vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 17:44:38 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.devops/llama-cli-intel.Dockerfile
#	.devops/llama-server-intel.Dockerfile
#	.github/workflows/build.yml
#	CMakePresets.json
#	Makefile
#	docs/backend/SYCL.md
#	docs/build.md
#	ggml/CMakeLists.txt
#	ggml/src/ggml-cpu/CMakeLists.txt
#	scripts/compare-llama-bench.py
#	scripts/sync-ggml-am.sh
#	scripts/sync-ggml.last
This commit is contained in:
Concedo 2024-11-16 17:20:14 +08:00
commit 590553ef07
18 changed files with 731 additions and 342 deletions
Download patch file Download diff file Expand all files Collapse all files

29
ggml/src/ggml-cpu/ggml-cpu.c
View file

@ -1472,8 +1472,12 @@ static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t
sumf += (ggml_float)_mm512_reduce_add_ps(c2);
#undef LOAD
#elif defined(__AVX2__)
#elif defined(__AVX2__) || defined(__AVX__)
#if defined(__AVX2__)
#define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16))
#else
#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1))
#endif
__m256 c1 = _mm256_setzero_ps();
__m256 c2 = _mm256_setzero_ps();
__m256 c3 = _mm256_setzero_ps();
@ -7358,6 +7362,7 @@ static void ggml_compute_forward_group_norm(
static void ggml_compute_forward_mul_mat_one_chunk(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
const enum ggml_type type,
const int64_t num_rows_per_vec_dot,
const int64_t ir0_start,
const int64_t ir0_end,
@ -7369,8 +7374,6 @@ static void ggml_compute_forward_mul_mat_one_chunk(
GGML_TENSOR_BINARY_OP_LOCALS
const enum ggml_type type = src0->type;
const bool src1_cont = ggml_is_contiguous(src1);
ggml_vec_dot_t const vec_dot = type_traits_cpu[type].vec_dot;
@ -7458,7 +7461,11 @@ static void ggml_compute_forward_mul_mat(
const int ith = params->ith;
const int nth = params->nth;
const enum ggml_type type = src0->type;
enum ggml_type type = src0->type;
if (src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
type = (enum ggml_type)(intptr_t)src0->extra;
}
enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type;
ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float;
@ -7505,15 +7512,15 @@ static void ggml_compute_forward_mul_mat(
if (src1_cont) {
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
nb01/ggml_type_size(type),
(const char *)src1->data + i12*nb12 + i13*nb13,
nb11/ggml_type_size(src1->type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
type,
src1->type,
dst->type))
goto UseGgmlGemm1;
@ -7566,15 +7573,15 @@ UseGgmlGemm1:;
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
nb01/ggml_type_size(type),
(const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
row_size/ggml_type_size(vec_dot_type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
type,
vec_dot_type,
dst->type))
goto UseGgmlGemm2;
@ -7659,7 +7666,7 @@ UseGgmlGemm2:;
const int64_t ir1_start = dr1 * ith1;
const int64_t ir1_end = MIN(ir1_start + dr1, nr1);
ggml_compute_forward_mul_mat_one_chunk(params, dst, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
ggml_compute_forward_mul_mat_one_chunk(params, dst, type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
if (nth >= nchunk0 * nchunk1) {
break;