-
Notifications
You must be signed in to change notification settings - Fork 310
/
clip.hpp
952 lines (823 loc) · 38.8 KB
/
clip.hpp
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
#ifndef __CLIP_HPP__
#define __CLIP_HPP__
#include "ggml_extend.hpp"
#include "model.h"
/*================================================== CLIPTokenizer ===================================================*/
std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remove_lora(std::string text) {
std::regex re("<lora:([^:]+):([^>]+)>");
std::smatch matches;
std::unordered_map<std::string, float> filename2multiplier;
while (std::regex_search(text, matches, re)) {
std::string filename = matches[1].str();
float multiplier = std::stof(matches[2].str());
text = std::regex_replace(text, re, "", std::regex_constants::format_first_only);
if (multiplier == 0.f) {
continue;
}
if (filename2multiplier.find(filename) == filename2multiplier.end()) {
filename2multiplier[filename] = multiplier;
} else {
filename2multiplier[filename] += multiplier;
}
}
return std::make_pair(filename2multiplier, text);
}
std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
std::vector<std::pair<int, std::u32string>> byte_unicode_pairs;
std::set<int> byte_set;
for (int b = static_cast<int>('!'); b <= static_cast<int>('~'); ++b) {
byte_set.insert(b);
byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
}
for (int b = 161; b <= 172; ++b) {
byte_set.insert(b);
byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
}
for (int b = 174; b <= 255; ++b) {
byte_set.insert(b);
byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
}
int n = 0;
for (int b = 0; b < 256; ++b) {
if (byte_set.find(b) == byte_set.end()) {
byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(n + 256)));
++n;
}
}
// LOG_DEBUG("byte_unicode_pairs %d", byte_unicode_pairs.size());
return byte_unicode_pairs;
}
// Ref: https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py
typedef std::function<bool(std::string&, std::vector<int32_t>&)> on_new_token_cb_t;
class CLIPTokenizer {
private:
std::map<int, std::u32string> byte_encoder;
std::map<std::u32string, int> byte_decoder;
std::map<std::u32string, int> encoder;
std::map<int, std::u32string> decoder;
std::map<std::pair<std::u32string, std::u32string>, int> bpe_ranks;
std::regex pat;
int encoder_len;
int bpe_len;
public:
const std::string UNK_TOKEN = "<|endoftext|>";
const std::string BOS_TOKEN = "<|startoftext|>";
const std::string EOS_TOKEN = "<|endoftext|>";
const std::string PAD_TOKEN = "<|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;
private:
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;
}
static std::set<std::pair<std::u32string, std::u32string>> get_pairs(const std::vector<std::u32string>& subwords) {
std::set<std::pair<std::u32string, std::u32string>> pairs;
if (subwords.size() == 0) {
return pairs;
}
std::u32string prev_subword = subwords[0];
for (int i = 1; i < subwords.size(); i++) {
std::u32string subword = subwords[i];
std::pair<std::u32string, std::u32string> pair(prev_subword, subword);
pairs.insert(pair);
prev_subword = subword;
}
return pairs;
}
public:
CLIPTokenizer(int pad_token_id = 49407, const std::string& merges_utf8_str = "")
: PAD_TOKEN_ID(pad_token_id) {
if (merges_utf8_str.size() > 0) {
load_from_merges(merges_utf8_str);
} else {
load_from_merges(ModelLoader::load_merges());
}
}
void load_from_merges(const std::string& merges_utf8_str) {
auto byte_unicode_pairs = bytes_to_unicode();
// printf("byte_unicode_pairs have %lu pairs \n", byte_unicode_pairs.size());
byte_encoder = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
for (auto& pair : byte_unicode_pairs) {
byte_decoder[pair.second] = pair.first;
}
// for (auto & pair: byte_unicode_pairs) {
// std::cout << pair.first << ": " << pair.second << std::endl;
// }
std::vector<std::u32string> merges;
size_t start = 0;
size_t pos;
std::u32string merges_utf32_str = utf8_to_utf32(merges_utf8_str);
while ((pos = merges_utf32_str.find('\n', start)) != std::string::npos) {
merges.push_back(merges_utf32_str.substr(start, pos - start));
start = pos + 1;
}
// LOG_DEBUG("merges size %llu", merges.size());
GGML_ASSERT(merges.size() == 48895);
merges = std::vector<std::u32string>(merges.begin() + 1, merges.end());
std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
for (const auto& merge : merges) {
size_t space_pos = merge.find(' ');
merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
// LOG_DEBUG("%s", utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
// printf("%s :: %s | %s \n", utf32_to_utf8(merge).c_str(), utf32_to_utf8(merge.substr(0, space_pos)).c_str(),
// utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
}
std::vector<std::u32string> vocab;
for (const auto& pair : byte_unicode_pairs) {
vocab.push_back(pair.second);
}
for (const auto& pair : byte_unicode_pairs) {
vocab.push_back(pair.second + utf8_to_utf32("</w>"));
}
for (const auto& merge : merge_pairs) {
vocab.push_back(merge.first + merge.second);
}
vocab.push_back(utf8_to_utf32("<|startoftext|>"));
vocab.push_back(utf8_to_utf32("<|endoftext|>"));
LOG_DEBUG("vocab size: %llu", vocab.size());
int i = 0;
for (const auto& token : vocab) {
encoder[token] = i;
decoder[i] = token;
i++;
}
encoder_len = i;
auto it = encoder.find(utf8_to_utf32("img</w>"));
if (it != encoder.end()) {
LOG_DEBUG(" trigger word img already in vocab");
} else {
LOG_DEBUG(" trigger word img not in vocab yet");
}
int rank = 0;
for (const auto& merge : merge_pairs) {
bpe_ranks[merge] = rank++;
}
bpe_len = rank;
};
void add_token(const std::string& text) {
std::u32string token = utf8_to_utf32(text);
auto it = encoder.find(token);
if (it != encoder.end()) {
encoder[token] = encoder_len;
decoder[encoder_len] = token;
encoder_len++;
}
}
std::u32string bpe(const std::u32string& token) {
std::vector<std::u32string> word;
for (int i = 0; i < token.size() - 1; i++) {
word.emplace_back(1, token[i]);
}
word.push_back(token.substr(token.size() - 1) + utf8_to_utf32("</w>"));
std::set<std::pair<std::u32string, std::u32string>> pairs = get_pairs(word);
if (pairs.empty()) {
return token + utf8_to_utf32("</w>");
}
while (true) {
auto min_pair_iter = std::min_element(pairs.begin(),
pairs.end(),
[&](const std::pair<std::u32string, std::u32string>& a,
const std::pair<std::u32string, std::u32string>& b) {
if (bpe_ranks.find(a) == bpe_ranks.end()) {
return false;
} else if (bpe_ranks.find(b) == bpe_ranks.end()) {
return true;
}
return bpe_ranks.at(a) < bpe_ranks.at(b);
});
const std::pair<std::u32string, std::u32string>& bigram = *min_pair_iter;
if (bpe_ranks.find(bigram) == bpe_ranks.end()) {
break;
}
std::u32string first = bigram.first;
std::u32string second = bigram.second;
std::vector<std::u32string> new_word;
int32_t i = 0;
while (i < word.size()) {
auto it = std::find(word.begin() + i, word.end(), first);
if (it == word.end()) {
new_word.insert(new_word.end(), word.begin() + i, word.end());
break;
}
new_word.insert(new_word.end(), word.begin() + i, it);
i = static_cast<int32_t>(std::distance(word.begin(), it));
if (word[i] == first && i < static_cast<int32_t>(word.size()) - 1 && word[i + 1] == second) {
new_word.push_back(first + second);
i += 2;
} else {
new_word.push_back(word[i]);
i += 1;
}
}
word = new_word;
if (word.size() == 1) {
break;
}
pairs = get_pairs(word);
}
std::u32string result;
for (int i = 0; i < word.size(); i++) {
result += word[i];
if (i != word.size() - 1) {
result += utf8_to_utf32(" ");
}
}
return result;
}
std::vector<int> tokenize(std::string text,
on_new_token_cb_t on_new_token_cb,
size_t max_length = 0,
bool padding = false) {
std::vector<int32_t> tokens = encode(text, on_new_token_cb);
tokens.insert(tokens.begin(), BOS_TOKEN_ID);
if (max_length > 0) {
if (tokens.size() > max_length - 1) {
tokens.resize(max_length - 1);
tokens.push_back(EOS_TOKEN_ID);
} else {
tokens.push_back(EOS_TOKEN_ID);
if (padding) {
tokens.insert(tokens.end(), max_length - tokens.size(), PAD_TOKEN_ID);
}
}
}
return tokens;
}
void pad_tokens(std::vector<int>& tokens,
std::vector<float>& weights,
size_t max_length = 0,
bool padding = false) {
if (max_length > 0 && padding) {
size_t n = std::ceil(tokens.size() * 1.0 / (max_length - 2));
if (n == 0) {
n = 1;
}
size_t length = max_length * n;
LOG_DEBUG("token length: %llu", length);
std::vector<int> new_tokens;
std::vector<float> new_weights;
new_tokens.push_back(BOS_TOKEN_ID);
new_weights.push_back(1.0);
int token_idx = 0;
for (int i = 1; i < length; i++) {
if (token_idx >= tokens.size()) {
break;
}
if (i % max_length == 0) {
new_tokens.push_back(BOS_TOKEN_ID);
new_weights.push_back(1.0);
} else if (i % max_length == max_length - 1) {
new_tokens.push_back(EOS_TOKEN_ID);
new_weights.push_back(1.0);
} else {
new_tokens.push_back(tokens[token_idx]);
new_weights.push_back(weights[token_idx]);
token_idx++;
}
}
new_tokens.push_back(EOS_TOKEN_ID);
new_weights.push_back(1.0);
tokens = new_tokens;
weights = new_weights;
if (padding) {
tokens.insert(tokens.end(), length - tokens.size(), PAD_TOKEN_ID);
weights.insert(weights.end(), length - weights.size(), 1.0);
}
}
}
std::string clean_up_tokenization(std::string& text) {
std::regex pattern(R"( ,)");
// Replace " ," with ","
std::string result = std::regex_replace(text, pattern, ",");
return result;
}
std::string decode(const std::vector<int>& tokens) {
std::string text = "";
for (int t : tokens) {
if (t == 49406 || t == 49407)
continue;
std::u32string ts = decoder[t];
// printf("%d, %s \n", t, utf32_to_utf8(ts).c_str());
std::string s = utf32_to_utf8(ts);
if (s.length() >= 4) {
if (ends_with(s, "</w>")) {
text += s.replace(s.length() - 4, s.length() - 1, "") + " ";
} else {
text += s;
}
} else {
text += " " + s;
}
}
// std::vector<unsigned char> bytes;
// for (auto c : text){
// bytes.push_back(byte_decoder[c]);
// }
// std::string s((char *)bytes.data());
// std::string s = "";
text = clean_up_tokenization(text);
return trim(text);
}
std::vector<int> encode(std::string text, on_new_token_cb_t on_new_token_cb) {
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)) {
bool skip = on_new_token_cb(str, bpe_tokens);
if (skip) {
continue;
}
for (auto& token : matches) {
std::string token_str = token.str();
std::u32string utf32_token;
for (int i = 0; i < token_str.length(); i++) {
unsigned char b = token_str[i];
utf32_token += byte_encoder[b];
}
auto bpe_strs = bpe(utf32_token);
size_t start = 0;
size_t pos;
while ((pos = bpe_strs.find(' ', start)) != std::u32string::npos) {
auto bpe_str = bpe_strs.substr(start, pos - start);
bpe_tokens.push_back(encoder[bpe_str]);
token_strs.push_back(utf32_to_utf8(bpe_str));
start = pos + 1;
}
auto bpe_str = bpe_strs.substr(start, bpe_strs.size() - start);
bpe_tokens.push_back(encoder[bpe_str]);
token_strs.push_back(utf32_to_utf8(bpe_str));
}
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());
// printf("split prompt \"%s\" to tokens %s \n", original_text.c_str(), ss.str().c_str());
return bpe_tokens;
}
};
/*================================================ FrozenCLIPEmbedder ================================================*/
// Ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/clip/modeling_clip.py
struct CLIPMLP : public GGMLBlock {
protected:
bool use_gelu;
public:
CLIPMLP(int64_t d_model, int64_t intermediate_size) {
blocks["fc1"] = std::shared_ptr<GGMLBlock>(new Linear(d_model, intermediate_size));
blocks["fc2"] = std::shared_ptr<GGMLBlock>(new Linear(intermediate_size, d_model));
if (d_model == 1024 || d_model == 1280) { // SD 2.x
use_gelu = true;
} else { // SD 1.x
use_gelu = false;
}
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
// x: [N, n_token, d_model]
auto fc1 = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
auto fc2 = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
x = fc1->forward(ctx, x);
if (use_gelu) {
x = ggml_gelu_inplace(ctx, x);
} else {
x = ggml_gelu_quick_inplace(ctx, x);
}
x = fc2->forward(ctx, x);
return x;
}
};
struct CLIPLayer : public GGMLBlock {
protected:
int64_t d_model; // hidden_size/embed_dim
int64_t n_head;
int64_t intermediate_size;
public:
CLIPLayer(int64_t d_model,
int64_t n_head,
int64_t intermediate_size)
: d_model(d_model),
n_head(n_head),
intermediate_size(intermediate_size) {
blocks["self_attn"] = std::shared_ptr<GGMLBlock>(new MultiheadAttention(d_model, n_head, true, true));
blocks["layer_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_model));
blocks["layer_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_model));
blocks["mlp"] = std::shared_ptr<GGMLBlock>(new CLIPMLP(d_model, intermediate_size));
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, bool mask = true) {
// x: [N, n_token, d_model]
auto self_attn = std::dynamic_pointer_cast<MultiheadAttention>(blocks["self_attn"]);
auto layer_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm1"]);
auto layer_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm2"]);
auto mlp = std::dynamic_pointer_cast<CLIPMLP>(blocks["mlp"]);
x = ggml_add(ctx, x, self_attn->forward(ctx, layer_norm1->forward(ctx, x), mask));
x = ggml_add(ctx, x, mlp->forward(ctx, layer_norm2->forward(ctx, x)));
return x;
}
};
struct CLIPEncoder : public GGMLBlock {
protected:
int64_t n_layer;
public:
CLIPEncoder(int64_t n_layer,
int64_t d_model,
int64_t n_head,
int64_t intermediate_size)
: n_layer(n_layer) {
for (int i = 0; i < n_layer; i++) {
std::string name = "layers." + std::to_string(i);
blocks[name] = std::shared_ptr<GGMLBlock>(new CLIPLayer(d_model, n_head, intermediate_size));
}
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, int clip_skip = -1, bool mask = true) {
// x: [N, n_token, d_model]
int layer_idx = n_layer - 1;
// LOG_DEBUG("clip_skip %d", clip_skip);
if (clip_skip > 0) {
layer_idx = n_layer - clip_skip;
}
for (int i = 0; i < n_layer; i++) {
// LOG_DEBUG("layer %d", i);
if (i == layer_idx + 1) {
break;
}
std::string name = "layers." + std::to_string(i);
auto layer = std::dynamic_pointer_cast<CLIPLayer>(blocks[name]);
x = layer->forward(ctx, x, mask); // [N, n_token, d_model]
// LOG_DEBUG("layer %d", i);
}
return x;
}
};
class CLIPEmbeddings : public GGMLBlock {
protected:
int64_t embed_dim;
int64_t vocab_size;
int64_t num_positions;
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
enum ggml_type token_wtype = (tensor_types.find(prefix + "token_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "token_embedding.weight"] : GGML_TYPE_F32;
enum ggml_type position_wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "position_embedding.weight"] : GGML_TYPE_F32;
params["token_embedding.weight"] = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
}
public:
CLIPEmbeddings(int64_t embed_dim,
int64_t vocab_size = 49408,
int64_t num_positions = 77)
: embed_dim(embed_dim),
vocab_size(vocab_size),
num_positions(num_positions) {
}
struct ggml_tensor* get_token_embed_weight() {
return params["token_embedding.weight"];
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* input_ids,
struct ggml_tensor* custom_embed_weight) {
// input_ids: [N, n_token]
auto token_embed_weight = params["token_embedding.weight"];
auto position_embed_weight = params["position_embedding.weight"];
GGML_ASSERT(input_ids->ne[0] == position_embed_weight->ne[1]);
input_ids = ggml_reshape_3d(ctx, input_ids, input_ids->ne[0], 1, input_ids->ne[1]);
auto token_embedding = ggml_get_rows(ctx, custom_embed_weight != NULL ? custom_embed_weight : token_embed_weight, input_ids);
token_embedding = ggml_reshape_3d(ctx, token_embedding, token_embedding->ne[0], token_embedding->ne[1], token_embedding->ne[3]);
// token_embedding + position_embedding
auto x = ggml_add(ctx,
token_embedding,
position_embed_weight); // [N, n_token, embed_dim]
return x;
}
};
class CLIPVisionEmbeddings : public GGMLBlock {
protected:
int64_t embed_dim;
int64_t num_channels;
int64_t patch_size;
int64_t image_size;
int64_t num_patches;
int64_t num_positions;
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
enum ggml_type patch_wtype = GGML_TYPE_F16; // tensor_types.find(prefix + "patch_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "patch_embedding.weight"] : GGML_TYPE_F16;
enum ggml_type class_wtype = GGML_TYPE_F32; // tensor_types.find(prefix + "class_embedding") != tensor_types.end() ? tensor_types[prefix + "class_embedding"] : GGML_TYPE_F32;
enum ggml_type position_wtype = GGML_TYPE_F32; // tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "position_embedding.weight"] : GGML_TYPE_F32;
params["patch_embedding.weight"] = ggml_new_tensor_4d(ctx, patch_wtype, patch_size, patch_size, num_channels, embed_dim);
params["class_embedding"] = ggml_new_tensor_1d(ctx, class_wtype, embed_dim);
params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
}
public:
CLIPVisionEmbeddings(int64_t embed_dim,
int64_t num_channels = 3,
int64_t patch_size = 14,
int64_t image_size = 224)
: embed_dim(embed_dim),
num_channels(num_channels),
patch_size(patch_size),
image_size(image_size) {
num_patches = (image_size / patch_size) * (image_size / patch_size);
num_positions = num_patches + 1;
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* pixel_values) {
// pixel_values: [N, num_channels, image_size, image_size]
// return: [N, num_positions, embed_dim]
GGML_ASSERT(pixel_values->ne[0] == image_size && pixel_values->ne[1] == image_size && pixel_values->ne[2] == num_channels);
auto patch_embed_weight = params["patch_embedding.weight"];
auto class_embed_weight = params["class_embedding"];
auto position_embed_weight = params["position_embedding.weight"];
// concat(patch_embedding, class_embedding) + position_embedding
struct ggml_tensor* patch_embedding;
int64_t N = pixel_values->ne[3];
patch_embedding = ggml_nn_conv_2d(ctx, pixel_values, patch_embed_weight, NULL, patch_size, patch_size); // [N, embed_dim, image_size // pacht_size, image_size // pacht_size]
patch_embedding = ggml_reshape_3d(ctx, patch_embedding, num_patches, embed_dim, N); // [N, embed_dim, num_patches]
patch_embedding = ggml_cont(ctx, ggml_permute(ctx, patch_embedding, 1, 0, 2, 3)); // [N, num_patches, embed_dim]
patch_embedding = ggml_reshape_4d(ctx, patch_embedding, 1, embed_dim, num_patches, N); // [N, num_patches, embed_dim, 1]
struct ggml_tensor* class_embedding = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, embed_dim, N);
class_embedding = ggml_repeat(ctx, class_embed_weight, class_embedding); // [N, embed_dim]
class_embedding = ggml_reshape_4d(ctx, class_embedding, 1, embed_dim, 1, N); // [N, 1, embed_dim, 1]
struct ggml_tensor* x = ggml_concat(ctx, class_embedding, patch_embedding, 2); // [N, num_positions, embed_dim, 1]
x = ggml_reshape_3d(ctx, x, embed_dim, num_positions, N); // [N, num_positions, embed_dim]
x = ggml_add(ctx, x, position_embed_weight);
return x; // [N, num_positions, embed_dim]
}
};
// OPENAI_CLIP_VIT_L_14: https://huggingface.co/openai/clip-vit-large-patch14/blob/main/config.json
// OPEN_CLIP_VIT_H_14: https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/blob/main/config.json
// OPEN_CLIP_VIT_BIGG_14: https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k/blob/main/config.json (CLIPTextModelWithProjection)
enum CLIPVersion {
OPENAI_CLIP_VIT_L_14, // SD 1.x and SDXL
OPEN_CLIP_VIT_H_14, // SD 2.x
OPEN_CLIP_VIT_BIGG_14, // SDXL
};
class CLIPTextModel : public GGMLBlock {
protected:
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
if (version == OPEN_CLIP_VIT_BIGG_14) {
enum ggml_type wtype = GGML_TYPE_F32; // tensor_types.find(prefix + "text_projection") != tensor_types.end() ? tensor_types[prefix + "text_projection"] : GGML_TYPE_F32;
params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
}
}
public:
CLIPVersion version = OPENAI_CLIP_VIT_L_14;
// network hparams
int32_t vocab_size = 49408;
int32_t n_token = 77; // max_position_embeddings
int32_t hidden_size = 768;
int32_t intermediate_size = 3072;
int32_t n_head = 12;
int32_t n_layer = 12; // num_hidden_layers
int32_t projection_dim = 1280; // only for OPEN_CLIP_VIT_BIGG_14
int32_t clip_skip = -1;
bool with_final_ln = true;
CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
int clip_skip_value = -1,
bool with_final_ln = true)
: version(version), with_final_ln(with_final_ln) {
if (version == OPEN_CLIP_VIT_H_14) {
hidden_size = 1024;
intermediate_size = 4096;
n_head = 16;
n_layer = 24;
} else if (version == OPEN_CLIP_VIT_BIGG_14) { // CLIPTextModelWithProjection
hidden_size = 1280;
intermediate_size = 5120;
n_head = 20;
n_layer = 32;
}
set_clip_skip(clip_skip_value);
blocks["embeddings"] = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token));
blocks["encoder"] = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
}
void set_clip_skip(int skip) {
if (skip <= 0) {
return;
}
clip_skip = skip;
}
struct ggml_tensor* get_token_embed_weight() {
auto embeddings = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
return embeddings->get_token_embed_weight();
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* input_ids,
struct ggml_tensor* tkn_embeddings,
size_t max_token_idx = 0,
bool return_pooled = false) {
// input_ids: [N, n_token]
auto embeddings = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
auto encoder = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
auto final_layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["final_layer_norm"]);
auto x = embeddings->forward(ctx, input_ids, tkn_embeddings); // [N, n_token, hidden_size]
x = encoder->forward(ctx, x, return_pooled ? -1 : clip_skip, true);
if (return_pooled || with_final_ln) {
x = final_layer_norm->forward(ctx, x);
}
if (return_pooled) {
auto text_projection = params["text_projection"];
ggml_tensor* pooled = ggml_view_1d(ctx, x, hidden_size, x->nb[1] * max_token_idx);
if (text_projection != NULL) {
pooled = ggml_nn_linear(ctx, pooled, text_projection, NULL);
} else {
LOG_DEBUG("Missing text_projection matrix, assuming identity...");
}
return pooled; // [hidden_size, 1, 1]
}
return x; // [N, n_token, hidden_size]
}
};
class CLIPVisionModel : public GGMLBlock {
public:
// network hparams
int32_t num_channels = 3;
int32_t patch_size = 14;
int32_t image_size = 224;
int32_t num_positions = 257; // (image_size / patch_size)^2 + 1
int32_t hidden_size = 1024;
int32_t intermediate_size = 4096;
int32_t n_head = 16;
int32_t n_layer = 24;
public:
CLIPVisionModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14) {
if (version == OPEN_CLIP_VIT_H_14) {
hidden_size = 1280;
intermediate_size = 5120;
n_head = 16;
n_layer = 32;
} else if (version == OPEN_CLIP_VIT_BIGG_14) {
hidden_size = 1664;
intermediate_size = 8192;
n_head = 16;
n_layer = 48;
}
blocks["embeddings"] = std::shared_ptr<GGMLBlock>(new CLIPVisionEmbeddings(hidden_size, num_channels, patch_size, image_size));
blocks["pre_layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
blocks["encoder"] = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
blocks["post_layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* pixel_values, bool return_pooled = true) {
// pixel_values: [N, num_channels, image_size, image_size]
auto embeddings = std::dynamic_pointer_cast<CLIPVisionEmbeddings>(blocks["embeddings"]);
auto pre_layernorm = std::dynamic_pointer_cast<LayerNorm>(blocks["pre_layernorm"]);
auto encoder = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
auto post_layernorm = std::dynamic_pointer_cast<LayerNorm>(blocks["post_layernorm"]);
auto x = embeddings->forward(ctx, pixel_values); // [N, num_positions, embed_dim]
x = pre_layernorm->forward(ctx, x);
x = encoder->forward(ctx, x, -1, false);
// print_ggml_tensor(x, true, "ClipVisionModel x: ");
auto last_hidden_state = x;
x = post_layernorm->forward(ctx, x); // [N, n_token, hidden_size]
GGML_ASSERT(x->ne[3] == 1);
if (return_pooled) {
ggml_tensor* pooled = ggml_cont(ctx, ggml_view_2d(ctx, x, x->ne[0], x->ne[2], x->nb[2], 0));
return pooled; // [N, hidden_size]
} else {
// return x; // [N, n_token, hidden_size]
return last_hidden_state; // [N, n_token, hidden_size]
}
}
};
class CLIPProjection : public UnaryBlock {
protected:
int64_t in_features;
int64_t out_features;
bool transpose_weight;
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
enum ggml_type wtype = tensor_types.find(prefix + "weight") != tensor_types.end() ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
if (transpose_weight) {
params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
} else {
params["weight"] = ggml_new_tensor_2d(ctx, wtype, in_features, out_features);
}
}
public:
CLIPProjection(int64_t in_features,
int64_t out_features,
bool transpose_weight = false)
: in_features(in_features),
out_features(out_features),
transpose_weight(transpose_weight) {}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
struct ggml_tensor* w = params["weight"];
if (transpose_weight) {
w = ggml_cont(ctx, ggml_transpose(ctx, w));
}
return ggml_nn_linear(ctx, x, w, NULL);
}
};
class CLIPVisionModelProjection : public GGMLBlock {
public:
int32_t hidden_size = 1024;
int32_t projection_dim = 768;
int32_t image_size = 224;
public:
CLIPVisionModelProjection(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
bool transpose_proj_w = false) {
if (version == OPEN_CLIP_VIT_H_14) {
hidden_size = 1280;
projection_dim = 1024;
} else if (version == OPEN_CLIP_VIT_BIGG_14) {
hidden_size = 1664;
}
blocks["vision_model"] = std::shared_ptr<GGMLBlock>(new CLIPVisionModel(version));
blocks["visual_projection"] = std::shared_ptr<GGMLBlock>(new CLIPProjection(hidden_size, projection_dim, transpose_proj_w));
}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* pixel_values) {
// pixel_values: [N, num_channels, image_size, image_size]
// return: [N, projection_dim]
auto vision_model = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
auto visual_projection = std::dynamic_pointer_cast<CLIPProjection>(blocks["visual_projection"]);
auto x = vision_model->forward(ctx, pixel_values); // [N, hidden_size]
x = visual_projection->forward(ctx, x); // [N, projection_dim]
return x; // [N, projection_dim]
}
};
struct CLIPTextModelRunner : public GGMLRunner {
CLIPTextModel model;
CLIPTextModelRunner(ggml_backend_t backend,
std::map<std::string, enum ggml_type>& tensor_types,
const std::string prefix,
CLIPVersion version = OPENAI_CLIP_VIT_L_14,
int clip_skip_value = 1,
bool with_final_ln = true)
: GGMLRunner(backend), model(version, clip_skip_value, with_final_ln) {
model.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() {
return "clip";
}
void set_clip_skip(int clip_skip) {
model.set_clip_skip(clip_skip);
}
void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
model.get_param_tensors(tensors, prefix);
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* input_ids,
struct ggml_tensor* embeddings,
size_t max_token_idx = 0,
bool return_pooled = false) {
size_t N = input_ids->ne[1];
size_t n_token = input_ids->ne[0];
if (input_ids->ne[0] > model.n_token) {
GGML_ASSERT(input_ids->ne[0] % model.n_token == 0);
input_ids = ggml_reshape_2d(ctx, input_ids, model.n_token, input_ids->ne[0] / model.n_token);
}
return model.forward(ctx, input_ids, embeddings, max_token_idx, return_pooled);
}
struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
int num_custom_embeddings = 0,
void* custom_embeddings_data = NULL,
size_t max_token_idx = 0,
bool return_pooled = false) {
struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
input_ids = to_backend(input_ids);
struct ggml_tensor* embeddings = NULL;
if (num_custom_embeddings > 0 && custom_embeddings_data != NULL) {
auto token_embed_weight = model.get_token_embed_weight();
auto custom_embeddings = ggml_new_tensor_2d(compute_ctx,
token_embed_weight->type,
model.hidden_size,
num_custom_embeddings);
set_backend_tensor_data(custom_embeddings, custom_embeddings_data);
// concatenate custom embeddings
embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 1);
}
struct ggml_tensor* hidden_states = forward(compute_ctx, input_ids, embeddings, max_token_idx, return_pooled);
ggml_build_forward_expand(gf, hidden_states);
return gf;
}
void compute(const int n_threads,
struct ggml_tensor* input_ids,
int num_custom_embeddings,
void* custom_embeddings_data,
size_t max_token_idx,
bool return_pooled,
ggml_tensor** output,
ggml_context* output_ctx = NULL) {
auto get_graph = [&]() -> struct ggml_cgraph* {
return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled);
};
GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
}
};
#endif // __CLIP_HPP__