From 0409ed2628a370ffb269521ef47fc23ac5fe863f Mon Sep 17 00:00:00 2001
From: "David M. Rogers" <predictivestatmech@gmail.com>
Date: Tue, 22 Sep 2026 09:43:29 -0400
Subject: [PATCH] Performance tune for gemma4-26b-a4b flash attention shape.
 (llama/28450)

---
 ggml/src/ggml-sycl/fattn-tile.hpp | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/ggml/src/ggml-sycl/fattn-tile.hpp b/ggml/src/ggml-sycl/fattn-tile.hpp
index 9ba529696..dcdcad88a 100644
--- a/ggml/src/ggml-sycl/fattn-tile.hpp
+++ b/ggml/src/ggml-sycl/fattn-tile.hpp
@@ -1173,6 +1173,10 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_sycl_context & ctx, ggm
             launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
             return;
         }
+        if (use_gqa_opt && gqa_ratio % 8 == 0) {
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
+            return;
+        }
         if (use_gqa_opt && gqa_ratio % 4 == 0) {
             launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
             return;