mirror/whisper.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Importance Matrix calculation (llama/4861)

Browse Source 
* imatrix: 1st version

* imatrix: WIP

* Cleanup

* Update examples/imatrix/imatrix.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Kawrakow 2024-01-12 06:59:57 +01:00 committed by Georgi Gerganov
parent d05b7ee90e
commit 3fa98f4395
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
2 changed files with 20 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
ggml.c
View File

@ -394,6 +394,12 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y); static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y);
static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y); static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y);
ggml_collect_imatrix_t g_imatrix_collect = NULL;
void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect) {
g_imatrix_collect = imatrix_collect;
}
static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
[GGML_TYPE_I8] = { [GGML_TYPE_I8] = {
.type_name = "i8", .type_name = "i8",
@ -9763,6 +9769,10 @@ static void ggml_compute_forward_mul_mat(
const int ith = params->ith; const int ith = params->ith;
const int nth = params->nth; const int nth = params->nth;
if (ith == 1 && g_imatrix_collect) {
g_imatrix_collect(src0, src1);
}
const enum ggml_type type = src0->type; const enum ggml_type type = src0->type;
const bool src1_cont = ggml_is_contiguous(src1); const bool src1_cont = ggml_is_contiguous(src1);
@ -10066,6 +10076,10 @@ static void ggml_compute_forward_mul_mat_id(
const struct ggml_tensor * src0_cur = dst->src[cur_a + 2]; const struct ggml_tensor * src0_cur = dst->src[cur_a + 2];
if (ith == 1 && g_imatrix_collect) {
g_imatrix_collect(src0_cur, src1);
}
const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
const size_t row_size = ggml_row_size(vec_dot_type, ne10); const size_t row_size = ggml_row_size(vec_dot_type, ne10);

6
ggml.h
View File

@ -2067,6 +2067,12 @@ extern "C" {
GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist); GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
//
// Importance matrix
//
typedef void(*ggml_collect_imatrix_t)(const struct ggml_tensor * src0, const struct ggml_tensor * src1);
GGML_API void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect);
// //
// gguf // gguf
// //