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

sycl : enhance set_rows to support q1_0, mxfp4, nvfp4 (#24564)

Browse Source
This commit is contained in:
Neo Zhang
2026-06-15 10:01:40 +03:00
committed by GitHub
parent 987fbd821d
commit 8872ab5467
4 changed files with 334 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/set_rows.cpp
+10 -2
View File
@@ -135,7 +135,7 @@ static void set_rows_sycl(
stream->parallel_for(
sycl::nd_range<1>(grid_size * block_size, block_size),
[=](sycl::nd_item<1> item_ct1) {
[=](sycl::nd_item<1> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
k_set_rows<TIn, TIdx, TOut>(
src0_d, src1_d, dst_d,
ne00, ne01, ne02,
@@ -202,6 +202,9 @@ static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * s
case GGML_TYPE_Q8_0:
set_rows_sycl_q<TIdx, block_q8_0, QK8_0, cpy_blck_f32_q8_0>(src0_d, src1_d, (block_q8_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
break;
case GGML_TYPE_Q1_0:
set_rows_sycl_q<TIdx, block_q1_0, QK1_0, cpy_blck_f32_q1_0>(src0_d, src1_d, (block_q1_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
break;
case GGML_TYPE_Q5_1:
set_rows_sycl_q<TIdx, block_q5_1, QK5_1, cpy_blck_f32_q5_1>(src0_d, src1_d, (block_q5_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
break;
@@ -217,7 +220,12 @@ static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * s
case GGML_TYPE_IQ4_NL:
set_rows_sycl_q<TIdx, block_iq4_nl, QK4_NL, cpy_blck_f32_iq4_nl>(src0_d, src1_d, (block_iq4_nl *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
break;
case GGML_TYPE_MXFP4:
set_rows_sycl_q<TIdx, block_mxfp4, QK_MXFP4, cpy_blck_f32_mxfp4>(src0_d, src1_d, (block_mxfp4 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
break;
case GGML_TYPE_NVFP4:
set_rows_sycl_q<TIdx, block_nvfp4, QK_NVFP4, cpy_blck_f32_nvfp4>(src0_d, src1_d, (block_nvfp4 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
break;
default:
GGML_ABORT("Unsupported tensor type!");
break;