ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-05-01 03:41:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

bitnet(scale in a separate tensor): mul -> scale on the CPU

Browse Source
This commit is contained in:
Kawrakow
2024-06-20 08:21:25 +03:00
parent e423af855f
commit 8b31c14e0d
2 changed files with 29 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
ggml.c
View File

@@ -10156,6 +10156,23 @@ static void ggml_compute_forward_mul_f32(
const int ith = params->ith;
const int nth = params->nth;
if (ggml_nelements(dst->src[1]) == 1 && ggml_is_contiguous(dst->src[0]) && ggml_is_contiguous(dst) &&
dst->src[0]->type == GGML_TYPE_F32 && dst->src[1]->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
int64_t nelements = ggml_nelements(dst->src[0]);
int64_t n_per_thread = (nelements + nth - 1)/nth;
n_per_thread = MAX(1024, n_per_thread);
int64_t start = n_per_thread*ith;
if (start >= nelements) return;
int64_t end = MIN(nelements, start + n_per_thread);
const float * src = (const float *)dst->src[0]->data + start;
float * res = (float *)dst->data + start;
if (res != src) {
memcpy(res, src, (end - start)*sizeof(float));
}
ggml_vec_scale_f32(end - start, res, *(const float *)dst->src[1]->data);
return;
}
const int64_t nr = ggml_nrows(src0);
GGML_TENSOR_BINARY_OP_LOCALS