Commit e01c9f187 for llama.cpp

commit e01c9f1875e8a96674756ed144aa9701498b83dd
Author: CC <65151926+realMoana@users.noreply.github.com>
Date:   Mon Nov 9 06:32:27 2026 -0500

    metal : prevent out-of-bounds writes to the Metal FA output buffer (#31662)

    Co-authored-by: Zhuomin Chen <moanachen@Mac.attlocal.net>

diff --git a/ggml/src/ggml-metal/kernels/fa_vec_common.metal b/ggml/src/ggml-metal/kernels/fa_vec_common.metal
index c81e4d09b..1a96c6f7e 100644
--- a/ggml/src/ggml-metal/kernels/fa_vec_common.metal
+++ b/ggml/src/ggml-metal/kernels/fa_vec_common.metal
@@ -131,9 +131,11 @@ kernel void kernel_flash_attn_ext_vec(
     }

     // zero out so
-    for (short qq = 0; qq < Q; ++qq) {
-        for (short i = 0; i < DV4/NL; ++i) {
-            so4[qq*DV4 + i*NL] = (o4_t) 0.0f;
+    if ((DV4/NL % NW == 0) || tiisg < NL) {
+        for (short qq = 0; qq < Q; ++qq) {
+            for (short i = 0; i < DV4/NL; ++i) {
+                so4[qq*DV4 + i*NL] = (o4_t) 0.0f;
+            }
         }
     }