-
Notifications
You must be signed in to change notification settings - Fork 48
/
quantize.cpp
413 lines (344 loc) · 13.2 KB
/
quantize.cpp
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
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
/*This script quantizes the weights of the 3 GPT encoders. 5 quantization types are
available:
- q4_0
- q4_1
- q5_0
- q5_1
- q8_0
Usage:
./quantize \
./ggml_weights/ggml_weights_text.bin \
./ggml_weights_q4/ggml_weights_text_quant.bin \
type
*/
#include "ggml.h"
#include "bark.h"
#include "bark-util.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <regex>
static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
{"q4_0", GGML_FTYPE_MOSTLY_Q4_0},
{"q4_1", GGML_FTYPE_MOSTLY_Q4_1},
{"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
{"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
{"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
};
void ggml_print_ftypes(FILE * fp) {
for (auto it = GGML_FTYPE_MAP.begin(); it != GGML_FTYPE_MAP.end(); it++) {
fprintf(fp, " type = \"%s\" or %d\n", it->first.c_str(), it->second);
}
}
enum ggml_ftype ggml_parse_ftype(const char * str) {
enum ggml_ftype ftype;
if (str[0] == 'q') {
const auto it = GGML_FTYPE_MAP.find(str);
if (it == GGML_FTYPE_MAP.end()) {
fprintf(stderr, "%s: unknown ftype '%s'\n", __func__, str);
return GGML_FTYPE_UNKNOWN;
}
ftype = it->second;
} else {
ftype = (enum ggml_ftype) atoi(str);
}
return ftype;
}
bool ggml_common_quantize_0(
std::ifstream & fin,
std::ofstream & fout,
const ggml_ftype ftype,
const std::vector<std::string> & to_quant,
const std::vector<std::string> & to_skip) {
ggml_type qtype = GGML_TYPE_F32;
switch (ftype) {
case GGML_FTYPE_MOSTLY_Q4_0: qtype = GGML_TYPE_Q4_0; break;
case GGML_FTYPE_MOSTLY_Q4_1: qtype = GGML_TYPE_Q4_1; break;
case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
case GGML_FTYPE_UNKNOWN:
case GGML_FTYPE_ALL_F32:
case GGML_FTYPE_MOSTLY_F16:
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
case GGML_FTYPE_MOSTLY_Q2_K:
case GGML_FTYPE_MOSTLY_Q3_K:
case GGML_FTYPE_MOSTLY_Q4_K:
case GGML_FTYPE_MOSTLY_Q5_K:
case GGML_FTYPE_MOSTLY_Q6_K:
{
fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
return false;
}
};
if (!ggml_is_quantized(qtype)) {
fprintf(stderr, "%s: invalid quantization type %d (%s)\n", __func__, qtype, ggml_type_name(qtype));
return false;
}
size_t total_size_org = 0;
size_t total_size_new = 0;
std::vector<float> work;
std::vector<uint8_t> data_u8;
std::vector<ggml_fp16_t> data_f16;
std::vector<float> data_f32;
std::vector<int64_t> hist_all(1 << 4, 0);
while (true) {
int32_t n_dims;
int32_t length;
int32_t ttype;
read_safe(fin, n_dims);
read_safe(fin, length);
read_safe(fin, ttype);
if (fin.eof()) {
break;
}
int32_t nelements = 1;
int32_t ne[4] = { 1, 1, 1, 1 };
for (int i = 0; i < n_dims; ++i) {
read_safe(fin, ne[i]);
nelements *= ne[i];
}
std::string name(length, 0);
fin.read(&name[0], length);
printf("%64s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype));
bool quantize = false;
// check if we should quantize this tensor
for (const auto & s : to_quant) {
if (std::regex_match(name, std::regex(s))) {
quantize = true;
break;
}
}
// check if we should skip this tensor
for (const auto & s : to_skip) {
if (std::regex_match(name, std::regex(s))) {
quantize = false;
break;
}
}
// quantize only 2D tensors
quantize &= (n_dims == 2);
if (quantize) {
if (ttype != GGML_TYPE_F32 && ttype != GGML_TYPE_F16) {
fprintf(stderr, "%s: unsupported ttype %d (%s) for integer quantization\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
return false;
}
if (ttype == GGML_TYPE_F16) {
data_f16.resize(nelements);
fin.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
data_f32.resize(nelements);
for (int i = 0; i < nelements; ++i) {
data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
}
} else {
data_f32.resize(nelements);
fin.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
}
ttype = qtype;
} else {
const int bpe = (ttype == 0) ? sizeof(float) : sizeof(uint16_t);
data_u8.resize(nelements*bpe);
fin.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
}
write_safe(fout, n_dims);
write_safe(fout, length);
write_safe(fout, ttype);
for (int i = 0; i < n_dims; ++i) {
write_safe(fout, ne[i]);
}
fout.write(&name[0], length);
if (quantize) {
work.resize(nelements); // for quantization
size_t cur_size = 0;
std::vector<int64_t> hist_cur(1 << 4, 0);
switch ((ggml_type) ttype) {
case GGML_TYPE_Q4_0:
{
cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q4_1:
{
cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q5_0:
{
cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q5_1:
{
cur_size = ggml_quantize_q5_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q8_0:
{
cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_I8:
case GGML_TYPE_I16:
case GGML_TYPE_I32:
case GGML_TYPE_Q8_1:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
case GGML_TYPE_Q8_K:
case GGML_TYPE_COUNT:
{
fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
return false;
}
}
fout.write(reinterpret_cast<char *>(work.data()), cur_size);
total_size_new += cur_size;
printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
for (int i = 0; i < (int) hist_cur.size(); ++i) {
hist_all[i] += hist_cur[i];
}
for (int i = 0; i < (int) hist_cur.size(); ++i) {
printf("%5.3f ", hist_cur[i] / (float)nelements);
}
printf("\n");
} else {
printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
total_size_new += data_u8.size();
}
total_size_org += nelements * sizeof(float);
}
printf("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
printf("%s: quant size = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
{
int64_t sum_all = 0;
for (int i = 0; i < (int) hist_all.size(); ++i) {
sum_all += hist_all[i];
}
printf("%s: hist: ", __func__);
for (int i = 0; i < (int) hist_all.size(); ++i) {
printf("%5.3f ", hist_all[i] / (float)sum_all);
}
printf("\n");
}
return true;
}
bool bark_model_quantize(
const std::string & fname_inp,
const std::string & fname_out,
ggml_ftype ftype) {
printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
gpt_model model;
auto fin = std::ifstream(fname_inp, std::ios::binary);
if (!fin) {
fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
return false;
}
auto fout = std::ofstream(fname_out, std::ios::binary);
if (!fout) {
fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
return false;
}
// verify magic
{
uint32_t magic;
fin.read((char *) &magic, sizeof(magic));
if (magic != GGML_FILE_MAGIC) {
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
return false;
}
fout.write((char *) &magic, sizeof(magic));
}
gpt_hparams hparams;
// load hparams
{
auto & hparams = model.hparams;
read_safe(fin, hparams.n_layer);
read_safe(fin, hparams.n_head);
read_safe(fin, hparams.n_embd);
read_safe(fin, hparams.block_size);
read_safe(fin, hparams.n_in_vocab);
read_safe(fin, hparams.n_out_vocab);
read_safe(fin, hparams.n_lm_heads);
read_safe(fin, hparams.n_wtes);
read_safe(fin, hparams.ftype);
const int32_t qntvr_src = hparams.ftype / GGML_QNT_VERSION_FACTOR;
int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
printf("%s: n_in_vocab = %d\n", __func__, hparams.n_in_vocab);
printf("%s: n_out_vocab = %d\n", __func__, hparams.n_out_vocab);
printf("%s: block_size = %d\n", __func__, hparams.block_size);
printf("%s: n_embd = %d\n", __func__, hparams.n_embd);
printf("%s: n_head = %d\n", __func__, hparams.n_head);
printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
printf("%s: n_lm_heads = %d\n", __func__, hparams.n_lm_heads);
printf("%s: n_wtes = %d\n", __func__, hparams.n_wtes);
printf("%s: ftype (src) = %d\n", __func__, hparams.ftype);
printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
write_safe(fout, hparams.n_layer);
write_safe(fout, hparams.n_head);
write_safe(fout, hparams.n_embd);
write_safe(fout, hparams.block_size);
write_safe(fout, hparams.n_in_vocab);
write_safe(fout, hparams.n_out_vocab);
write_safe(fout, hparams.n_lm_heads);
write_safe(fout, hparams.n_wtes);
write_safe(fout, ftype_dst);
}
// regexes of tensor names to be quantized
const std::vector<std::string> to_quant = {
"model/wte/.*",
"model/lm_head/.*",
"model/h.*/attn/c_attn/w",
"model/h.*/attn/c_proj/w",
"model/h.*/mlp/c_fc/w",
"model/h.*/mlp/c_proj/w",
};
if (!ggml_common_quantize_0(fin, fout, ftype, to_quant, {})) {
fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
return false;
}
fin.close();
fout.close();
return true;
}
int main(int argc, char ** argv) {
if (argc != 4) {
fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
ggml_print_ftypes(stderr);
return 1;
}
// needed to initialize f16 tables
{
struct ggml_init_params params = { 0, NULL, false };
struct ggml_context * ctx = ggml_init(params);
ggml_free(ctx);
}
const std::string fname_inp = argv[1];
const std::string fname_out = argv[2];
const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
const int64_t t_main_start_us = ggml_time_us();
int64_t t_quantize_us = 0;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!bark_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
return 1;
}
t_quantize_us = ggml_time_us() - t_start_us;
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n");
printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
return 0;
}