From d077b4c21466cfad678b07b05b557599f4db3974 Mon Sep 17 00:00:00 2001
From: Ozymandias_EBON <112784549+johnkarlhill@users.noreply.github.com>
Date: Fri, 28 Aug 2026 03:58:58 -0500
Subject: [PATCH] sycl: use TILE for quantized KV decode on BMG (#26689)

Route quantized KV decode to TILE on Xe2 (BMG) only, keep VEC on other archs until validated there.
---
 ggml/src/ggml-sycl/fattn.cpp | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-sycl/fattn.cpp b/ggml/src/ggml-sycl/fattn.cpp
index a85eb721f..a85bca7cb 100644
--- a/ggml/src/ggml-sycl/fattn.cpp
+++ b/ggml/src/ggml-sycl/fattn.cpp
@@ -104,7 +104,6 @@ enum best_fattn_kernel {
 
 
 static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const ggml_tensor * dst) {
-    GGML_UNUSED(device);
 #ifndef SYCL_FLASH_ATTN
     GGML_UNUSED(dst);
     return BEST_FATTN_KERNEL_NONE;
@@ -263,6 +262,11 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
             }
         } else {
             if (Q->ne[1] <= 2) {
+                // TILE is faster for quantized KV decode on Xe2 (BMG); keep VEC on untested archs
+                const gpu_arch arch = ggml_sycl_info().devices[device].hw_info.arch;
+                if (arch == gpu_arch::intel_gpu_bmg_g21 || arch == gpu_arch::intel_gpu_bmg_g31) {
+                    return BEST_FATTN_KERNEL_TILE;
+                }
                 return BEST_FATTN_KERNEL_VEC;
             }
         }