vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Explicitly state scaled data type

Browse source
This commit is contained in:
Aidan 2024-02-27 12:02:41 +00:00
parent b2aaee3500
commit abed262eeb
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml-sycl.cpp
View file

@ -13401,8 +13401,8 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
main_stream->submit([&](sycl::handler &cgh) { main_stream->submit([&](sycl::handler &cgh) {
const void **ptrs_src_get = ptrs_src.get(); const void **ptrs_src_get = ptrs_src.get();
void **ptrs_dst_get = ptrs_dst.get(); void **ptrs_dst_get = ptrs_dst.get();
auto nb12_scaled = src1->type == GGML_TYPE_F16 ? nb12 : nb12 / 2; size_t nb12_scaled = src1->type == GGML_TYPE_F16 ? nb12 : nb12 / 2;
auto nb13_scaled = src1->type == GGML_TYPE_F16 ? nb13 : nb13 / 2; size_t nb13_scaled = src1->type == GGML_TYPE_F16 ? nb13 : nb13 / 2;
cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) { [=](sycl::nd_item<3> item_ct1) {
k_compute_batched_ptrs( k_compute_batched_ptrs(