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

sycl : support MUL_MAT and OUT_PROD with Q1_0 (#24721)

Browse Source
This commit is contained in:
Neo Zhang
2026-06-18 11:17:37 +03:00
committed by GitHub
parent 6ec59ddaea
commit dd69db2924
7 changed files with 247 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/convert.cpp
+6
View File
@@ -642,6 +642,8 @@ static void convert_unary_sycl(const void * vx, dst_t * y, const int64_t k, dpct
to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
switch (type) {
case GGML_TYPE_Q1_0:
return dequantize_block_sycl<QK1_0, QR1_0, dequantize_q1_0>;
case GGML_TYPE_Q4_0:
if (dst->src[0]->extra &&
((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
@@ -724,6 +726,8 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
switch (type) {
case GGML_TYPE_Q1_0:
return dequantize_block_sycl<QK1_0, QR1_0, dequantize_q1_0>;
case GGML_TYPE_Q4_0:
if (dst->src[0]->extra &&
((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
@@ -830,6 +834,8 @@ to_fp16_nc_sycl_t ggml_get_to_fp16_nc_sycl(ggml_type type) {
case GGML_TYPE_BF16:
return convert_unary_nc_sycl<sycl::ext::oneapi::bfloat16>;
#endif
case GGML_TYPE_Q1_0:
return dequantize_block_nc_sycl<QK1_0, QR1_0, dequantize_q1_0>;
case GGML_TYPE_Q4_0:
return dequantize_block_nc_sycl<QK4_0, QR4_0, dequantize_q4_0>;
case GGML_TYPE_Q4_1: