ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-16 05:48:09 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

ggml : refactor online repacking (llama/10446)

Browse Source 
* rename ggml-cpu-aarch64.c to .cpp

* reformat extra cpu backend.

- clean Q4_0_N_M and IQ4_0_N_M
  - remove from "file" tensor type
  - allow only with dynamic repack

- extract cpu extra bufts and convert to C++
  - hbm
  - "aarch64"

- more generic use of extra buffer
  - generalise extra_supports_op
  - new API for "cpu-accel":
     - amx
     - aarch64

* clang-format

* Clean Q4_0_N_M ref

Enable restrict on C++

* add op GGML_OP_MUL_MAT_ID for Q4_0_N_M with runtime repack

* added/corrected control on tensor size for Q4 repacking.

* Update ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp

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

* Update ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp

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

* add debug logs on repacks.

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Djip007
2024-12-07 13:37:50 +01:00
committed by Georgi Gerganov
parent 4a6d52efe6
commit e990d1b791
23 changed files with 4794 additions and 702 deletions
Download Patch File Download Diff File Expand all files Collapse all files

86
ggml/src/ggml.c
View File

@ -8,7 +8,10 @@
// FIXME: required here for quantization functions
#include "ggml-quants.h"
#include "ggml-aarch64.h"
#ifdef GGML_USE_CPU_HBM
#include <hbwmalloc.h>
#endif
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <malloc.h> // using malloc.h with MSC/MINGW
@ -788,32 +791,23 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
.to_float = (ggml_to_float_t) ggml_bf16_to_fp32_row,
.from_float_ref = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref,
},
[GGML_TYPE_Q4_0_4_4] = {
.type_name = "q4_0_4x4",
.blck_size = QK4_0,
.blck_size_interleave = 4,
.type_size = sizeof(block_q4_0),
.is_quantized = true,
.to_float = NULL,
.from_float_ref = NULL,
[31] = { // GGML_TYPE_Q4_0_4_4
.type_name = "TYPE_Q4_0_4_4 REMOVED, use Q4_0 with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[GGML_TYPE_Q4_0_4_8] = {
.type_name = "q4_0_4x8",
.blck_size = QK4_0,
.blck_size_interleave = 8,
.type_size = sizeof(block_q4_0),
.is_quantized = true,
.to_float = NULL,
.from_float_ref = NULL,
[32] = { // GGML_TYPE_Q4_0_4_8
.type_name = "TYPE_Q4_0_4_8 REMOVED, use Q4_0 with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[GGML_TYPE_Q4_0_8_8] = {
.type_name = "q4_0_8x8",
.blck_size = QK4_0,
.blck_size_interleave = 8,
.type_size = sizeof(block_q4_0),
.is_quantized = true,
.to_float = NULL,
.from_float_ref = NULL,
[33] = { // GGML_TYPE_Q4_0_8_8
.type_name = "TYPE_Q4_0_8_8 REMOVED, use Q4_0 with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[GGML_TYPE_TQ1_0] = {
.type_name = "tq1_0",
@ -831,14 +825,23 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
.to_float = (ggml_to_float_t) dequantize_row_tq2_0,
.from_float_ref = (ggml_from_float_t) quantize_row_tq2_0_ref,
},
[GGML_TYPE_IQ4_NL_4_4] = {
.type_name = "iq4_nl_4x4",
.blck_size = QK4_NL,
.blck_size_interleave = 4,
.type_size = sizeof(block_iq4_nl),
.is_quantized = true,
.to_float = NULL,
.from_float_ref = NULL,
[36] = { // GGML_TYPE_IQ4_NL_4_4
.type_name = "TYPE_IQ4_NL_4_4 REMOVED, use IQ4_NL with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[37] = { // GGML_TYPE_IQ4_NL_4_8
.type_name = "TYPE_IQ4_NL_4_8 REMOVED, use IQ4_NL with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[38] = { // GGML_TYPE_IQ4_NL_8_8
.type_name = "TYPE_IQ4_NL_8_8 REMOVED, use IQ4_NL with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
};
@ -1270,9 +1273,6 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
case GGML_FTYPE_MOSTLY_IQ4_XS: wtype = GGML_TYPE_IQ4_XS; break;
case GGML_FTYPE_MOSTLY_IQ3_S: wtype = GGML_TYPE_IQ3_S; break;
case GGML_FTYPE_MOSTLY_IQ2_S: wtype = GGML_TYPE_IQ2_S; break;
case GGML_FTYPE_MOSTLY_Q4_0_4_4: wtype = GGML_TYPE_Q4_0_4_4; break;
case GGML_FTYPE_MOSTLY_Q4_0_4_8: wtype = GGML_TYPE_Q4_0_4_8; break;
case GGML_FTYPE_MOSTLY_Q4_0_8_8: wtype = GGML_TYPE_Q4_0_8_8; break;
case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break;
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
}
@ -6304,9 +6304,6 @@ size_t ggml_quantize_chunk(
case GGML_TYPE_IQ1_M: result = quantize_iq1_m (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_IQ4_NL: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_IQ4_XS: result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_0_8_8: result = quantize_q4_0_8x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_F16:
{
size_t elemsize = sizeof(ggml_fp16_t);
@ -6838,7 +6835,16 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
(int64_t) info->ne[2] *
(int64_t) info->ne[3];
if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) {
if (ggml_blck_size(info->type) == 0 ) {
// this tensor type support have been removed:
fprintf(stderr, "%s: tensor '%s' of type %d: %s\n",
__func__, info->name.data, (int) info->type, ggml_type_name(info->type));
fclose(file);
gguf_free(ctx);
return NULL;
}
if (ne % ggml_blck_size(info->type) != 0) {
fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
__func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
fclose(file);