CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-03 09:22:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

sycl: reordered Q4_K MMVQ (#13109)

Browse Source
This commit is contained in:
Svetlozar Georgiev
2025-05-15 16:35:44 +01:00
committed by GitHub
parent 9c404ed54c
commit 64bb51cf90
7 changed files with 280 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-sycl/dmmv.cpp
View File

@@ -1129,7 +1129,13 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
dequantize_mul_mat_vec_q3_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q4_K:
dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
// reorder is currently not supported for dmmv
GGML_ABORT("Unimplemented dequantize case case for q4_k reorder");
} else {
dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
}
break;
case GGML_TYPE_Q5_K:
dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);