ggml : fix llamafile sgemm wdata offsets (llama/6710)

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-04-16 23:50:22 +03:00
parent 7a4f7d825e
commit c97796aa0f

11
ggml.c
View File

@ -33,12 +33,8 @@
#include <unistd.h>
#endif
#ifndef GGML_USE_LLAMAFILE
#ifdef __ARM_FEATURE_MATMUL_INT8
#define GGML_USE_LLAMAFILE 0
#else
#define GGML_USE_LLAMAFILE 1
#endif
#undef GGML_USE_LLAMAFILE
#endif
#if defined(_MSC_VER)
@ -10941,8 +10937,9 @@ UseGgmlGemm1:;
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
(const char *)wdata + (nb12/ggml_type_size(src1->type)*ggml_type_size(vec_dot_type)*i12 +
nb13/ggml_type_size(src1->type)*ggml_type_size(vec_dot_type)*i13),
(const char *)wdata + ggml_row_size(vec_dot_type,
nb12/ggml_type_size(src1->type)*i12 +
nb13/ggml_type_size(src1->type)*i13),
row_size/ggml_type_size(vec_dot_type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),