(ss[8*cc + ty*TF + 2*tx + 1]);
-                    }
+                        if (logit_softcap != 0.0f) {
+                            ss[8*cc + ty*TF + 2*tx + 0] = logit_softcap*precise::tanh(ss[8*cc + ty*TF + 2*tx + 0]);
+                            ss[8*cc + ty*TF + 2*tx + 1] = logit_softcap*precise::tanh(ss[8*cc + ty*TF + 2*tx + 1]);
+                        }
 
-                    if (mask != q) {
-                        // mqk = mqk + mask*slope
-                        ss[8*cc + ty*TF + 2*tx + 0] += slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 0];
-                        ss[8*cc + ty*TF + 2*tx + 1] += slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 1];
+                        if (mask != q) {
+                            // mqk = mqk + mask*slope
+                            ss[8*cc + ty*TF + 2*tx + 0] += slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 0];
+                            ss[8*cc + ty*TF + 2*tx + 1] += slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 1];
+                        }
                     }
                 }
             }