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
# Original code from Comfy, https://github.com/comfyanonymous/ComfyUI
# Modified for improved VAE tiling and aspect ratio handling

import torch

from ldm_patched.modules import model_management
from ldm_patched.ldm.models.autoencoder import AutoencoderKL, AutoencodingEngine
import yaml

import ldm_patched.modules.utils

from . import clip_vision
from . import gligen
from . import diffusers_convert
from . import model_base
from . import model_detection

from . import sd1_clip
from . import sd2_clip
from . import sdxl_clip

import ldm_patched.modules.model_patcher
import ldm_patched.modules.lora
import ldm_patched.t2ia.adapter
import ldm_patched.modules.supported_models_base
import ldm_patched.taesd.taesd


def load_model_weights(model, sd):
    m, u = model.load_state_dict(sd, strict=False)
    m = set(m)
    unexpected_keys = set(u)

    k = list(sd.keys())
    for x in k:
        if x not in unexpected_keys:
            w = sd.pop(x)
            del w
    if len(m) > 0:
        print("extra", m)
    return model


def load_clip_weights(model, sd):
    k = list(sd.keys())
    for x in k:
        if x.startswith("cond_stage_model.transformer.") and not x.startswith("cond_stage_model.transformer.text_model."):
            y = x.replace("cond_stage_model.transformer.", "cond_stage_model.transformer.text_model.")
            sd[y] = sd.pop(x)

    if 'cond_stage_model.transformer.text_model.embeddings.position_ids' in sd:
        ids = sd['cond_stage_model.transformer.text_model.embeddings.position_ids']
        if ids.dtype == torch.float32:
            sd['cond_stage_model.transformer.text_model.embeddings.position_ids'] = ids.round()

    sd = ldm_patched.modules.utils.transformers_convert(sd, "cond_stage_model.model.", "cond_stage_model.transformer.text_model.", 24)
    return load_model_weights(model, sd)


def load_lora_for_models(model, clip, lora, strength_model, strength_clip, filename='default'):
    model_flag = type(model.model).__name__ if model is not None else 'default'

    key_map = {}
    if model is not None and strength_model != 0:
        key_map = ldm_patched.modules.lora.model_lora_keys_unet(model.model, key_map)
    if clip is not None and strength_clip != 0:
        key_map = ldm_patched.modules.lora.model_lora_keys_clip(clip.cond_stage_model, key_map)

    if not key_map:
        return (model, clip)

    loaded = ldm_patched.modules.lora.load_lora(lora, key_map)

    if model is not None and strength_model != 0:
        new_modelpatcher = model.clone()
        loaded_keys_unet = new_modelpatcher.add_patches(loaded, strength_model)
    else:
        new_modelpatcher = model
        loaded_keys_unet = set()

    if clip is not None and strength_clip != 0:
        new_clip = clip.clone()
        loaded_keys_clip = new_clip.add_patches(loaded, strength_clip)
    else:
        new_clip = clip
        loaded_keys_clip = set()

    if loaded_keys_unet or loaded_keys_clip:
        total_loaded_keys = len(loaded_keys_unet) + len(loaded_keys_clip)
        print(f'[LORA] Loaded {filename} for {model_flag} with {total_loaded_keys} keys at weight {strength_clip}')

    return (new_modelpatcher, new_clip)


class CLIP:
    def __init__(self, target=None, embedding_directory=None, no_init=False):
        if no_init:
            return
        params = target.params.copy()
        clip = target.clip
        tokenizer = target.tokenizer

        load_device = model_management.text_encoder_device()
        offload_device = model_management.text_encoder_offload_device()
        params['device'] = offload_device
        params['dtype'] = model_management.text_encoder_dtype(load_device)

        self.cond_stage_model = clip(**(params))
        self.tokenizer = tokenizer(embedding_directory=embedding_directory)
        self.patcher = ldm_patched.modules.model_patcher.ModelPatcher(
            self.cond_stage_model,
            load_device=load_device,
            offload_device=offload_device
        )
        self.layer_idx = None

    def clone(self):
        n = CLIP(no_init=True)
        n.patcher = self.patcher.clone()
        n.cond_stage_model = self.cond_stage_model
        n.tokenizer = self.tokenizer
        n.layer_idx = self.layer_idx
        return n

    def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
        return self.patcher.add_patches(patches, strength_patch, strength_model)

    def clip_layer(self, layer_idx):
        self.layer_idx = layer_idx

    def tokenize(self, text, return_word_ids=False):
        return self.tokenizer.tokenize_with_weights(text, return_word_ids)

    def encode_from_tokens(self, tokens, return_pooled=False):
        self.cond_stage_model.reset_clip_options()
        if self.layer_idx is not None:
            self.cond_stage_model.set_clip_options({"layer": self.layer_idx})

        if return_pooled == "unprojected":
            self.cond_stage_model.set_clip_options({"projected_pooled": False})

        self.load_model()
        cond, pooled = self.cond_stage_model.encode_token_weights(tokens)
        if return_pooled:
            return cond, pooled
        return cond

    def encode(self, text):
        tokens = self.tokenize(text)
        return self.encode_from_tokens(tokens)

    def load_sd(self, sd):
        return self.cond_stage_model.load_sd(sd)

    def get_sd(self):
        return self.cond_stage_model.state_dict()

    def load_model(self):
        model_management.load_model_gpu(self.patcher)
        return self.patcher

    def get_key_patches(self):
        return self.patcher.get_key_patches()


class VAE:
    # Conservative defaults for 12GB
    STRIPE_ASPECT_THRESHOLD = 2.0
    MIN_TILE_SIZE = 32
    DEFAULT_TILE_SIZE_LATENT = 64  # 512px output tiles
    DEFAULT_OVERLAP_LATENT = 8     # 64px overlap

    def __init__(self, sd=None, device=None, config=None, dtype=None, no_init=False):
        if no_init:
            return

        if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys():
            sd = diffusers_convert.convert_vae_state_dict(sd)

        self.memory_used_encode = lambda shape, dtype: (1767 * shape[2] * shape[3]) * model_management.dtype_size(dtype)
        self.memory_used_decode = lambda shape, dtype: (2178 * shape[2] * shape[3] * 64) * model_management.dtype_size(dtype)
        self.downscale_ratio = 8
        self.latent_channels = 4

        if config is None:
            if "decoder.mid.block_1.mix_factor" in sd:
                encoder_config = {
                    'double_z': True, 'z_channels': 4, 'resolution': 256,
                    'in_channels': 3, 'out_ch': 3, 'ch': 128,
                    'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2,
                    'attn_resolutions': [], 'dropout': 0.0
                }
                decoder_config = encoder_config.copy()
                decoder_config["video_kernel_size"] = [3, 1, 1]
                decoder_config["alpha"] = 0.0
                self.first_stage_model = AutoencodingEngine(
                    regularizer_config={'target': "ldm_patched.ldm.models.autoencoder.DiagonalGaussianRegularizer"},
                    encoder_config={'target': "ldm_patched.ldm.modules.diffusionmodules.model.Encoder", 'params': encoder_config},
                    decoder_config={'target': "ldm_patched.ldm.modules.temporal_ae.VideoDecoder", 'params': decoder_config}
                )
            elif "taesd_decoder.1.weight" in sd:
                self.first_stage_model = ldm_patched.taesd.taesd.TAESD()
            else:
                ddconfig = {
                    'double_z': True, 'z_channels': 4, 'resolution': 256,
                    'in_channels': 3, 'out_ch': 3, 'ch': 128,
                    'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2,
                    'attn_resolutions': [], 'dropout': 0.0
                }

                if 'encoder.down.2.downsample.conv.weight' not in sd:
                    ddconfig['ch_mult'] = [1, 2, 4]
                    self.downscale_ratio = 4

                self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=4)
        else:
            self.first_stage_model = AutoencoderKL(**(config['params']))

        self.first_stage_model = self.first_stage_model.eval()

        m, u = self.first_stage_model.load_state_dict(sd, strict=False)
        if len(m) > 0:
            print("Missing VAE keys", m)
        if len(u) > 0:
            print("Leftover VAE keys", u)

        if device is None:
            device = model_management.vae_device()
        self.device = device
        offload_device = model_management.vae_offload_device()

        if dtype is None:
            dtype = model_management.vae_dtype()
        self.vae_dtype = dtype
        self.first_stage_model.to(self.vae_dtype)
        self.output_device = model_management.intermediate_device()

        self.patcher = ldm_patched.modules.model_patcher.ModelPatcher(
            self.first_stage_model,
            load_device=self.device,
            offload_device=offload_device
        )

    def clone(self):
        n = VAE(no_init=True)
        n.patcher = self.patcher.clone()
        n.memory_used_encode = self.memory_used_encode
        n.memory_used_decode = self.memory_used_decode
        n.downscale_ratio = self.downscale_ratio
        n.latent_channels = self.latent_channels
        n.first_stage_model = self.first_stage_model
        n.device = self.device
        n.vae_dtype = self.vae_dtype
        n.output_device = self.output_device
        return n

    def _decode_tile_to_cpu(self, samples):
        """Decode a single tile and return result on CPU"""
        samples_gpu = samples.to(self.vae_dtype).to(self.device)
        decoded = self.first_stage_model.decode(samples_gpu)
        # Normalize, clamp, move to CPU immediately
        result = torch.clamp((decoded.float() + 1.0) / 2.0, 0.0, 1.0).cpu()
        # Let GPU memory be freed naturally
        return result

    def _encode_tile_to_cpu(self, pixels):
        """Encode a single tile and return result on CPU"""
        pixels_gpu = (2.0 * pixels - 1.0).to(self.vae_dtype).to(self.device)
        encoded = self.first_stage_model.encode(pixels_gpu)
        result = encoded.float().cpu()
        return result

    def _make_linear_ramp(self, size, direction='up'):
        """Create a 1D linear ramp on CPU"""
        if direction == 'up':
            return torch.linspace(0.0, 1.0, size, dtype=torch.float32)
        else:
            return torch.linspace(1.0, 0.0, size, dtype=torch.float32)

    def _decode_tiled_cpu_accumulate(self, samples, tile_x, tile_y, overlap):
        """
        Tiled decode with CPU accumulation.
        GPU only processes one tile at a time.
        """
        b, c, h, w = samples.shape
        out_h = h * self.downscale_ratio
        out_w = w * self.downscale_ratio
        overlap_px = overlap * self.downscale_ratio

        # All accumulation happens on CPU - uses system RAM
        output = torch.zeros((b, 3, out_h, out_w), dtype=torch.float32)
        weights = torch.zeros((1, 1, out_h, out_w), dtype=torch.float32)

        step_x = max(1, tile_x - overlap)
        step_y = max(1, tile_y - overlap)

        # Build tile list
        tiles = []
        y = 0
        while y < h:
            x = 0
            while x < w:
                x_end = min(x + tile_x, w)
                y_end = min(y + tile_y, h)
                tiles.append((x, y, x_end, y_end))
                x += step_x
            y += step_y

        pbar = ldm_patched.modules.utils.ProgressBar(len(tiles), title='VAE decode')

        for (x, y, x_end, y_end) in tiles:
            # Extract tile (stays on CPU initially if samples on CPU, or moved)
            tile = samples[:, :, y:y_end, x:x_end]

            # Decode - GPU work happens here, result comes back to CPU
            decoded = self._decode_tile_to_cpu(tile)

            # Output coordinates
            ox = x * self.downscale_ratio
            oy = y * self.downscale_ratio
            ox_end = x_end * self.downscale_ratio
            oy_end = y_end * self.downscale_ratio

            th = oy_end - oy
            tw = ox_end - ox

            # Build blend mask on CPU
            blend = torch.ones((1, 1, th, tw), dtype=torch.float32)

            # Left edge blend
            if x > 0 and overlap_px > 0:
                ramp_len = min(overlap_px, tw)
                ramp = self._make_linear_ramp(ramp_len, 'up')
                blend[:, :, :, :ramp_len] *= ramp.view(1, 1, 1, -1)

            # Right edge blend
            if x_end < w and overlap_px > 0:
                ramp_len = min(overlap_px, tw)
                ramp = self._make_linear_ramp(ramp_len, 'down')
                blend[:, :, :, -ramp_len:] *= ramp.view(1, 1, 1, -1)

            # Top edge blend
            if y > 0 and overlap_px > 0:
                ramp_len = min(overlap_px, th)
                ramp = self._make_linear_ramp(ramp_len, 'up')
                blend[:, :, :ramp_len, :] *= ramp.view(1, 1, -1, 1)

            # Bottom edge blend
            if y_end < h and overlap_px > 0:
                ramp_len = min(overlap_px, th)
                ramp = self._make_linear_ramp(ramp_len, 'down')
                blend[:, :, -ramp_len:, :] *= ramp.view(1, 1, -1, 1)

            # Accumulate on CPU
            output[:, :, oy:oy_end, ox:ox_end] += decoded * blend
            weights[:, :, oy:oy_end, ox:ox_end] += blend

            pbar.update(1)

        # Normalize
        result = output / weights.clamp(min=1e-8)

        return result

    def _encode_tiled_cpu_accumulate(self, pixel_samples, tile_x, tile_y, overlap):
        """
        Tiled encode with CPU accumulation.
        """
        b, c, h, w = pixel_samples.shape
        out_h = h // self.downscale_ratio
        out_w = w // self.downscale_ratio
        overlap_latent = overlap // self.downscale_ratio

        output = torch.zeros((b, self.latent_channels, out_h, out_w), dtype=torch.float32)
        weights = torch.zeros((1, 1, out_h, out_w), dtype=torch.float32)

        step_x = max(self.downscale_ratio, tile_x - overlap)
        step_y = max(self.downscale_ratio, tile_y - overlap)

        tiles = []
        y = 0
        while y < h:
            x = 0
            while x < w:
                x_end = min(x + tile_x, w)
                y_end = min(y + tile_y, h)
                tiles.append((x, y, x_end, y_end))
                x += step_x
            y += step_y

        pbar = ldm_patched.modules.utils.ProgressBar(len(tiles), title='VAE encode')

        for (x, y, x_end, y_end) in tiles:
            tile = pixel_samples[:, :, y:y_end, x:x_end]
            encoded = self._encode_tile_to_cpu(tile)

            ox = x // self.downscale_ratio
            oy = y // self.downscale_ratio
            ox_end = x_end // self.downscale_ratio
            oy_end = y_end // self.downscale_ratio

            th = oy_end - oy
            tw = ox_end - ox

            blend = torch.ones((1, 1, th, tw), dtype=torch.float32)

            if x > 0 and overlap_latent > 0:
                ramp_len = min(overlap_latent, tw)
                ramp = self._make_linear_ramp(ramp_len, 'up')
                blend[:, :, :, :ramp_len] *= ramp.view(1, 1, 1, -1)

            if x_end < w and overlap_latent > 0:
                ramp_len = min(overlap_latent, tw)
                ramp = self._make_linear_ramp(ramp_len, 'down')
                blend[:, :, :, -ramp_len:] *= ramp.view(1, 1, 1, -1)

            if y > 0 and overlap_latent > 0:
                ramp_len = min(overlap_latent, th)
                ramp = self._make_linear_ramp(ramp_len, 'up')
                blend[:, :, :ramp_len, :] *= ramp.view(1, 1, -1, 1)

            if y_end < h and overlap_latent > 0:
                ramp_len = min(overlap_latent, th)
                ramp = self._make_linear_ramp(ramp_len, 'down')
                blend[:, :, -ramp_len:, :] *= ramp.view(1, 1, -1, 1)

            output[:, :, oy:oy_end, ox:ox_end] += encoded * blend
            weights[:, :, oy:oy_end, ox:ox_end] += blend

            pbar.update(1)

        result = output / weights.clamp(min=1e-8)

        return result

    def decode_tiled_(self, samples, tile_x=None, tile_y=None, overlap=None):
        """Tiled VAE decode - simple and memory efficient"""
        b, c, h, w = samples.shape

        # Use conservative defaults
        if tile_x is None:
            tile_x = self.DEFAULT_TILE_SIZE_LATENT
        if tile_y is None:
            tile_y = self.DEFAULT_TILE_SIZE_LATENT
        if overlap is None:
            overlap = self.DEFAULT_OVERLAP_LATENT

        # Clamp to image size
        tile_x = min(tile_x, w)
        tile_y = min(tile_y, h)

        # Single tile case
        if w <= tile_x and h <= tile_y:
            return self._decode_tile_to_cpu(samples).to(self.output_device)

        # Clear GPU memory once before starting
        if self.device.type != 'cpu':
            torch.cuda.empty_cache()

        result = self._decode_tiled_cpu_accumulate(samples, tile_x, tile_y, overlap)

        return result.to(self.output_device)

    def encode_tiled_(self, pixel_samples, tile_x=None, tile_y=None, overlap=None):
        """Tiled VAE encode - simple and memory efficient"""
        b, c, h, w = pixel_samples.shape

        # Defaults in pixel space
        if tile_x is None:
            tile_x = self.DEFAULT_TILE_SIZE_LATENT * self.downscale_ratio  # 512
        if tile_y is None:
            tile_y = self.DEFAULT_TILE_SIZE_LATENT * self.downscale_ratio  # 512
        if overlap is None:
            overlap = self.DEFAULT_OVERLAP_LATENT * self.downscale_ratio   # 64

        tile_x = min(tile_x, w)
        tile_y = min(tile_y, h)

        if w <= tile_x and h <= tile_y:
            return self._encode_tile_to_cpu(pixel_samples).to(self.output_device)

        if self.device.type != 'cpu':
            torch.cuda.empty_cache()

        result = self._encode_tiled_cpu_accumulate(pixel_samples, tile_x, tile_y, overlap)

        return result.to(self.output_device)

    def decode_inner(self, samples_in):
        if model_management.VAE_ALWAYS_TILED:
            return self.decode_tiled(samples_in).to(self.output_device)

        try:
            memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
            free_memory = model_management.get_free_memory(self.device)
            batch_number = int(free_memory / memory_used)
            batch_number = max(1, batch_number)

            pixel_samples = torch.empty(
                (samples_in.shape[0], 3,
                 round(samples_in.shape[2] * self.downscale_ratio),
                 round(samples_in.shape[3] * self.downscale_ratio)),
                device=self.output_device
            )

            for x in range(0, samples_in.shape[0], batch_number):
                samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
                pixel_samples[x:x+batch_number] = torch.clamp(
                    (self.first_stage_model.decode(samples).to(self.output_device).float() + 1.0) / 2.0,
                    min=0.0, max=1.0
                )

        except model_management.OOM_EXCEPTION:
            print("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
            if self.device.type != 'cpu':
                torch.cuda.empty_cache()
            pixel_samples = self.decode_tiled_(samples_in)

        pixel_samples = pixel_samples.to(self.output_device).movedim(1, -1)
        return pixel_samples

    def decode(self, samples_in):
        wrapper = self.patcher.model_options.get('model_vae_decode_wrapper', None)
        if wrapper is None:
            return self.decode_inner(samples_in)
        else:
            return wrapper(self.decode_inner, samples_in)

    def decode_tiled(self, samples, tile_x=None, tile_y=None, overlap=None):
        """Public tiled decode method"""
        model_management.load_model_gpu(self.patcher)
        output = self.decode_tiled_(samples, tile_x, tile_y, overlap)
        return output.movedim(1, -1)

    def encode_inner(self, pixel_samples):
        if model_management.VAE_ALWAYS_TILED:
            return self.encode_tiled(pixel_samples)

        regulation = self.patcher.model_options.get("model_vae_regulation", None)
        pixel_samples = pixel_samples.movedim(-1, 1)

        try:
            memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
            free_memory = model_management.get_free_memory(self.device)
            batch_number = int(free_memory / memory_used)
            batch_number = max(1, batch_number)

            samples = torch.empty(
                (pixel_samples.shape[0], self.latent_channels,
                 round(pixel_samples.shape[2] // self.downscale_ratio),
                 round(pixel_samples.shape[3] // self.downscale_ratio)),
                device=self.output_device
            )

            for x in range(0, pixel_samples.shape[0], batch_number):
                pixels_in = (2.0 * pixel_samples[x:x+batch_number] - 1.0).to(self.vae_dtype).to(self.device)
                samples[x:x+batch_number] = self.first_stage_model.encode(pixels_in, regulation).to(self.output_device).float()

        except model_management.OOM_EXCEPTION:
            print("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.")
            if self.device.type != 'cpu':
                torch.cuda.empty_cache()
            samples = self.encode_tiled_(pixel_samples)

        return samples

    def encode(self, pixel_samples):
        wrapper = self.patcher.model_options.get('model_vae_encode_wrapper', None)
        if wrapper is None:
            return self.encode_inner(pixel_samples)
        else:
            return wrapper(self.encode_inner, pixel_samples)

    def encode_tiled(self, pixel_samples, tile_x=None, tile_y=None, overlap=None):
        """Public tiled encode method"""
        model_management.load_model_gpu(self.patcher)
        pixel_samples = pixel_samples.movedim(-1, 1)
        samples = self.encode_tiled_(pixel_samples, tile_x=tile_x, tile_y=tile_y, overlap=overlap)
        return samples

    def get_sd(self):
        return self.first_stage_model.state_dict()


class StyleModel:
    def __init__(self, model, device="cpu"):
        self.model = model

    def get_cond(self, input):
        return self.model(input.last_hidden_state)


def load_style_model(ckpt_path):
    model_data = ldm_patched.modules.utils.load_torch_file(ckpt_path, safe_load=True)
    keys = model_data.keys()
    if "style_embedding" in keys:
        model = ldm_patched.t2ia.adapter.StyleAdapter(width=1024, context_dim=768, num_head=8, n_layes=3, num_token=8)
    else:
        raise Exception("invalid style model {}".format(ckpt_path))
    model.load_state_dict(model_data)
    return StyleModel(model)


def load_clip(ckpt_paths, embedding_directory=None):
    clip_data = []
    for p in ckpt_paths:
        clip_data.append(ldm_patched.modules.utils.load_torch_file(p, safe_load=True))

    class EmptyClass:
        pass

    for i in range(len(clip_data)):
        if "transformer.resblocks.0.ln_1.weight" in clip_data[i]:
            clip_data[i] = ldm_patched.modules.utils.transformers_convert(clip_data[i], "", "text_model.", 32)

    clip_target = EmptyClass()
    clip_target.params = {}

    if len(clip_data) == 1:
        if "text_model.encoder.layers.30.mlp.fc1.weight" in clip_data[0]:
            clip_target.clip = sdxl_clip.SDXLRefinerClipModel
            clip_target.tokenizer = sdxl_clip.SDXLTokenizer
        elif "text_model.encoder.layers.22.mlp.fc1.weight" in clip_data[0]:
            clip_target.clip = sd2_clip.SD2ClipModel
            clip_target.tokenizer = sd2_clip.SD2Tokenizer
        else:
            clip_target.clip = sd1_clip.SD1ClipModel
            clip_target.tokenizer = sd1_clip.SD1Tokenizer
    else:
        clip_target.clip = sdxl_clip.SDXLClipModel
        clip_target.tokenizer = sdxl_clip.SDXLTokenizer

    clip = CLIP(clip_target, embedding_directory=embedding_directory)
    for c in clip_data:
        m, u = clip.load_sd(c)
        if len(m) > 0:
            print("clip missing:", m)
        if len(u) > 0:
            print("clip unexpected:", u)
    return clip


def load_gligen(ckpt_path):
    data = ldm_patched.modules.utils.load_torch_file(ckpt_path, safe_load=True)
    model = gligen.load_gligen(data)
    if model_management.should_use_fp16():
        model = model.half()
    return ldm_patched.modules.model_patcher.ModelPatcher(
        model,
        load_device=model_management.get_torch_device(),
        offload_device=model_management.unet_offload_device()
    )


def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_clip=True, embedding_directory=None, state_dict=None, config=None):
    if config is None:
        with open(config_path, 'r') as stream:
            config = yaml.safe_load(stream)

    model_config_params = config['model']['params']
    clip_config = model_config_params['cond_stage_config']
    scale_factor = model_config_params['scale_factor']
    vae_config = model_config_params['first_stage_config']

    fp16 = False
    if "unet_config" in model_config_params:
        if "params" in model_config_params["unet_config"]:
            unet_config = model_config_params["unet_config"]["params"]
            if "use_fp16" in unet_config:
                fp16 = unet_config.pop("use_fp16")
                if fp16:
                    unet_config["dtype"] = torch.float16

    noise_aug_config = None
    if "noise_aug_config" in model_config_params:
        noise_aug_config = model_config_params["noise_aug_config"]

    model_type = model_base.ModelType.EPS

    if "parameterization" in model_config_params:
        if model_config_params["parameterization"] == "v":
            model_type = model_base.ModelType.V_PREDICTION

    clip = None
    vae = None

    class WeightsLoader(torch.nn.Module):
        pass

    if state_dict is None:
        state_dict = ldm_patched.modules.utils.load_torch_file(ckpt_path)

    class EmptyClass:
        pass

    model_config = ldm_patched.modules.supported_models_base.BASE({})

    from . import latent_formats
    model_config.latent_format = latent_formats.SD15(scale_factor=scale_factor)
    model_config.unet_config = model_detection.convert_config(unet_config)

    if config['model']["target"].endswith("ImageEmbeddingConditionedLatentDiffusion"):
        model = model_base.SD21UNCLIP(model_config, noise_aug_config["params"], model_type=model_type)
    else:
        model = model_base.BaseModel(model_config, model_type=model_type)

    if config['model']["target"].endswith("LatentInpaintDiffusion"):
        model.set_inpaint()

    if fp16:
        model = model.half()

    offload_device = model_management.unet_offload_device()
    model = model.to(offload_device)
    model.load_model_weights(state_dict, "model.diffusion_model.")

    if output_vae:
        vae_sd = ldm_patched.modules.utils.state_dict_prefix_replace(state_dict, {"first_stage_model.": ""}, filter_keys=True)
        vae = VAE(sd=vae_sd, config=vae_config)

    if output_clip:
        w = WeightsLoader()
        clip_target = EmptyClass()
        clip_target.params = clip_config.get("params", {})
        if clip_config["target"].endswith("FrozenOpenCLIPEmbedder"):
            clip_target.clip = sd2_clip.SD2ClipModel
            clip_target.tokenizer = sd2_clip.SD2Tokenizer
            clip = CLIP(clip_target, embedding_directory=embedding_directory)
            w.cond_stage_model = clip.cond_stage_model.clip_h
        elif clip_config["target"].endswith("FrozenCLIPEmbedder"):
            clip_target.clip = sd1_clip.SD1ClipModel
            clip_target.tokenizer = sd1_clip.SD1Tokenizer
            clip = CLIP(clip_target, embedding_directory=embedding_directory)
            w.cond_stage_model = clip.cond_stage_model.clip_l
        load_clip_weights(w, state_dict)

    return (
        ldm_patched.modules.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device),
        clip,
        vae
    )


def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True):
    sd = ldm_patched.modules.utils.load_torch_file(ckpt_path)
    sd_keys = sd.keys()
    clip = None
    clipvision = None
    vae = None
    model = None
    model_patcher = None
    clip_target = None

    parameters = ldm_patched.modules.utils.calculate_parameters(sd, "model.diffusion_model.")
    unet_dtype = model_management.unet_dtype(model_params=parameters)
    load_device = model_management.get_torch_device()
    manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device)

    class WeightsLoader(torch.nn.Module):
        pass

    model_config = model_detection.model_config_from_unet(sd, "model.diffusion_model.", unet_dtype)
    if model_config is None:
        raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))

    model_config.set_manual_cast(manual_cast_dtype)

    if model_config.clip_vision_prefix is not None:
        if output_clipvision:
            clipvision = clip_vision.load_clipvision_from_sd(sd, model_config.clip_vision_prefix, True)

    if output_model:
        inital_load_device = model_management.unet_inital_load_device(parameters, unet_dtype)
        offload_device = model_management.unet_offload_device()
        model = model_config.get_model(sd, "model.diffusion_model.", device=inital_load_device)
        model.load_model_weights(sd, "model.diffusion_model.")

    if output_vae:
        vae_sd = ldm_patched.modules.utils.state_dict_prefix_replace(sd, {k: "" for k in model_config.vae_key_prefix}, filter_keys=True)
        vae_sd = model_config.process_vae_state_dict(vae_sd)
        vae = VAE(sd=vae_sd)

    if output_clip:
        w = WeightsLoader()
        clip_target = model_config.clip_target()
        if clip_target is not None:
            clip = CLIP(clip_target, embedding_directory=embedding_directory)
            w.cond_stage_model = clip.cond_stage_model
            sd = model_config.process_clip_state_dict(sd)
            load_model_weights(w, sd)

    left_over = sd.keys()
    if len(left_over) > 0:
        print("left over keys:", left_over)

    if output_model:
        model_patcher = ldm_patched.modules.model_patcher.ModelPatcher(
            model,
            load_device=load_device,
            offload_device=model_management.unet_offload_device(),
            current_device=inital_load_device
        )
        if inital_load_device != torch.device("cpu"):
            print("loaded straight to GPU")
            model_management.load_model_gpu(model_patcher)

    return (model_patcher, clip, vae, clipvision)


def load_unet_state_dict(sd):
    parameters = ldm_patched.modules.utils.calculate_parameters(sd)
    unet_dtype = model_management.unet_dtype(model_params=parameters)
    load_device = model_management.get_torch_device()
    manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device)

    if "input_blocks.0.0.weight" in sd:
        model_config = model_detection.model_config_from_unet(sd, "", unet_dtype)
        if model_config is None:
            return None
        new_sd = sd
    else:
        model_config = model_detection.model_config_from_diffusers_unet(sd, unet_dtype)
        if model_config is None:
            return None

        diffusers_keys = ldm_patched.modules.utils.unet_to_diffusers(model_config.unet_config)

        new_sd = {}
        for k in diffusers_keys:
            if k in sd:
                new_sd[diffusers_keys[k]] = sd.pop(k)
            else:
                print(diffusers_keys[k], k)

    offload_device = model_management.unet_offload_device()
    model_config.set_manual_cast(manual_cast_dtype)
    model = model_config.get_model(new_sd, "")
    model = model.to(offload_device)
    model.load_model_weights(new_sd, "")

    left_over = sd.keys()
    if len(left_over) > 0:
        print("left over keys in unet:", left_over)

    return ldm_patched.modules.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device)


def load_unet(unet_path):
    sd = ldm_patched.modules.utils.load_torch_file(unet_path)
    model = load_unet_state_dict(sd)
    if model is None:
        print("ERROR UNSUPPORTED UNET", unet_path)
        raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
    return model


def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, metadata=None):
    clip_sd = None
    load_models = [model]
    if clip is not None:
        load_models.append(clip.load_model())
        clip_sd = clip.get_sd()

    model_management.load_models_gpu(load_models)
    clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None
    sd = model.model.state_dict_for_saving(clip_sd, vae.get_sd(), clip_vision_sd)
    ldm_patched.modules.utils.save_torch_file(sd, output_path, metadata=metadata)
Edit

Pub: 16 Feb 2026 14:27 UTC

Views: 18