Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-02-07 08:53:16 +01:00
Code Issues Packages Projects Releases Wiki Activity

Upgrade v1 format to v2 by leveraging quantize

Browse Source
This commit is contained in:
Howard Su 2023-05-17 23:39:39 +08:00
parent fab49c685e
commit b8d69650dc
3 changed files with 49 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
ggml.c
View File

@ -813,6 +813,34 @@ typedef struct {
} block_q8_1; } block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
void quantize_upgrade(enum ggml_type type, void* data, size_t size) {
if (type == GGML_TYPE_Q4_0) {
int qk = ggml_blck_size(type);
const size_t nb = size / sizeof(block_q4_0);
block_q4_0 *blk = (block_q4_0 *)data;
block_q4_0 new_blk;
for (size_t i = 0; i < nb ; i++) {
for (size_t j = 0; j < qk/4; j++)
{
// old: d0, d1, d2, d3, d4, ....... d_half, d_half1
// new: d0, d_half, d1, d_half1
uint8_t d1;
uint8_t d2;
d1 = blk[i].qs[0 + j];
d2 = blk[i].qs[qk/4 + j];
new_blk.qs[0 + j * 2] = (d1 & 0x0f) | ((d2 & 0x0f) << 4);
new_blk.qs[1 + j * 2] = (d1 >> 4) | (d2 & 0xf0);
}
memcpy(blk[i].qs, new_blk.qs, sizeof(new_blk.qs));
}
}
}
// reference implementation for deterministic creation of model files // reference implementation for deterministic creation of model files
static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) { static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) {
static const int qk = QK4_0; static const int qk = QK4_0;

1
ggml.h
View File

@ -1086,6 +1086,7 @@ 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);
GGML_API void quantize_upgrade(enum ggml_type type, void* data, size_t size);
// //
// system info // system info
// //

21
llama.cpp
View File

@ -2074,7 +2074,26 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
size_t new_size; size_t new_size;
llama_buffer work; llama_buffer work;
if (!quantize) { if (model_loader->file_loaders.at(0)->file_version == LLAMA_FILE_VERSION_GGJT_V1 && quantize) {
if (tensor.type == GGML_TYPE_Q4_0 && quantized_type == GGML_TYPE_Q4_0) {
// convet
new_type = tensor.type;
new_data = tensor.data;
new_size = tensor.size;
quantize_upgrade(new_type, new_data, new_size);
printf("Upgrade - size = %8.3f MB\n", tensor.size/1024.0/1024.0);
}
else if (tensor.type == GGML_TYPE_Q4_1 && quantized_type == GGML_TYPE_Q4_1) {
new_type = tensor.type;
new_data = tensor.data;
new_size = tensor.size;
quantize_upgrade(new_type, new_data, new_size);
printf("Upgrade - size = %8.3f MB\n", tensor.size/1024.0/1024.0);
}
else {
throw format("type %s unsupported for quantization format upgrade", ggml_type_name(tensor.type));
}
} else if (!quantize) {
new_type = tensor.type; new_type = tensor.type;
new_data = tensor.data; new_data = tensor.data;
new_size = tensor.size; new_size = tensor.size;