Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

sycl: use TILE for quantized KV decode on BMG (#26689)

Browse Source 
Route quantized KV decode to TILE on Xe2 (BMG) only, keep VEC on other archs until validated there.
This commit is contained in:
Ozymandias_EBON
2026-08-28 11:58:58 +03:00
committed by GitHub
parent be876204aa
commit d077b4c214
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/fattn.cpp
+5 -1
View File
@@ -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) { static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const ggml_tensor * dst) {
GGML_UNUSED(device);
#ifndef SYCL_FLASH_ATTN #ifndef SYCL_FLASH_ATTN
GGML_UNUSED(dst); GGML_UNUSED(dst);
return BEST_FATTN_KERNEL_NONE; return BEST_FATTN_KERNEL_NONE;
@@ -263,6 +262,11 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
} }
} else { } else {
if (Q->ne[1] <= 2) { 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; return BEST_FATTN_KERNEL_VEC;
} }
} }