forked from leejet/stable-diffusion.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathconvert.cpp
More file actions
183 lines (156 loc) · 6.94 KB
/
Copy pathconvert.cpp
File metadata and controls
183 lines (156 loc) · 6.94 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
#include <cstring>
#include <mutex>
#include <regex>
#include <vector>
#include "model_io/gguf_io.h"
#include "model_io/safetensors_io.h"
#include "model_loader.h"
#include "util.h"
#include "ggml_extend_backend.h"
static ggml_type get_export_tensor_type(ModelLoader& model_loader,
const TensorStorage& tensor_storage,
ggml_type type,
const TensorTypeRules& tensor_type_rules) {
const std::string& name = tensor_storage.name;
ggml_type tensor_type = tensor_storage.type;
ggml_type dst_type = type;
for (const auto& tensor_type_rule : tensor_type_rules) {
std::regex pattern(tensor_type_rule.first);
if (std::regex_search(name, pattern)) {
dst_type = tensor_type_rule.second;
break;
}
}
if (model_loader.tensor_should_be_converted(tensor_storage, dst_type)) {
tensor_type = dst_type;
}
return tensor_type;
}
static bool load_tensors_for_export(ModelLoader& model_loader,
ggml_context* ggml_ctx,
ggml_type type,
const TensorTypeRules& tensor_type_rules,
std::vector<TensorWriteInfo>& tensors) {
std::mutex tensor_mutex;
auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
const std::string& name = tensor_storage.name;
ggml_type tensor_type = get_export_tensor_type(model_loader, tensor_storage, type, tensor_type_rules);
std::lock_guard<std::mutex> lock(tensor_mutex);
ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, tensor_type, tensor_storage.n_dims, tensor_storage.ne);
if (tensor == nullptr) {
LOG_ERROR("ggml_new_tensor failed");
return false;
}
ggml_set_name(tensor, name.c_str());
if (!tensor->data) {
GGML_ASSERT(ggml_nelements(tensor) == 0);
// Avoid crashing writers by setting a dummy pointer for zero-sized tensors.
LOG_DEBUG("setting dummy pointer for zero-sized tensor %s", name.c_str());
tensor->data = ggml_get_mem_buffer(ggml_ctx);
}
TensorWriteInfo write_info;
write_info.tensor = tensor;
write_info.n_dims = tensor_storage.n_dims;
for (int i = 0; i < tensor_storage.n_dims; ++i) {
write_info.ne[i] = tensor_storage.ne[i];
}
*dst_tensor = tensor;
tensors.push_back(std::move(write_info));
return true;
};
bool success = model_loader.load_tensors(on_new_tensor_cb);
LOG_INFO("load tensors done");
return success;
}
static bool init_convert_path(ModelLoader& model_loader, const char* path, const char* prefix, bool& loaded_any) {
if (path == nullptr || strlen(path) == 0) {
return true;
}
if (!model_loader.init_from_file(path, prefix)) {
LOG_ERROR("init model loader from file failed: '%s'", path);
return false;
}
loaded_any = true;
return true;
}
static bool export_loaded_model(ModelLoader& model_loader,
const char* output_path,
sd_type_t output_type,
const char* tensor_type_rules) {
ggml_type type = sd_type_to_ggml_type(output_type);
bool output_is_safetensors = ends_with(output_path, ".safetensors");
TensorTypeRules type_rules = parse_tensor_type_rules(tensor_type_rules);
auto backend = sd_backend_cpu_init();
size_t mem_size = 1 * 1024 * 1024; // for padding
mem_size += model_loader.get_tensor_storage_map().size() * ggml_tensor_overhead();
mem_size += model_loader.get_params_mem_size(backend, type);
LOG_INFO("model tensors mem size: %.2fMB", mem_size / 1024.f / 1024.f);
ggml_context* ggml_ctx = ggml_init({mem_size, nullptr, false});
if (ggml_ctx == nullptr) {
LOG_ERROR("ggml_init failed for converter");
ggml_backend_free(backend);
return false;
}
std::vector<TensorWriteInfo> tensors;
bool success = load_tensors_for_export(model_loader, ggml_ctx, type, type_rules, tensors);
ggml_backend_free(backend);
std::string error;
if (success) {
if (output_is_safetensors) {
success = write_safetensors_file(output_path, tensors, &error);
} else {
success = write_gguf_file(output_path, tensors, &error);
}
}
if (!success && !error.empty()) {
LOG_ERROR("%s", error.c_str());
}
ggml_free(ggml_ctx);
return success;
}
bool convert_with_components(const char* model_path,
const char* clip_l_path,
const char* clip_g_path,
const char* t5xxl_path,
const char* diffusion_model_path,
const char* vae_path,
const char* output_path,
sd_type_t output_type,
const char* tensor_type_rules,
bool convert_name) {
ModelLoader model_loader;
bool loaded_any = false;
if (!init_convert_path(model_loader, model_path, "", loaded_any) ||
!init_convert_path(model_loader, clip_l_path, "text_encoders.clip_l.transformer.", loaded_any) ||
!init_convert_path(model_loader, clip_g_path, "text_encoders.clip_g.transformer.", loaded_any) ||
!init_convert_path(model_loader, t5xxl_path, "text_encoders.t5xxl.transformer.", loaded_any) ||
!init_convert_path(model_loader, diffusion_model_path, "model.diffusion_model.", loaded_any) ||
!init_convert_path(model_loader, vae_path, "vae.", loaded_any)) {
return false;
}
if (!loaded_any) {
LOG_ERROR("no input model path provided for convert");
return false;
}
if (convert_name) {
model_loader.convert_tensors_name();
}
return export_loaded_model(model_loader, output_path, output_type, tensor_type_rules);
}
bool convert(const char* input_path,
const char* vae_path,
const char* output_path,
sd_type_t output_type,
const char* tensor_type_rules,
bool convert_name) {
return convert_with_components(input_path,
nullptr,
nullptr,
nullptr,
nullptr,
vae_path,
output_path,
output_type,
tensor_type_rules,
convert_name);
}