upstream-archive/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-08-04 01:00:43 -05:00
Code Issues Packages Projects Releases Wiki Activity

feat: add imatrix support (#633)

Browse Source
This commit is contained in:
stduhpf
2026-07-01 17:26:37 +02:00
committed by GitHub
parent 484baa41e5
commit 1a13107bac
13 changed files with 661 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/model_loader.cpp
View File

@@ -20,6 +20,7 @@
#include "model_io/torch_legacy_io.h"
#include "model_io/torch_zip_io.h"
#include "model_loader.h"
#include "runtime/imatrix.h"
#include "stable-diffusion.h"
#include "core/ggml_extend_backend.h"
@@ -156,7 +157,8 @@ void convert_tensor(void* src,
void* dst,
ggml_type dst_type,
int nrows,
int n_per_row) {
int n_per_row,
std::vector<float> imatrix = {}) {
int n = nrows * n_per_row;
if (src_type == dst_type) {
size_t nbytes = n * ggml_type_size(src_type) / ggml_blck_size(src_type);
@@ -165,7 +167,7 @@ void convert_tensor(void* src,
if (dst_type == GGML_TYPE_F16) {
ggml_fp32_to_fp16_row((float*)src, (ggml_fp16_t*)dst, n);
} else {
std::vector<float> imatrix(n_per_row, 1.0f); // dummy importance matrix
imatrix.resize(n_per_row, 1.0f);
const float* im = imatrix.data();
ggml_quantize_chunk(dst_type, (float*)src, dst, 0, nrows, n_per_row, im);
}
@@ -195,7 +197,7 @@ void convert_tensor(void* src,
if (dst_type == GGML_TYPE_F16) {
ggml_fp32_to_fp16_row((float*)src_data_f32, (ggml_fp16_t*)dst, n);
} else {
std::vector<float> imatrix(n_per_row, 1.0f); // dummy importance matrix
imatrix.resize(n_per_row, 1.0f);
const float* im = imatrix.data();
ggml_quantize_chunk(dst_type, (float*)src_data_f32, dst, 0, nrows, n_per_row, im);
}
@@ -970,6 +972,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
size_t total_tensors_processed = 0;
const int64_t t_start = start_time;
int last_n_threads = 1;
SDVersion imatrix_version = (version_ == VERSION_COUNT) ? get_sd_version() : version_;
for (size_t file_index = 0; file_index < file_data.size(); ++file_index) {
auto& fdata = file_data[file_index];
@@ -1154,12 +1157,15 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
failed = true;
return;
}
std::string processed_name = convert_tensor_name(tensor_storage.name, imatrix_version);
std::vector<float> imatrix = get_imatrix_collector().get_values(processed_name);
convert_tensor((void*)target_buf,
tensor_storage.type,
convert_buf,
dst_tensor->type,
(int)tensor_storage.nelements() / (int)tensor_storage.ne[0],
(int)tensor_storage.ne[0]);
(int)tensor_storage.ne[0],
std::move(imatrix));
} else {
convert_buf = read_buf;
}