-
Notifications
You must be signed in to change notification settings - Fork 366
/
Copy pathstable-diffusion.cpp
1799 lines (1620 loc) · 77.8 KB
/
stable-diffusion.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
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 "ggml_extend.hpp"
#include "model.h"
#include "rng.hpp"
#include "rng_philox.hpp"
#include "stable-diffusion.h"
#include "util.h"
#include "conditioner.hpp"
#include "control.hpp"
#include "denoiser.hpp"
#include "diffusion_model.hpp"
#include "esrgan.hpp"
#include "lora.hpp"
#include "pmid.hpp"
#include "tae.hpp"
#include "vae.hpp"
#define STB_IMAGE_IMPLEMENTATION
#define STB_IMAGE_STATIC
#include "stb_image.h"
// #define STB_IMAGE_WRITE_IMPLEMENTATION
// #define STB_IMAGE_WRITE_STATIC
// #include "stb_image_write.h"
const char* model_version_to_str[] = {
"SD 1.x",
"SD 2.x",
"SDXL",
"SVD",
"SD3.x",
"Flux"};
const char* sampling_methods_str[] = {
"Euler A",
"Euler",
"Heun",
"DPM2",
"DPM++ (2s)",
"DPM++ (2M)",
"modified DPM++ (2M)",
"iPNDM",
"iPNDM_v",
"LCM",
};
/*================================================== Helper Functions ================================================*/
void calculate_alphas_cumprod(float* alphas_cumprod,
float linear_start = 0.00085f,
float linear_end = 0.0120,
int timesteps = TIMESTEPS) {
float ls_sqrt = sqrtf(linear_start);
float le_sqrt = sqrtf(linear_end);
float amount = le_sqrt - ls_sqrt;
float product = 1.0f;
for (int i = 0; i < timesteps; i++) {
float beta = ls_sqrt + amount * ((float)i / (timesteps - 1));
product *= 1.0f - powf(beta, 2.0f);
alphas_cumprod[i] = product;
}
}
/*=============================================== StableDiffusionGGML ================================================*/
class StableDiffusionGGML {
public:
ggml_backend_t backend = NULL; // general backend
ggml_backend_t clip_backend = NULL;
ggml_backend_t control_net_backend = NULL;
ggml_backend_t vae_backend = NULL;
ggml_type model_wtype = GGML_TYPE_COUNT;
ggml_type conditioner_wtype = GGML_TYPE_COUNT;
ggml_type diffusion_model_wtype = GGML_TYPE_COUNT;
ggml_type vae_wtype = GGML_TYPE_COUNT;
SDVersion version;
bool vae_decode_only = false;
bool free_params_immediately = false;
std::shared_ptr<RNG> rng = std::make_shared<STDDefaultRNG>();
int n_threads = -1;
float scale_factor = 0.18215f;
std::shared_ptr<Conditioner> cond_stage_model;
std::shared_ptr<FrozenCLIPVisionEmbedder> clip_vision; // for svd
std::shared_ptr<DiffusionModel> diffusion_model;
std::shared_ptr<AutoEncoderKL> first_stage_model;
std::shared_ptr<TinyAutoEncoder> tae_first_stage;
std::shared_ptr<ControlNet> control_net;
std::shared_ptr<PhotoMakerIDEncoder> pmid_model;
std::shared_ptr<LoraModel> pmid_lora;
std::shared_ptr<PhotoMakerIDEmbed> pmid_id_embeds;
std::string taesd_path;
bool use_tiny_autoencoder = false;
bool vae_tiling = false;
bool stacked_id = false;
std::map<std::string, struct ggml_tensor*> tensors;
std::string lora_model_dir;
// lora_name => multiplier
std::unordered_map<std::string, float> curr_lora_state;
std::shared_ptr<Denoiser> denoiser = std::make_shared<CompVisDenoiser>();
StableDiffusionGGML() = default;
StableDiffusionGGML(int n_threads,
bool vae_decode_only,
bool free_params_immediately,
std::string lora_model_dir,
rng_type_t rng_type)
: n_threads(n_threads),
vae_decode_only(vae_decode_only),
free_params_immediately(free_params_immediately),
lora_model_dir(lora_model_dir) {
if (rng_type == STD_DEFAULT_RNG) {
rng = std::make_shared<STDDefaultRNG>();
} else if (rng_type == CUDA_RNG) {
rng = std::make_shared<PhiloxRNG>();
}
}
~StableDiffusionGGML() {
if (clip_backend != backend) {
ggml_backend_free(clip_backend);
}
if (control_net_backend != backend) {
ggml_backend_free(control_net_backend);
}
if (vae_backend != backend) {
ggml_backend_free(vae_backend);
}
ggml_backend_free(backend);
}
bool load_from_file(const std::string& model_path,
const std::string& clip_l_path,
const std::string& clip_g_path,
const std::string& t5xxl_path,
const std::string& diffusion_model_path,
const std::string& vae_path,
const std::string control_net_path,
const std::string embeddings_path,
const std::string id_embeddings_path,
const std::string& taesd_path,
bool vae_tiling_,
ggml_type wtype,
schedule_t schedule,
bool clip_on_cpu,
bool control_net_cpu,
bool vae_on_cpu,
bool diffusion_flash_attn) {
use_tiny_autoencoder = taesd_path.size() > 0;
#ifdef SD_USE_CUBLAS
LOG_DEBUG("Using CUDA backend");
backend = ggml_backend_cuda_init(0);
#endif
#ifdef SD_USE_METAL
LOG_DEBUG("Using Metal backend");
ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
backend = ggml_backend_metal_init();
#endif
#ifdef SD_USE_VULKAN
LOG_DEBUG("Using Vulkan backend");
for (int device = 0; device < ggml_backend_vk_get_device_count(); ++device) {
backend = ggml_backend_vk_init(device);
}
if (!backend) {
LOG_WARN("Failed to initialize Vulkan backend");
}
#endif
#ifdef SD_USE_SYCL
LOG_DEBUG("Using SYCL backend");
backend = ggml_backend_sycl_init(0);
#endif
if (!backend) {
LOG_DEBUG("Using CPU backend");
backend = ggml_backend_cpu_init();
}
ModelLoader model_loader;
vae_tiling = vae_tiling_;
if (model_path.size() > 0) {
LOG_INFO("loading model from '%s'", model_path.c_str());
if (!model_loader.init_from_file(model_path)) {
LOG_ERROR("init model loader from file failed: '%s'", model_path.c_str());
}
}
if (clip_l_path.size() > 0) {
LOG_INFO("loading clip_l from '%s'", clip_l_path.c_str());
if (!model_loader.init_from_file(clip_l_path, "text_encoders.clip_l.transformer.")) {
LOG_WARN("loading clip_l from '%s' failed", clip_l_path.c_str());
}
}
if (clip_g_path.size() > 0) {
LOG_INFO("loading clip_g from '%s'", clip_g_path.c_str());
if (!model_loader.init_from_file(clip_g_path, "text_encoders.clip_g.transformer.")) {
LOG_WARN("loading clip_g from '%s' failed", clip_g_path.c_str());
}
}
if (t5xxl_path.size() > 0) {
LOG_INFO("loading t5xxl from '%s'", t5xxl_path.c_str());
if (!model_loader.init_from_file(t5xxl_path, "text_encoders.t5xxl.transformer.")) {
LOG_WARN("loading t5xxl from '%s' failed", t5xxl_path.c_str());
}
}
if (diffusion_model_path.size() > 0) {
LOG_INFO("loading diffusion model from '%s'", diffusion_model_path.c_str());
if (!model_loader.init_from_file(diffusion_model_path, "model.diffusion_model.")) {
LOG_WARN("loading diffusion model from '%s' failed", diffusion_model_path.c_str());
}
}
if (vae_path.size() > 0) {
LOG_INFO("loading vae from '%s'", vae_path.c_str());
if (!model_loader.init_from_file(vae_path, "vae.")) {
LOG_WARN("loading vae from '%s' failed", vae_path.c_str());
}
}
version = model_loader.get_sd_version();
if (version == VERSION_COUNT) {
LOG_ERROR("get sd version from file failed: '%s'", model_path.c_str());
return false;
}
LOG_INFO("Version: %s ", model_version_to_str[version]);
if (wtype == GGML_TYPE_COUNT) {
model_wtype = model_loader.get_sd_wtype();
if (model_wtype == GGML_TYPE_COUNT) {
model_wtype = GGML_TYPE_F32;
LOG_WARN("can not get mode wtype frome weight, use f32");
}
conditioner_wtype = model_loader.get_conditioner_wtype();
if (conditioner_wtype == GGML_TYPE_COUNT) {
conditioner_wtype = wtype;
}
diffusion_model_wtype = model_loader.get_diffusion_model_wtype();
if (diffusion_model_wtype == GGML_TYPE_COUNT) {
diffusion_model_wtype = wtype;
}
vae_wtype = model_loader.get_vae_wtype();
if (vae_wtype == GGML_TYPE_COUNT) {
vae_wtype = wtype;
}
} else {
model_wtype = wtype;
conditioner_wtype = wtype;
diffusion_model_wtype = wtype;
vae_wtype = wtype;
model_loader.set_wtype_override(wtype);
}
if (version == VERSION_SDXL) {
vae_wtype = GGML_TYPE_F32;
model_loader.set_wtype_override(GGML_TYPE_F32, "vae.");
}
LOG_INFO("Weight type: %s", model_wtype != SD_TYPE_COUNT ? ggml_type_name(model_wtype) : "??");
LOG_INFO("Conditioner weight type: %s", conditioner_wtype != SD_TYPE_COUNT ? ggml_type_name(conditioner_wtype) : "??");
LOG_INFO("Diffusion model weight type: %s", diffusion_model_wtype != SD_TYPE_COUNT ? ggml_type_name(diffusion_model_wtype) : "??");
LOG_INFO("VAE weight type: %s", vae_wtype != SD_TYPE_COUNT ? ggml_type_name(vae_wtype) : "??");
LOG_DEBUG("ggml tensor size = %d bytes", (int)sizeof(ggml_tensor));
if (version == VERSION_SDXL) {
scale_factor = 0.13025f;
if (vae_path.size() == 0 && taesd_path.size() == 0) {
LOG_WARN(
"!!!It looks like you are using SDXL model. "
"If you find that the generated images are completely black, "
"try specifying SDXL VAE FP16 Fix with the --vae parameter. "
"You can find it here: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors");
}
} else if (sd_version_is_sd3(version)) {
scale_factor = 1.5305f;
} else if (sd_version_is_flux(version)) {
scale_factor = 0.3611;
// TODO: shift_factor
}
if (version == VERSION_SVD) {
clip_vision = std::make_shared<FrozenCLIPVisionEmbedder>(backend, model_loader.tensor_storages_types);
clip_vision->alloc_params_buffer();
clip_vision->get_param_tensors(tensors);
diffusion_model = std::make_shared<UNetModel>(backend, model_loader.tensor_storages_types, version);
diffusion_model->alloc_params_buffer();
diffusion_model->get_param_tensors(tensors);
first_stage_model = std::make_shared<AutoEncoderKL>(backend, model_loader.tensor_storages_types, "first_stage_model", vae_decode_only, true, version);
LOG_DEBUG("vae_decode_only %d", vae_decode_only);
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
} else {
clip_backend = backend;
bool use_t5xxl = false;
if (sd_version_is_dit(version)) {
use_t5xxl = true;
}
if (!ggml_backend_is_cpu(backend) && use_t5xxl && conditioner_wtype != GGML_TYPE_F32) {
clip_on_cpu = true;
LOG_INFO("set clip_on_cpu to true");
}
if (clip_on_cpu && !ggml_backend_is_cpu(backend)) {
LOG_INFO("CLIP: Using CPU backend");
clip_backend = ggml_backend_cpu_init();
}
if (diffusion_flash_attn) {
LOG_INFO("Using flash attention in the diffusion model");
}
if (sd_version_is_sd3(version)) {
if (diffusion_flash_attn) {
LOG_WARN("flash attention in this diffusion model is currently unsupported!");
}
cond_stage_model = std::make_shared<SD3CLIPEmbedder>(clip_backend, model_loader.tensor_storages_types);
diffusion_model = std::make_shared<MMDiTModel>(backend, model_loader.tensor_storages_types);
} else if (sd_version_is_flux(version)) {
cond_stage_model = std::make_shared<FluxCLIPEmbedder>(clip_backend, model_loader.tensor_storages_types);
diffusion_model = std::make_shared<FluxModel>(backend, model_loader.tensor_storages_types, diffusion_flash_attn);
} else {
if (id_embeddings_path.find("v2") != std::string::npos) {
cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend, model_loader.tensor_storages_types, embeddings_path, version, PM_VERSION_2);
} else {
cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend, model_loader.tensor_storages_types, embeddings_path, version);
}
diffusion_model = std::make_shared<UNetModel>(backend, model_loader.tensor_storages_types, version, diffusion_flash_attn);
}
cond_stage_model->alloc_params_buffer();
cond_stage_model->get_param_tensors(tensors);
diffusion_model->alloc_params_buffer();
diffusion_model->get_param_tensors(tensors);
if (!use_tiny_autoencoder) {
if (vae_on_cpu && !ggml_backend_is_cpu(backend)) {
LOG_INFO("VAE Autoencoder: Using CPU backend");
vae_backend = ggml_backend_cpu_init();
} else {
vae_backend = backend;
}
first_stage_model = std::make_shared<AutoEncoderKL>(vae_backend, model_loader.tensor_storages_types, "first_stage_model", vae_decode_only, false, version);
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
} else {
tae_first_stage = std::make_shared<TinyAutoEncoder>(backend, model_loader.tensor_storages_types, "decoder.layers", vae_decode_only);
}
// first_stage_model->get_param_tensors(tensors, "first_stage_model.");
if (control_net_path.size() > 0) {
ggml_backend_t controlnet_backend = NULL;
if (control_net_cpu && !ggml_backend_is_cpu(backend)) {
LOG_DEBUG("ControlNet: Using CPU backend");
controlnet_backend = ggml_backend_cpu_init();
} else {
controlnet_backend = backend;
}
control_net = std::make_shared<ControlNet>(controlnet_backend, model_loader.tensor_storages_types, version);
}
if (id_embeddings_path.find("v2") != std::string::npos) {
pmid_model = std::make_shared<PhotoMakerIDEncoder>(backend, model_loader.tensor_storages_types, "pmid", version, PM_VERSION_2);
LOG_INFO("using PhotoMaker Version 2");
} else {
pmid_model = std::make_shared<PhotoMakerIDEncoder>(backend, model_loader.tensor_storages_types, "pmid", version);
}
if (id_embeddings_path.size() > 0) {
pmid_lora = std::make_shared<LoraModel>(backend, id_embeddings_path, "");
if (!pmid_lora->load_from_file(true)) {
LOG_WARN("load photomaker lora tensors from %s failed", id_embeddings_path.c_str());
return false;
}
LOG_INFO("loading stacked ID embedding (PHOTOMAKER) model file from '%s'", id_embeddings_path.c_str());
if (!model_loader.init_from_file(id_embeddings_path, "pmid.")) {
LOG_WARN("loading stacked ID embedding from '%s' failed", id_embeddings_path.c_str());
} else {
stacked_id = true;
}
}
if (stacked_id) {
if (!pmid_model->alloc_params_buffer()) {
LOG_ERROR(" pmid model params buffer allocation failed");
return false;
}
pmid_model->get_param_tensors(tensors, "pmid");
}
}
struct ggml_init_params params;
params.mem_size = static_cast<size_t>(10 * 1024) * 1024; // 10M
params.mem_buffer = NULL;
params.no_alloc = false;
// LOG_DEBUG("mem_size %u ", params.mem_size);
struct ggml_context* ctx = ggml_init(params); // for alphas_cumprod and is_using_v_parameterization check
GGML_ASSERT(ctx != NULL);
ggml_tensor* alphas_cumprod_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, TIMESTEPS);
calculate_alphas_cumprod((float*)alphas_cumprod_tensor->data);
// load weights
LOG_DEBUG("loading weights");
int64_t t0 = ggml_time_ms();
std::set<std::string> ignore_tensors;
tensors["alphas_cumprod"] = alphas_cumprod_tensor;
if (use_tiny_autoencoder) {
ignore_tensors.insert("first_stage_model.");
}
if (stacked_id) {
ignore_tensors.insert("lora.");
}
if (vae_decode_only) {
ignore_tensors.insert("first_stage_model.encoder");
ignore_tensors.insert("first_stage_model.quant");
}
if (version == VERSION_SVD) {
ignore_tensors.insert("conditioner.embedders.3");
}
bool success = model_loader.load_tensors(tensors, backend, ignore_tensors);
if (!success) {
LOG_ERROR("load tensors from model loader failed");
ggml_free(ctx);
return false;
}
// LOG_DEBUG("model size = %.2fMB", total_size / 1024.0 / 1024.0);
if (version == VERSION_SVD) {
// diffusion_model->test();
// first_stage_model->test();
// return false;
} else {
size_t clip_params_mem_size = cond_stage_model->get_params_buffer_size();
size_t unet_params_mem_size = diffusion_model->get_params_buffer_size();
size_t vae_params_mem_size = 0;
if (!use_tiny_autoencoder) {
vae_params_mem_size = first_stage_model->get_params_buffer_size();
} else {
if (!tae_first_stage->load_from_file(taesd_path)) {
return false;
}
vae_params_mem_size = tae_first_stage->get_params_buffer_size();
}
size_t control_net_params_mem_size = 0;
if (control_net) {
if (!control_net->load_from_file(control_net_path)) {
return false;
}
control_net_params_mem_size = control_net->get_params_buffer_size();
}
size_t pmid_params_mem_size = 0;
if (stacked_id) {
pmid_params_mem_size = pmid_model->get_params_buffer_size();
}
size_t total_params_ram_size = 0;
size_t total_params_vram_size = 0;
if (ggml_backend_is_cpu(clip_backend)) {
total_params_ram_size += clip_params_mem_size + pmid_params_mem_size;
} else {
total_params_vram_size += clip_params_mem_size + pmid_params_mem_size;
}
if (ggml_backend_is_cpu(backend)) {
total_params_ram_size += unet_params_mem_size;
} else {
total_params_vram_size += unet_params_mem_size;
}
if (ggml_backend_is_cpu(vae_backend)) {
total_params_ram_size += vae_params_mem_size;
} else {
total_params_vram_size += vae_params_mem_size;
}
if (ggml_backend_is_cpu(control_net_backend)) {
total_params_ram_size += control_net_params_mem_size;
} else {
total_params_vram_size += control_net_params_mem_size;
}
size_t total_params_size = total_params_ram_size + total_params_vram_size;
LOG_INFO(
"total params memory size = %.2fMB (VRAM %.2fMB, RAM %.2fMB): "
"clip %.2fMB(%s), unet %.2fMB(%s), vae %.2fMB(%s), controlnet %.2fMB(%s), pmid %.2fMB(%s)",
total_params_size / 1024.0 / 1024.0,
total_params_vram_size / 1024.0 / 1024.0,
total_params_ram_size / 1024.0 / 1024.0,
clip_params_mem_size / 1024.0 / 1024.0,
ggml_backend_is_cpu(clip_backend) ? "RAM" : "VRAM",
unet_params_mem_size / 1024.0 / 1024.0,
ggml_backend_is_cpu(backend) ? "RAM" : "VRAM",
vae_params_mem_size / 1024.0 / 1024.0,
ggml_backend_is_cpu(vae_backend) ? "RAM" : "VRAM",
control_net_params_mem_size / 1024.0 / 1024.0,
ggml_backend_is_cpu(control_net_backend) ? "RAM" : "VRAM",
pmid_params_mem_size / 1024.0 / 1024.0,
ggml_backend_is_cpu(clip_backend) ? "RAM" : "VRAM");
}
int64_t t1 = ggml_time_ms();
LOG_INFO("loading model from '%s' completed, taking %.2fs", model_path.c_str(), (t1 - t0) * 1.0f / 1000);
// check is_using_v_parameterization_for_sd2
bool is_using_v_parameterization = false;
if (version == VERSION_SD2) {
if (is_using_v_parameterization_for_sd2(ctx)) {
is_using_v_parameterization = true;
}
} else if (version == VERSION_SVD) {
// TODO: V_PREDICTION_EDM
is_using_v_parameterization = true;
}
if (sd_version_is_sd3(version)) {
LOG_INFO("running in FLOW mode");
denoiser = std::make_shared<DiscreteFlowDenoiser>();
} else if (sd_version_is_flux(version)) {
LOG_INFO("running in Flux FLOW mode");
float shift = 1.0f; // TODO: validate
for (auto pair : model_loader.tensor_storages_types) {
if (pair.first.find("model.diffusion_model.guidance_in.in_layer.weight") != std::string::npos) {
shift = 1.15f;
break;
}
}
denoiser = std::make_shared<FluxFlowDenoiser>(shift);
} else if (is_using_v_parameterization) {
LOG_INFO("running in v-prediction mode");
denoiser = std::make_shared<CompVisVDenoiser>();
} else {
LOG_INFO("running in eps-prediction mode");
}
if (schedule != DEFAULT) {
switch (schedule) {
case DISCRETE:
LOG_INFO("running with discrete schedule");
denoiser->schedule = std::make_shared<DiscreteSchedule>();
break;
case KARRAS:
LOG_INFO("running with Karras schedule");
denoiser->schedule = std::make_shared<KarrasSchedule>();
break;
case EXPONENTIAL:
LOG_INFO("running exponential schedule");
denoiser->schedule = std::make_shared<ExponentialSchedule>();
break;
case AYS:
LOG_INFO("Running with Align-Your-Steps schedule");
denoiser->schedule = std::make_shared<AYSSchedule>();
denoiser->schedule->version = version;
break;
case GITS:
LOG_INFO("Running with GITS schedule");
denoiser->schedule = std::make_shared<GITSSchedule>();
denoiser->schedule->version = version;
break;
case DEFAULT:
// Don't touch anything.
break;
default:
LOG_ERROR("Unknown schedule %i", schedule);
abort();
}
}
auto comp_vis_denoiser = std::dynamic_pointer_cast<CompVisDenoiser>(denoiser);
if (comp_vis_denoiser) {
for (int i = 0; i < TIMESTEPS; i++) {
comp_vis_denoiser->sigmas[i] = std::sqrt((1 - ((float*)alphas_cumprod_tensor->data)[i]) / ((float*)alphas_cumprod_tensor->data)[i]);
comp_vis_denoiser->log_sigmas[i] = std::log(comp_vis_denoiser->sigmas[i]);
}
}
LOG_DEBUG("finished loaded file");
ggml_free(ctx);
return true;
}
bool is_using_v_parameterization_for_sd2(ggml_context* work_ctx) {
struct ggml_tensor* x_t = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 4, 1);
ggml_set_f32(x_t, 0.5);
struct ggml_tensor* c = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 1024, 2, 1, 1);
ggml_set_f32(c, 0.5);
struct ggml_tensor* timesteps = ggml_new_tensor_1d(work_ctx, GGML_TYPE_F32, 1);
ggml_set_f32(timesteps, 999);
int64_t t0 = ggml_time_ms();
struct ggml_tensor* out = ggml_dup_tensor(work_ctx, x_t);
diffusion_model->compute(n_threads, x_t, timesteps, c, NULL, NULL, NULL, -1, {}, 0.f, &out);
diffusion_model->free_compute_buffer();
double result = 0.f;
{
float* vec_x = (float*)x_t->data;
float* vec_out = (float*)out->data;
int64_t n = ggml_nelements(out);
for (int i = 0; i < n; i++) {
result += ((double)vec_out[i] - (double)vec_x[i]);
}
result /= n;
}
int64_t t1 = ggml_time_ms();
LOG_DEBUG("check is_using_v_parameterization_for_sd2, taking %.2fs", (t1 - t0) * 1.0f / 1000);
return result < -1;
}
void apply_lora(const std::string& lora_name, float multiplier) {
int64_t t0 = ggml_time_ms();
std::string st_file_path = path_join(lora_model_dir, lora_name + ".safetensors");
std::string ckpt_file_path = path_join(lora_model_dir, lora_name + ".ckpt");
std::string file_path;
if (file_exists(st_file_path)) {
file_path = st_file_path;
} else if (file_exists(ckpt_file_path)) {
file_path = ckpt_file_path;
} else {
LOG_WARN("can not find %s or %s for lora %s", st_file_path.c_str(), ckpt_file_path.c_str(), lora_name.c_str());
return;
}
LoraModel lora(backend, file_path);
if (!lora.load_from_file()) {
LOG_WARN("load lora tensors from %s failed", file_path.c_str());
return;
}
lora.multiplier = multiplier;
lora.apply(tensors, n_threads);
lora.free_params_buffer();
int64_t t1 = ggml_time_ms();
LOG_INFO("lora '%s' applied, taking %.2fs", lora_name.c_str(), (t1 - t0) * 1.0f / 1000);
}
void apply_loras(const std::unordered_map<std::string, float>& lora_state) {
if (lora_state.size() > 0 && model_wtype != GGML_TYPE_F16 && model_wtype != GGML_TYPE_F32) {
LOG_WARN("In quantized models when applying LoRA, the images have poor quality.");
}
std::unordered_map<std::string, float> lora_state_diff;
for (auto& kv : lora_state) {
const std::string& lora_name = kv.first;
float multiplier = kv.second;
if (curr_lora_state.find(lora_name) != curr_lora_state.end()) {
float curr_multiplier = curr_lora_state[lora_name];
float multiplier_diff = multiplier - curr_multiplier;
if (multiplier_diff != 0.f) {
lora_state_diff[lora_name] = multiplier_diff;
}
} else {
lora_state_diff[lora_name] = multiplier;
}
}
LOG_INFO("Attempting to apply %lu LoRAs", lora_state.size());
for (auto& kv : lora_state_diff) {
apply_lora(kv.first, kv.second);
}
curr_lora_state = lora_state;
}
ggml_tensor* id_encoder(ggml_context* work_ctx,
ggml_tensor* init_img,
ggml_tensor* prompts_embeds,
ggml_tensor* id_embeds,
std::vector<bool>& class_tokens_mask) {
ggml_tensor* res = NULL;
pmid_model->compute(n_threads, init_img, prompts_embeds, id_embeds, class_tokens_mask, &res, work_ctx);
return res;
}
SDCondition get_svd_condition(ggml_context* work_ctx,
sd_image_t init_image,
int width,
int height,
int fps = 6,
int motion_bucket_id = 127,
float augmentation_level = 0.f,
bool force_zero_embeddings = false) {
// c_crossattn
int64_t t0 = ggml_time_ms();
struct ggml_tensor* c_crossattn = NULL;
{
if (force_zero_embeddings) {
c_crossattn = ggml_new_tensor_1d(work_ctx, GGML_TYPE_F32, clip_vision->vision_model.projection_dim);
ggml_set_f32(c_crossattn, 0.f);
} else {
sd_image_f32_t image = sd_image_t_to_sd_image_f32_t(init_image);
sd_image_f32_t resized_image = clip_preprocess(image, clip_vision->vision_model.image_size);
free(image.data);
image.data = NULL;
ggml_tensor* pixel_values = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, resized_image.width, resized_image.height, 3, 1);
sd_image_f32_to_tensor(resized_image.data, pixel_values, false);
free(resized_image.data);
resized_image.data = NULL;
// print_ggml_tensor(pixel_values);
clip_vision->compute(n_threads, pixel_values, &c_crossattn, work_ctx);
// print_ggml_tensor(c_crossattn);
}
}
// c_concat
struct ggml_tensor* c_concat = NULL;
{
if (force_zero_embeddings) {
c_concat = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width / 8, height / 8, 4, 1);
ggml_set_f32(c_concat, 0.f);
} else {
ggml_tensor* init_img = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width, height, 3, 1);
if (width != init_image.width || height != init_image.height) {
sd_image_f32_t image = sd_image_t_to_sd_image_f32_t(init_image);
sd_image_f32_t resized_image = resize_sd_image_f32_t(image, width, height);
free(image.data);
image.data = NULL;
sd_image_f32_to_tensor(resized_image.data, init_img, false);
free(resized_image.data);
resized_image.data = NULL;
} else {
sd_image_to_tensor(init_image.data, init_img);
}
if (augmentation_level > 0.f) {
struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, init_img);
ggml_tensor_set_f32_randn(noise, rng);
// encode_pixels += torch.randn_like(pixels) * augmentation_level
ggml_tensor_scale(noise, augmentation_level);
ggml_tensor_add(init_img, noise);
}
ggml_tensor* moments = encode_first_stage(work_ctx, init_img);
c_concat = get_first_stage_encoding(work_ctx, moments);
}
}
// y
struct ggml_tensor* y = NULL;
{
y = ggml_new_tensor_1d(work_ctx, GGML_TYPE_F32, diffusion_model->get_adm_in_channels());
int out_dim = 256;
int fps_id = fps - 1;
std::vector<float> timesteps = {(float)fps_id, (float)motion_bucket_id, augmentation_level};
set_timestep_embedding(timesteps, y, out_dim);
}
int64_t t1 = ggml_time_ms();
LOG_DEBUG("computing svd condition graph completed, taking %" PRId64 " ms", t1 - t0);
return {c_crossattn, y, c_concat};
}
ggml_tensor* sample(ggml_context* work_ctx,
ggml_tensor* init_latent,
ggml_tensor* noise,
SDCondition cond,
SDCondition uncond,
ggml_tensor* control_hint,
float control_strength,
float min_cfg,
float cfg_scale,
float guidance,
sample_method_t method,
const std::vector<float>& sigmas,
int start_merge_step,
SDCondition id_cond,
std::vector<int> skip_layers = {},
float slg_scale = 0,
float skip_layer_start = 0.01,
float skip_layer_end = 0.2,
std::function<void(int, ggml_tensor*, SDVersion)> step_callback = nullptr) {
size_t steps = sigmas.size() - 1;
// noise = load_tensor_from_file(work_ctx, "./rand0.bin");
// print_ggml_tensor(noise);
struct ggml_tensor* x = ggml_dup_tensor(work_ctx, init_latent);
copy_ggml_tensor(x, init_latent);
x = denoiser->noise_scaling(sigmas[0], noise, x);
struct ggml_tensor* noised_input = ggml_dup_tensor(work_ctx, noise);
bool has_unconditioned = cfg_scale != 1.0 && uncond.c_crossattn != NULL;
bool has_skiplayer = slg_scale != 0.0 && skip_layers.size() > 0;
// denoise wrapper
struct ggml_tensor* out_cond = ggml_dup_tensor(work_ctx, x);
struct ggml_tensor* out_uncond = NULL;
struct ggml_tensor* out_skip = NULL;
if (has_unconditioned) {
out_uncond = ggml_dup_tensor(work_ctx, x);
}
if (has_skiplayer) {
if (sd_version_is_dit(version)) {
out_skip = ggml_dup_tensor(work_ctx, x);
} else {
has_skiplayer = false;
LOG_WARN("SLG is incompatible with %s models", model_version_to_str[version]);
}
}
struct ggml_tensor* denoised = ggml_dup_tensor(work_ctx, x);
auto denoise = [&](ggml_tensor* input, float sigma, int step) -> ggml_tensor* {
if (step == 1) {
pretty_progress(0, (int)steps, 0);
}
int64_t t0 = ggml_time_us();
std::vector<float> scaling = denoiser->get_scalings(sigma);
GGML_ASSERT(scaling.size() == 3);
float c_skip = scaling[0];
float c_out = scaling[1];
float c_in = scaling[2];
float t = denoiser->sigma_to_t(sigma);
std::vector<float> timesteps_vec(x->ne[3], t); // [N, ]
auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
std::vector<float> guidance_vec(x->ne[3], guidance);
auto guidance_tensor = vector_to_ggml_tensor(work_ctx, guidance_vec);
copy_ggml_tensor(noised_input, input);
// noised_input = noised_input * c_in
ggml_tensor_scale(noised_input, c_in);
std::vector<struct ggml_tensor*> controls;
if (control_hint != NULL) {
control_net->compute(n_threads, noised_input, control_hint, timesteps, cond.c_crossattn, cond.c_vector);
controls = control_net->controls;
// print_ggml_tensor(controls[12]);
// GGML_ASSERT(0);
}
if (start_merge_step == -1 || step <= start_merge_step) {
// cond
diffusion_model->compute(n_threads,
noised_input,
timesteps,
cond.c_crossattn,
cond.c_concat,
cond.c_vector,
guidance_tensor,
-1,
controls,
control_strength,
&out_cond);
} else {
diffusion_model->compute(n_threads,
noised_input,
timesteps,
id_cond.c_crossattn,
cond.c_concat,
id_cond.c_vector,
guidance_tensor,
-1,
controls,
control_strength,
&out_cond);
}
float* negative_data = NULL;
if (has_unconditioned) {
// uncond
if (control_hint != NULL) {
control_net->compute(n_threads, noised_input, control_hint, timesteps, uncond.c_crossattn, uncond.c_vector);
controls = control_net->controls;
}
diffusion_model->compute(n_threads,
noised_input,
timesteps,
uncond.c_crossattn,
uncond.c_concat,
uncond.c_vector,
guidance_tensor,
-1,
controls,
control_strength,
&out_uncond);
negative_data = (float*)out_uncond->data;
}
int step_count = sigmas.size();
bool is_skiplayer_step = has_skiplayer && step > (int)(skip_layer_start * step_count) && step < (int)(skip_layer_end * step_count);
float* skip_layer_data = NULL;
if (is_skiplayer_step) {
LOG_DEBUG("Skipping layers at step %d\n", step);
// skip layer (same as conditionned)
diffusion_model->compute(n_threads,
noised_input,
timesteps,
cond.c_crossattn,
cond.c_concat,
cond.c_vector,
guidance_tensor,
-1,
controls,
control_strength,
&out_skip,
NULL,
skip_layers);
skip_layer_data = (float*)out_skip->data;
}
float* vec_denoised = (float*)denoised->data;
float* vec_input = (float*)input->data;
float* positive_data = (float*)out_cond->data;
int ne_elements = (int)ggml_nelements(denoised);
for (int i = 0; i < ne_elements; i++) {
float latent_result = positive_data[i];
if (has_unconditioned) {
// out_uncond + cfg_scale * (out_cond - out_uncond)
int64_t ne3 = out_cond->ne[3];
if (min_cfg != cfg_scale && ne3 != 1) {
int64_t i3 = i / out_cond->ne[0] * out_cond->ne[1] * out_cond->ne[2];
float scale = min_cfg + (cfg_scale - min_cfg) * (i3 * 1.0f / ne3);
} else {
latent_result = negative_data[i] + cfg_scale * (positive_data[i] - negative_data[i]);
}
}
if (is_skiplayer_step) {
latent_result = latent_result + (positive_data[i] - skip_layer_data[i]) * slg_scale;
}
// v = latent_result, eps = latent_result
// denoised = (v * c_out + input * c_skip) or (input + eps * c_out)
vec_denoised[i] = latent_result * c_out + vec_input[i] * c_skip;
}
int64_t t1 = ggml_time_us();
if (step > 0) {
pretty_progress(step, (int)steps, (t1 - t0) / 1000000.f);
// LOG_INFO("step %d sampling completed taking %.2fs", step, (t1 - t0) * 1.0f / 1000000);
}
if (step_callback != nullptr) {
step_callback(step, denoised, version);
}
return denoised;
};
sample_k_diffusion(method, denoise, work_ctx, x, sigmas, rng);
x = denoiser->inverse_noise_scaling(sigmas[sigmas.size() - 1], x);
if (control_net) {
control_net->free_control_ctx();
control_net->free_compute_buffer();
}
diffusion_model->free_compute_buffer();
return x;
}
// ldm.models.diffusion.ddpm.LatentDiffusion.get_first_stage_encoding
ggml_tensor* get_first_stage_encoding(ggml_context* work_ctx, ggml_tensor* moments) {
// ldm.modules.distributions.distributions.DiagonalGaussianDistribution.sample
ggml_tensor* latent = ggml_new_tensor_4d(work_ctx, moments->type, moments->ne[0], moments->ne[1], moments->ne[2] / 2, moments->ne[3]);
struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, latent);
ggml_tensor_set_f32_randn(noise, rng);
// noise = load_tensor_from_file(work_ctx, "noise.bin");
{
float mean = 0;
float logvar = 0;
float value = 0;
float std_ = 0;
for (int i = 0; i < latent->ne[3]; i++) {
for (int j = 0; j < latent->ne[2]; j++) {
for (int k = 0; k < latent->ne[1]; k++) {
for (int l = 0; l < latent->ne[0]; l++) {
mean = ggml_tensor_get_f32(moments, l, k, j, i);
logvar = ggml_tensor_get_f32(moments, l, k, j + (int)latent->ne[2], i);
logvar = std::max(-30.0f, std::min(logvar, 20.0f));
std_ = std::exp(0.5f * logvar);
value = mean + std_ * ggml_tensor_get_f32(noise, l, k, j, i);
value = value * scale_factor;
// printf("%d %d %d %d -> %f\n", i, j, k, l, value);
ggml_tensor_set_f32(latent, value, l, k, j, i);
}
}
}
}
}
return latent;
}
ggml_tensor* compute_first_stage(ggml_context* work_ctx, ggml_tensor* x, bool decode) {
int64_t W = x->ne[0];
int64_t H = x->ne[1];
int64_t C = 8;