forked from leejet/stable-diffusion.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathstable-diffusion.cpp
More file actions
3723 lines (3130 loc) · 151 KB
/
Copy pathstable-diffusion.cpp
File metadata and controls
3723 lines (3130 loc) · 151 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
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
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#include <assert.h>
#include <algorithm>
#include <cstring>
#include <fstream>
#include <iostream>
#include <iterator>
#include <map>
#include <random>
#include <regex>
#include <set>
#include <sstream>
#include <string>
#include <unordered_map>
#include <vector>
#include "ggml/ggml.h"
#include "stable-diffusion.h"
static SDLogLevel log_level = SDLogLevel::INFO;
#define __FILENAME__ "stable-diffusion.cpp"
#define SD_LOG(level, format, ...) \
do { \
if (level < log_level) { \
break; \
} \
if (level == SDLogLevel::DEBUG) { \
printf("[DEBUG] %s:%-4d - " format "\n", __FILENAME__, __LINE__, ##__VA_ARGS__); \
} else if (level == SDLogLevel::INFO) { \
printf("[INFO] %s:%-4d - " format "\n", __FILENAME__, __LINE__, ##__VA_ARGS__); \
} else if (level == SDLogLevel::WARN) { \
fprintf(stderr, "[WARN] %s:%-4d - " format "\n", __FILENAME__, __LINE__, ##__VA_ARGS__); \
} else if (level == SDLogLevel::ERROR) { \
fprintf(stderr, "[ERROR] %s:%-4d - " format "\n", __FILENAME__, __LINE__, ##__VA_ARGS__); \
} \
} while (0)
#define LOG_DEBUG(format, ...) SD_LOG(SDLogLevel::DEBUG, format, ##__VA_ARGS__)
#define LOG_INFO(format, ...) SD_LOG(SDLogLevel::INFO, format, ##__VA_ARGS__)
#define LOG_WARN(format, ...) SD_LOG(SDLogLevel::WARN, format, ##__VA_ARGS__)
#define LOG_ERROR(format, ...) SD_LOG(SDLogLevel::ERROR, format, ##__VA_ARGS__)
#define GGML_FILE_MAGIC 0x67676d6c
#define TIMESTEPS 1000
/*================================================== Helper Functions ================================================*/
void set_sd_log_level(SDLogLevel level) {
log_level = level;
}
std::string sd_get_system_info() {
std::stringstream ss;
ss << "System Info: \n";
ss << " BLAS = " << ggml_cpu_has_blas() << std::endl;
ss << " SSE3 = " << ggml_cpu_has_sse3() << std::endl;
ss << " AVX = " << ggml_cpu_has_avx() << std::endl;
ss << " AVX2 = " << ggml_cpu_has_avx2() << std::endl;
ss << " AVX512 = " << ggml_cpu_has_avx512() << std::endl;
ss << " AVX512_VBMI = " << ggml_cpu_has_avx512_vbmi() << std::endl;
ss << " AVX512_VNNI = " << ggml_cpu_has_avx512_vnni() << std::endl;
ss << " FMA = " << ggml_cpu_has_fma() << std::endl;
ss << " NEON = " << ggml_cpu_has_neon() << std::endl;
ss << " ARM_FMA = " << ggml_cpu_has_arm_fma() << std::endl;
ss << " F16C = " << ggml_cpu_has_f16c() << std::endl;
ss << " FP16_VA = " << ggml_cpu_has_fp16_va() << std::endl;
ss << " WASM_SIMD = " << ggml_cpu_has_wasm_simd() << std::endl;
ss << " VSX = " << ggml_cpu_has_vsx() << std::endl;
return ss.str();
}
ggml_tensor* load_tensor_from_file(ggml_context* ctx, const std::string& file_path) {
std::ifstream file(file_path, std::ios::binary);
if (!file.is_open()) {
LOG_ERROR("failed to open '%s'", file_path.c_str());
return NULL;
}
int32_t n_dims;
int32_t length;
int32_t ttype;
file.read(reinterpret_cast<char*>(&n_dims), sizeof(n_dims));
file.read(reinterpret_cast<char*>(&length), sizeof(length));
file.read(reinterpret_cast<char*>(&ttype), sizeof(ttype));
if (file.eof()) {
LOG_ERROR("incomplete file '%s'", file_path.c_str());
return NULL;
}
int32_t nelements = 1;
int32_t ne[4] = {1, 1, 1, 1};
for (int i = 0; i < n_dims; ++i) {
file.read(reinterpret_cast<char*>(&ne[i]), sizeof(ne[i]));
nelements *= ne[i];
}
std::string name(length, 0);
file.read(&name[0], length);
ggml_tensor* tensor = ggml_new_tensor_4d(ctx, (ggml_type)ttype, ne[0], ne[1], ne[2], ne[3]);
const size_t bpe = ggml_type_size(ggml_type(ttype));
file.read(reinterpret_cast<char*>(tensor->data), ggml_nbytes(tensor));
return tensor;
}
static std::default_random_engine generator;
void set_random_seed(int seed) {
generator.seed(seed);
}
void ggml_tensor_set_f32_randn(struct ggml_tensor* tensor) {
float mean = 0.0;
float stddev = 1.0;
std::normal_distribution<float> distribution(mean, stddev);
for (int i = 0; i < ggml_nelements(tensor); i++) {
float random_number = distribution(generator);
ggml_set_f32_1d(tensor, i, random_number);
}
}
// set tensor[i, j, k, l]
// set tensor[l]
// set tensor[k, l]
// set tensor[j, k, l]
void ggml_tensor_set_f32(struct ggml_tensor* tensor, float value, int l, int k = 0, int j = 0, int i = 0) {
GGML_ASSERT(tensor->nb[0] == sizeof(float));
*(float*)((char*)(tensor->data) + i * tensor->nb[3] + j * tensor->nb[2] + k * tensor->nb[1] + l * tensor->nb[0]) = value;
}
float ggml_tensor_get_f32(const ggml_tensor* tensor, int l, int k = 0, int j = 0, int i = 0) {
GGML_ASSERT(tensor->nb[0] == sizeof(float));
return *(float*)((char*)(tensor->data) + i * tensor->nb[3] + j * tensor->nb[2] + k * tensor->nb[1] + l * tensor->nb[0]);
}
void print_ggml_tensor(struct ggml_tensor* tensor, bool shape_only = false) {
printf("shape(%zu, %zu, %zu, %zu)\n", tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
if (shape_only) {
return;
}
int range = 3;
for (int i = 0; i < tensor->ne[3]; i++) {
if (i >= range && i + range < tensor->ne[3]) {
continue;
}
for (int j = 0; j < tensor->ne[2]; j++) {
if (j >= range && j + range < tensor->ne[2]) {
continue;
}
for (int k = 0; k < tensor->ne[1]; k++) {
if (k >= range && k + range < tensor->ne[1]) {
continue;
}
for (int l = 0; l < tensor->ne[0]; l++) {
if (l >= range && l + range < tensor->ne[0]) {
continue;
}
printf(" [%d, %d, %d, %d] = %f\n", i, j, k, l, ggml_tensor_get_f32(tensor, l, k, j, i));
}
}
}
}
}
void copy_ggml_tensor(
struct ggml_tensor* dst,
const struct ggml_tensor* src) {
dst->nb[0] = src->nb[0];
dst->nb[1] = src->nb[1];
dst->nb[2] = src->nb[2];
dst->nb[3] = src->nb[3];
memcpy(((char*)dst->data), ((char*)src->data), ggml_nbytes(dst));
}
// Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
void set_timestep_embedding(struct ggml_tensor* timesteps, struct ggml_tensor* embedding, int dim, int max_period = 10000) {
// timesteps: [N,]
// embedding: [(dim + 1)/2, N]
int half = dim / 2;
std::vector<float> freqs(half);
for (int i = 0; i < half; ++i) {
freqs[i] = (float)std::exp(-std::log(max_period) * i / half);
}
for (int i = 0; i < timesteps->ne[0]; ++i) {
for (int j = 0; j < half; ++j) {
float arg = ggml_get_f32_1d(timesteps, i) * freqs[j];
ggml_tensor_set_f32(embedding, std::cos(arg), j, i);
ggml_tensor_set_f32(embedding, std::sin(arg), j + half, i);
}
if (dim % 2 != 0) {
*(float*)((char*)embedding->data + i * embedding->nb[1] + dim * embedding->nb[0]) = 0;
}
}
}
struct ggml_tensor* new_timestep_embedding(struct ggml_context* ctx, struct ggml_tensor* timesteps, int dim, int max_period = 10000) {
// timesteps: [N,]
// embedding: [(dim + 1)/2, N]
int acutual_dim = dim;
if (dim % 2 != 0) {
acutual_dim = dim + 1;
}
struct ggml_tensor* embedding = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, acutual_dim, timesteps->ne[0]);
if (!ggml_get_no_alloc(ctx)) {
set_timestep_embedding(timesteps, embedding, dim, max_period);
}
return embedding;
}
std::vector<uint8_t> ggml_to_image_vec(struct ggml_tensor* t) {
int64_t w = t->ne[0];
int64_t h = t->ne[1];
int64_t c = t->ne[2];
std::vector<uint8_t> vec;
vec.resize(w * h * c);
uint8_t* data = (uint8_t*)vec.data();
for (int i = 0; i < h; i++) {
for (int j = 0; j < w; j++) {
for (int k = 0; k < c; k++) {
float value = ggml_tensor_get_f32(t, j, i, k);
value = (value + 1.0f) * 0.5f;
if (value < 0) {
value = 0;
} else if (value > 1) {
value = 1;
}
value *= 255.f;
*(data + i * w * c + j * c + k) = (uint8_t)value;
}
}
}
return vec;
}
void image_vec_to_ggml(const std::vector<uint8_t>& vec,
struct ggml_tensor* t) {
int64_t w = t->ne[0];
int64_t h = t->ne[1];
int64_t c = t->ne[2];
uint8_t* data = (uint8_t*)vec.data();
for (int i = 0; i < h; i++) {
for (int j = 0; j < w; j++) {
for (int k = 0; k < c; k++) {
float value = *(data + i * w * c + j * c + k);
value = value / 255.f;
value = 2 * value - 1;
ggml_tensor_set_f32(t, value, j, i, k);
}
}
}
}
struct ggml_tensor * ggml_group_norm_32(struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_group_norm(ctx, a, 32);
}
/*================================================== CLIPTokenizer ===================================================*/
const std::string UNK_TOKEN = "<|endoftext|>";
const std::string BOS_TOKEN = "<|startoftext|>";
const std::string EOS_TOKEN = "<|endoftext|>";
const std::string PAD_TOEKN = "<|endoftext|>";
const int UNK_TOKEN_ID = 49407;
const int BOS_TOKEN_ID = 49406;
const int EOS_TOKEN_ID = 49407;
const int PAD_TOKEN_ID = 49407;
// Ref: https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py
// TODO: implement bpe
class CLIPTokenizer {
private:
std::map<std::string, int32_t> encoder;
std::regex pat;
static std::string strip(const std::string& str) {
std::string::size_type start = str.find_first_not_of(" \t\n\r\v\f");
std::string::size_type end = str.find_last_not_of(" \t\n\r\v\f");
if (start == std::string::npos) {
// String contains only whitespace characters
return "";
}
return str.substr(start, end - start + 1);
}
static std::string whitespace_clean(std::string text) {
text = std::regex_replace(text, std::regex(R"(\s+)"), " ");
text = strip(text);
return text;
}
public:
CLIPTokenizer() = default;
std::string bpe(std::string token) {
std::string word = token + "</w>";
if (encoder.find(word) != encoder.end()) {
return word;
} else if (encoder.find(token) != encoder.end()) {
return token;
}
return UNK_TOKEN;
}
void add_token(std::string token, int32_t token_id) {
encoder[token] = token_id;
}
std::vector<int> tokenize(std::string text, size_t max_length = 0, bool padding = false) {
std::vector<int32_t> tokens = encode(text);
tokens.insert(tokens.begin(), BOS_TOKEN_ID);
if (max_length > 0) {
if (tokens.size() > max_length - 1) {
tokens.resize(max_length - 1);
} else {
if (padding) {
tokens.insert(tokens.end(), max_length - 1 - tokens.size(), PAD_TOKEN_ID);
}
}
}
tokens.push_back(EOS_TOKEN_ID);
return tokens;
}
std::vector<int> encode(std::string text) {
std::string original_text = text;
std::vector<int32_t> bpe_tokens;
text = whitespace_clean(text);
std::transform(text.begin(), text.end(), text.begin(), [](unsigned char c) { return std::tolower(c); });
std::regex pat(R"(<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[[:alpha:]]+|[[:digit:]]|[^[:space:][:alpha:][:digit:]]+)",
std::regex::icase);
std::smatch matches;
std::string str = text;
std::vector<std::string> token_strs;
while (std::regex_search(str, matches, pat)) {
for (auto& token : matches) {
std::istringstream iss(bpe(token));
std::vector<std::string> tokens{std::istream_iterator<std::string>{iss},
std::istream_iterator<std::string>{}};
for (const auto& bpe_token : tokens) {
bpe_tokens.push_back(encoder[bpe_token]);
token_strs.push_back(bpe_token);
}
}
str = matches.suffix();
}
std::stringstream ss;
ss << "[";
for (auto token : token_strs) {
ss << "\"" << token << "\", ";
}
ss << "]";
LOG_DEBUG("split prompt \"%s\" to tokens %s", original_text.c_str(), ss.str().c_str());
return bpe_tokens;
}
};
// Ref: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/cad87bf4e3e0b0a759afa94e933527c3123d59bc/modules/prompt_parser.py#L345
//
// Parses a string with attention tokens and returns a list of pairs: text and its associated weight.
// Accepted tokens are:
// (abc) - increases attention to abc by a multiplier of 1.1
// (abc:3.12) - increases attention to abc by a multiplier of 3.12
// [abc] - decreases attention to abc by a multiplier of 1.1
// \( - literal character '('
// \[ - literal character '['
// \) - literal character ')'
// \] - literal character ']'
// \\ - literal character '\'
// anything else - just text
//
// >>> parse_prompt_attention('normal text')
// [['normal text', 1.0]]
// >>> parse_prompt_attention('an (important) word')
// [['an ', 1.0], ['important', 1.1], [' word', 1.0]]
// >>> parse_prompt_attention('(unbalanced')
// [['unbalanced', 1.1]]
// >>> parse_prompt_attention('\(literal\]')
// [['(literal]', 1.0]]
// >>> parse_prompt_attention('(unnecessary)(parens)')
// [['unnecessaryparens', 1.1]]
// >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).')
// [['a ', 1.0],
// ['house', 1.5730000000000004],
// [' ', 1.1],
// ['on', 1.0],
// [' a ', 1.1],
// ['hill', 0.55],
// [', sun, ', 1.1],
// ['sky', 1.4641000000000006],
// ['.', 1.1]]
std::vector<std::pair<std::string, float>> parse_prompt_attention(const std::string& text) {
std::vector<std::pair<std::string, float>> res;
std::vector<int> round_brackets;
std::vector<int> square_brackets;
float round_bracket_multiplier = 1.1f;
float square_bracket_multiplier = 1 / 1.1f;
std::regex re_attention(R"(\\\(|\\\)|\\\[|\\\]|\\\\|\\|\(|\[|:([+-]?[.\d]+)\)|\)|\]|[^\\()\[\]:]+|:)");
std::regex re_break(R"(\s*\bBREAK\b\s*)");
auto multiply_range = [&](int start_position, float multiplier) {
for (int p = start_position; p < res.size(); ++p) {
res[p].second *= multiplier;
}
};
std::smatch m;
std::string remaining_text = text;
while (std::regex_search(remaining_text, m, re_attention)) {
std::string text = m[0];
std::string weight = m[1];
if (text == "(") {
round_brackets.push_back(res.size());
} else if (text == "[") {
square_brackets.push_back(res.size());
} else if (!weight.empty()) {
if (!round_brackets.empty()) {
multiply_range(round_brackets.back(), std::stod(weight));
round_brackets.pop_back();
}
} else if (text == ")" && !round_brackets.empty()) {
multiply_range(round_brackets.back(), round_bracket_multiplier);
round_brackets.pop_back();
} else if (text == "]" && !square_brackets.empty()) {
multiply_range(square_brackets.back(), square_bracket_multiplier);
square_brackets.pop_back();
} else if (text == "\\(") {
res.push_back({text.substr(1), 1.0f});
} else {
res.push_back({text, 1.0f});
}
remaining_text = m.suffix();
}
for (int pos : round_brackets) {
multiply_range(pos, round_bracket_multiplier);
}
for (int pos : square_brackets) {
multiply_range(pos, square_bracket_multiplier);
}
if (res.empty()) {
res.push_back({"", 1.0f});
}
int i = 0;
while (i + 1 < res.size()) {
if (res[i].second == res[i + 1].second) {
res[i].first += res[i + 1].first;
res.erase(res.begin() + i + 1);
} else {
++i;
}
}
return res;
}
/*================================================ FrozenCLIPEmbedder ================================================*/
struct ResidualAttentionBlock {
int32_t n_head;
int32_t d_model;
int32_t hidden_size; // n_head * d_model
int32_t intermediate_size;
// attention
struct ggml_tensor* q_w; // [hidden_size, hidden_size]
struct ggml_tensor* q_b; // [hidden_size, ]
struct ggml_tensor* k_w; // [hidden_size, hidden_size]
struct ggml_tensor* k_b; // [hidden_size, ]
struct ggml_tensor* v_w; // [hidden_size, hidden_size]
struct ggml_tensor* v_b; // [hidden_size, ]
struct ggml_tensor* out_w; // [hidden_size, hidden_size]
struct ggml_tensor* out_b; // [hidden_size, ]
// layer norm 1
struct ggml_tensor* ln1_w; // [hidden_size, ]
struct ggml_tensor* ln1_b; // [hidden_size, ]
// mlp
struct ggml_tensor* fc1_w; // [intermediate_size, hidden_size]
struct ggml_tensor* fc1_b; // [intermediate_size, ]
struct ggml_tensor* fc2_w; // [hidden_size, intermediate_size]
struct ggml_tensor* fc2_b; // [hidden_size, ]
// layer norm 2
struct ggml_tensor* ln2_w; // [hidden_size, ]
struct ggml_tensor* ln2_b; // [hidden_size, ]
size_t compute_params_mem_size(ggml_type wtype) {
double mem_size = 0;
mem_size += 4 * hidden_size * hidden_size * ggml_type_sizef(wtype); // q_w/k_w/v_w/out_w
mem_size += 8 * hidden_size * ggml_type_sizef(GGML_TYPE_F32); // q_b/k_b/v_b/out_b/ln1_w/ln1_b/ln2_w/ln2_b
mem_size += 2 * hidden_size * intermediate_size * ggml_type_sizef(wtype); // fc1_w/fc2_w
mem_size += intermediate_size * ggml_type_sizef(GGML_TYPE_F32); // fc1_b
mem_size += hidden_size * ggml_type_sizef(GGML_TYPE_F32); // fc2_b
mem_size += 16 * ggml_tensor_overhead(); // tensor overhead
return static_cast<size_t>(mem_size);
}
void init_params(struct ggml_context* ctx, ggml_type wtype) {
ln1_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
ln1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
q_w = ggml_new_tensor_2d(ctx, wtype, hidden_size, hidden_size);
q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
k_w = ggml_new_tensor_2d(ctx, wtype, hidden_size, hidden_size);
k_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
v_w = ggml_new_tensor_2d(ctx, wtype, hidden_size, hidden_size);
v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
out_w = ggml_new_tensor_2d(ctx, wtype, hidden_size, hidden_size);
out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
fc1_w = ggml_new_tensor_2d(ctx, wtype, hidden_size, intermediate_size);
fc1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, intermediate_size);
fc2_w = ggml_new_tensor_2d(ctx, wtype, intermediate_size, hidden_size);
fc2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
ln2_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
ln2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
}
void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
tensors[prefix + "self_attn.q_proj.weight"] = q_w;
tensors[prefix + "self_attn.q_proj.bias"] = q_b;
tensors[prefix + "self_attn.k_proj.weight"] = k_w;
tensors[prefix + "self_attn.k_proj.bias"] = k_b;
tensors[prefix + "self_attn.v_proj.weight"] = v_w;
tensors[prefix + "self_attn.v_proj.bias"] = v_b;
tensors[prefix + "self_attn.out_proj.weight"] = out_w;
tensors[prefix + "self_attn.out_proj.bias"] = out_b;
tensors[prefix + "layer_norm1.weight"] = ln1_w;
tensors[prefix + "layer_norm1.bias"] = ln1_b;
tensors[prefix + "layer_norm2.weight"] = ln2_w;
tensors[prefix + "layer_norm2.bias"] = ln2_b;
tensors[prefix + "mlp.fc1.weight"] = fc1_w;
tensors[prefix + "mlp.fc1.bias"] = fc1_b;
tensors[prefix + "mlp.fc2.weight"] = fc2_w;
tensors[prefix + "mlp.fc2.bias"] = fc2_b;
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
// x: [N, n_token, hidden_size]
int64_t N = x->ne[2];
int64_t n_token = x->ne[1];
int64_t hidden_size = n_head * d_model;
struct ggml_tensor* r = x;
// layer norm 1
{
x = ggml_norm(ctx, x);
x = ggml_add(ctx,
ggml_mul(ctx, ggml_repeat(ctx, ln1_w, x), x),
ggml_repeat(ctx, ln1_b, x));
}
// self-attention
{
struct ggml_tensor* q = ggml_add(ctx,
ggml_repeat(ctx, q_b, x),
ggml_mul_mat(ctx, q_w, x));
q = ggml_scale_inplace(ctx, q, ggml_new_f32(ctx, 1.0f / sqrt((float)d_model)));
q = ggml_reshape_4d(ctx, q, d_model, n_head, n_token, N); // [N, n_token, n_head, d_model]
q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3)); // [N, n_head, n_token, d_model]
q = ggml_reshape_3d(ctx, q, d_model, n_token, n_head * N); // [N * n_head, n_token, d_model]
struct ggml_tensor* k = ggml_add(ctx,
ggml_repeat(ctx, k_b, x),
ggml_mul_mat(ctx, k_w, x));
k = ggml_reshape_4d(ctx, k, d_model, n_head, n_token, N); // [N, n_token, n_head, d_model]
k = ggml_cont(ctx, ggml_permute(ctx, k, 0, 2, 1, 3)); // [N, n_head, n_token, d_model]
k = ggml_reshape_3d(ctx, k, d_model, n_token, n_head); // [N * n_head, n_token, d_model]
struct ggml_tensor* v = ggml_add(ctx,
ggml_repeat(ctx, v_b, x),
ggml_mul_mat(ctx, v_w, x));
v = ggml_reshape_4d(ctx, v, d_model, n_head, n_token, N); // [N, n_token, n_head, d_model]
v = ggml_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3)); // [N, n_head, d_model, n_token]
v = ggml_reshape_3d(ctx, v, n_token, d_model, n_head * N); // [N * n_head, d_model, n_token]
struct ggml_tensor* kq = ggml_mul_mat(ctx, k, q); // [N * n_head, n_token, n_token]
kq = ggml_diag_mask_inf_inplace(ctx, kq, 0);
kq = ggml_soft_max_inplace(ctx, kq);
struct ggml_tensor* kqv = ggml_mul_mat(ctx, v, kq); // [N * n_head, n_token, d_model]
kqv = ggml_reshape_4d(ctx, kqv, d_model, n_token, n_head, N);
kqv = ggml_cont(ctx, ggml_permute(ctx, kqv, 0, 2, 1, 3)); // [N, n_token, n_head, d_model]
x = ggml_reshape_2d(ctx, kqv, d_model * n_head, n_token * N); // // [N * n_token, d_model * n_head]
}
// attention output
x = ggml_add(ctx, ggml_repeat(ctx, out_b, x), ggml_mul_mat(ctx, out_w, x));
// residual
x = ggml_add(ctx, x, r);
r = x;
// layer norm 2
{
x = ggml_norm(ctx, x);
x = ggml_add(ctx, ggml_mul(ctx, ggml_repeat(ctx, ln2_w, x), x),
ggml_repeat(ctx, ln2_b, x));
}
// mlp
x = ggml_mul_mat(ctx, fc1_w, x);
x = ggml_add(ctx, ggml_repeat(ctx, fc1_b, x), x);
x = ggml_gelu_quick_inplace(ctx, x);
x = ggml_mul_mat(ctx, fc2_w, x);
x = ggml_add(ctx, ggml_repeat(ctx, fc2_b, x), x);
// residual 2
x = ggml_add(ctx, x, r);
return x;
}
};
struct CLIPTextModel {
// network hparams
int32_t vocab_size = 49408;
int32_t max_position_embeddings = 77;
int32_t hidden_size = 768;
int32_t intermediate_size = 3072;
int32_t projection_dim = 768;
int32_t n_head = 12; // num_attention_heads
int32_t num_hidden_layers = 12;
// embeddings
struct ggml_tensor* position_ids;
struct ggml_tensor* token_embed_weight;
struct ggml_tensor* position_embed_weight;
// transformer
ResidualAttentionBlock resblocks[12];
struct ggml_tensor* final_ln_w;
struct ggml_tensor* final_ln_b;
CLIPTextModel() {
int d_model = hidden_size / n_head; // 64
for (int i = 0; i < num_hidden_layers; i++) {
resblocks[i].d_model = d_model;
resblocks[i].n_head = n_head;
resblocks[i].hidden_size = hidden_size;
resblocks[i].intermediate_size = intermediate_size;
}
}
size_t compute_params_mem_size(ggml_type wtype) {
double mem_size = 0;
mem_size += hidden_size * max_position_embeddings * ggml_type_sizef(GGML_TYPE_I32); // position_ids
mem_size += hidden_size * vocab_size * ggml_type_sizef(wtype); // token_embed_weight
mem_size += hidden_size * max_position_embeddings * ggml_type_sizef(wtype); // position_embed_weight
for (int i = 0; i < num_hidden_layers; i++) {
mem_size += resblocks[i].compute_params_mem_size(wtype);
}
mem_size += 2 * hidden_size * ggml_type_sizef(GGML_TYPE_F32); // final_ln_w/b
mem_size += ggml_tensor_overhead(); // object overhead
return static_cast<size_t>(mem_size);
}
void init_params(struct ggml_context* ctx, ggml_type wtype) {
position_ids = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, max_position_embeddings);
for (int i = 0; i < max_position_embeddings; i++) {
ggml_set_i32_1d(position_ids, i, i);
}
token_embed_weight = ggml_new_tensor_2d(ctx, wtype, hidden_size, vocab_size);
position_embed_weight = ggml_new_tensor_2d(ctx, wtype, hidden_size, max_position_embeddings);
for (int i = 0; i < num_hidden_layers; i++) {
resblocks[i].init_params(ctx, wtype);
}
final_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
final_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
}
void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
tensors[prefix + "embeddings.token_embedding.weight"] = token_embed_weight;
tensors[prefix + "embeddings.position_embedding.weight"] = position_embed_weight;
tensors[prefix + "final_layer_norm.weight"] = final_ln_w;
tensors[prefix + "final_layer_norm.bias"] = final_ln_b;
for (int i = 0; i < num_hidden_layers; i++) {
resblocks[i].map_by_name(tensors, prefix + "encoder.layers." + std::to_string(i) + ".");
}
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* input_ids) {
// input_ids: [N, n_token]
GGML_ASSERT(input_ids->ne[0] <= position_ids->ne[0]);
// token_embedding + position_embedding
struct ggml_tensor* x;
x = ggml_add(ctx,
ggml_get_rows(ctx, token_embed_weight, input_ids),
ggml_get_rows(ctx,
position_embed_weight,
ggml_view_1d(ctx, position_ids, input_ids->ne[0], 0))); // [N, n_token, hidden_size]
// transformer
for (int i = 0; i < num_hidden_layers; i++) {
x = resblocks[i].forward(ctx, x); // [N, n_token, hidden_size]
}
// final layer norm
{
x = ggml_norm(ctx, x);
x = ggml_add(ctx, ggml_mul(ctx, ggml_repeat(ctx, final_ln_w, x), x),
ggml_repeat(ctx, final_ln_b, x));
}
return x; // [N, n_token, hidden_size]
}
};
// ldm.modules.encoders.modules.FrozenCLIPEmbedder
struct FrozenCLIPEmbedder {
CLIPTokenizer tokenizer;
CLIPTextModel text_model;
struct ggml_tensor* forward(struct ggml_context* ctx, const std::string& prompt) {
std::vector<int32_t> tokens = tokenizer.tokenize(prompt, text_model.max_position_embeddings, true);
struct ggml_tensor* input_ids = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, tokens.size());
memcpy(input_ids->data, tokens.data(), tokens.size() * ggml_element_size(input_ids));
struct ggml_tensor* hidden_states = text_model.forward(ctx, input_ids);
return hidden_states;
}
};
// Ref: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/cad87bf4e3e0b0a759afa94e933527c3123d59bc/modules/sd_hijack_clip.py#L283
struct FrozenCLIPEmbedderWithCustomWords {
CLIPTokenizer tokenizer;
CLIPTextModel text_model;
std::pair<std::vector<int>, std::vector<float>> tokenize(std::string text,
size_t max_length = 0,
bool padding = false) {
auto parsed_attention = parse_prompt_attention(text);
{
std::stringstream ss;
ss << "[";
for (const auto& item : parsed_attention) {
ss << "['" << item.first << "', " << item.second << "], ";
}
ss << "]";
LOG_DEBUG("parse '%s' to %s", text.c_str(), ss.str().c_str());
}
std::vector<int> tokens;
std::vector<float> weights;
for (const auto& item : parsed_attention) {
const std::string& curr_text = item.first;
float curr_weight = item.second;
std::vector<int> curr_tokens = tokenizer.encode(curr_text);
tokens.insert(tokens.end(), curr_tokens.begin(), curr_tokens.end());
weights.insert(weights.end(), curr_tokens.size(), curr_weight);
}
tokens.insert(tokens.begin(), BOS_TOKEN_ID);
weights.insert(weights.begin(), 1.0);
if (max_length > 0) {
if (tokens.size() > max_length - 1) {
tokens.resize(max_length - 1);
weights.resize(max_length - 1);
} else {
if (padding) {
tokens.insert(tokens.end(), max_length - 1 - tokens.size(), PAD_TOKEN_ID);
weights.insert(weights.end(), max_length - 1 - weights.size(), 1.0);
}
}
}
tokens.push_back(EOS_TOKEN_ID);
weights.push_back(1.0);
// for (int i = 0; i < tokens.size(); i++) {
// std::cout << tokens[i] << ":" << weights[i] << ", ";
// }
// std::cout << std::endl;
return {tokens, weights};
}
};
/*==================================================== UnetModel =====================================================*/
struct ResBlock {
// network hparams
int channels; // model_channels * (1, 1, 1, 2, 2, 4, 4, 4)
int emb_channels; // time_embed_dim
int out_channels; // mult * model_channels
// network params
// in_layers
struct ggml_tensor* in_layer_0_w; // [channels, ]
struct ggml_tensor* in_layer_0_b; // [channels, ]
// in_layer_1 is nn.SILU()
struct ggml_tensor* in_layer_2_w; // [out_channels, channels, 3, 3]
struct ggml_tensor* in_layer_2_b; // [out_channels, ]
// emb_layers
// emb_layer_0 is nn.SILU()
struct ggml_tensor* emb_layer_1_w; // [out_channels, emb_channels]
struct ggml_tensor* emb_layer_1_b; // [out_channels, ]
// out_layers
struct ggml_tensor* out_layer_0_w; // [out_channels, ]
struct ggml_tensor* out_layer_0_b; // [out_channels, ]
// out_layer_1 is nn.SILU()
// out_layer_2 is nn.Dropout(), p = 0 for inference
struct ggml_tensor* out_layer_3_w; // [out_channels, out_channels, 3, 3]
struct ggml_tensor* out_layer_3_b; // [out_channels, ]
// skip connection, only if out_channels != channels
struct ggml_tensor* skip_w; // [out_channels, channels, 1, 1]
struct ggml_tensor* skip_b; // [out_channels, ]
size_t compute_params_mem_size(ggml_type wtype) {
double mem_size = 0;
mem_size += 2 * channels * ggml_type_sizef(GGML_TYPE_F32); // in_layer_0_w/b
mem_size += out_channels * channels * 3 * 3 * ggml_type_sizef(GGML_TYPE_F16); // in_layer_2_w
mem_size += 5 * out_channels * ggml_type_sizef(GGML_TYPE_F32); // in_layer_2_b/emb_layer_1_b/out_layer_0_w/out_layer_0_b/out_layer_3_b
mem_size += out_channels * emb_channels * ggml_type_sizef(wtype); // emb_layer_1_w
mem_size += out_channels * out_channels * 3 * 3 * ggml_type_sizef(GGML_TYPE_F16); // out_layer_3_w
mem_size += 10 * ggml_tensor_overhead(); // object overhead
if (out_channels != channels) {
mem_size += out_channels * channels * 1 * 1 * ggml_type_sizef(GGML_TYPE_F16); // skip_w
mem_size += out_channels * ggml_type_sizef(GGML_TYPE_F32); // skip_b
mem_size += 2 * ggml_tensor_overhead(); // object overhead
}
return static_cast<size_t>(mem_size);
}
void init_params(struct ggml_context* ctx, ggml_type wtype) {
in_layer_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, channels);
in_layer_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, channels);
in_layer_2_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, channels, out_channels);
in_layer_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
emb_layer_1_w = ggml_new_tensor_2d(ctx, wtype, emb_channels, out_channels);
emb_layer_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
out_layer_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
out_layer_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
out_layer_3_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, out_channels, out_channels);
out_layer_3_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
if (out_channels != channels) {
skip_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, 1, channels, out_channels);
skip_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
}
}
void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
tensors[prefix + "in_layers.0.weight"] = in_layer_0_w;
tensors[prefix + "in_layers.0.bias"] = in_layer_0_b;
tensors[prefix + "in_layers.2.weight"] = in_layer_2_w;
tensors[prefix + "in_layers.2.bias"] = in_layer_2_b;
tensors[prefix + "emb_layers.1.weight"] = emb_layer_1_w;
tensors[prefix + "emb_layers.1.bias"] = emb_layer_1_b;
tensors[prefix + "out_layers.0.weight"] = out_layer_0_w;
tensors[prefix + "out_layers.0.bias"] = out_layer_0_b;
tensors[prefix + "out_layers.3.weight"] = out_layer_3_w;
tensors[prefix + "out_layers.3.bias"] = out_layer_3_b;
if (out_channels != channels) {
tensors[prefix + "skip_connection.weight"] = skip_w;
tensors[prefix + "skip_connection.bias"] = skip_b;
}
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* emb) {
// x: [N, channels, h, w]
// emb: [N, emb_channels]
// in_layers
// group norm 32
auto h = ggml_group_norm_32(ctx, x);
h = ggml_add(ctx,
ggml_mul(ctx,
ggml_repeat(ctx,
ggml_reshape_4d(ctx, in_layer_0_w, 1, 1, in_layer_0_w->ne[0], 1),
h),
h),
ggml_repeat(ctx,
ggml_reshape_4d(ctx, in_layer_0_b, 1, 1, in_layer_0_b->ne[0], 1),
h));
// silu
h = ggml_silu_inplace(ctx, h);
// conv2d
h = ggml_conv_2d(ctx, in_layer_2_w, h, 1, 1, 1, 1, 1, 1);
h = ggml_add(ctx,
h,
ggml_repeat(ctx,
ggml_reshape_4d(ctx, in_layer_2_b, 1, 1, in_layer_2_b->ne[0], 1),
h)); // [N, out_channels, h, w]
// emb_layers
auto emb_out = ggml_silu(ctx, emb);
emb_out = ggml_mul_mat(ctx, emb_layer_1_w, emb_out);
emb_out = ggml_add(ctx, ggml_repeat(ctx, emb_layer_1_b, emb_out), emb_out); // [N, out_channels]
emb_out = ggml_reshape_4d(ctx, emb_out, 1, 1, emb_out->ne[0], emb_out->ne[1]); // [N, out_channels, 1, 1]
emb_out = ggml_repeat(ctx, emb_out, h); // [N, out_channels, h, w]
// out_layers
h = ggml_add(ctx, h, emb_out);
// group norm 32
h = ggml_group_norm_inplace(ctx, h, 32);
h = ggml_add(ctx,
ggml_mul(ctx, ggml_repeat(ctx, ggml_reshape_4d(ctx, out_layer_0_w, 1, 1, out_layer_0_w->ne[0], 1), h), h),
ggml_repeat(ctx, ggml_reshape_4d(ctx, out_layer_0_b, 1, 1, out_layer_0_b->ne[0], 1), h));
// silu
h = ggml_silu_inplace(ctx, h);
// dropout, skip for inference
// conv2d
h = ggml_conv_2d(ctx, out_layer_3_w, h, 1, 1, 1, 1, 1, 1);
h = ggml_add(ctx,
h,
ggml_repeat(ctx,
ggml_reshape_4d(ctx, out_layer_3_b, 1, 1, out_layer_3_b->ne[0], 1),
h)); // [N, out_channels, h, w
// skip connection
if (out_channels != channels) {
x = ggml_conv_2d(ctx, skip_w, x, 1, 1, 0, 0, 1, 1);
x = ggml_add(ctx,
x,
ggml_repeat(ctx,
ggml_reshape_4d(ctx, skip_b, 1, 1, skip_b->ne[0], 1),
x)); // [N, out_channels, h, w]
}
h = ggml_add(ctx, h, x);
return h; // [N, out_channels, h, w]
}
};
struct SpatialTransformer {
int in_channels; // mult * model_channels
int n_head; // num_heads
int d_head; // in_channels // n_heads
int depth = 1; // 1
int context_dim = 768; // hidden_size
// group norm
struct ggml_tensor* norm_w; // [in_channels,]
struct ggml_tensor* norm_b; // [in_channels,]
// proj_in
struct ggml_tensor* proj_in_w; // [in_channels, in_channels, 1, 1]
struct ggml_tensor* proj_in_b; // [in_channels,]
// transformer
struct
{
// layer norm 1
struct ggml_tensor* norm1_w; // [in_channels, ]
struct ggml_tensor* norm1_b; // [in_channels, ]
// attn1
struct ggml_tensor* attn1_q_w; // [in_channels, in_channels]
struct ggml_tensor* attn1_k_w; // [in_channels, in_channels]
struct ggml_tensor* attn1_v_w; // [in_channels, in_channels]
struct ggml_tensor* attn1_out_w; // [in_channels, in_channels]
struct ggml_tensor* attn1_out_b; // [in_channels, ]
// layer norm 2
struct ggml_tensor* norm2_w; // [in_channels, ]
struct ggml_tensor* norm2_b; // [in_channels, ]
// attn2