File size: 21,003 Bytes
436faa6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import contextlib
import math

from comfy import model_management
from .ldm.util import instantiate_from_config
from .ldm.models.autoencoder import AutoencoderKL
import yaml

import comfy.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 comfy.model_patcher
import comfy.lora
import comfy.t2i_adapter.adapter
import comfy.supported_models_base

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("missing", 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 = comfy.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):
    key_map = comfy.lora.model_lora_keys_unet(model.model)
    key_map = comfy.lora.model_lora_keys_clip(clip.cond_stage_model, key_map)
    loaded = comfy.lora.load_lora(lora, key_map)
    new_modelpatcher = model.clone()
    k = new_modelpatcher.add_patches(loaded, strength_model)
    new_clip = clip.clone()
    k1 = new_clip.add_patches(loaded, strength_clip)
    k = set(k)
    k1 = set(k1)
    for x in loaded:
        if (x not in k) and (x not in k1):
            print("NOT LOADED", x)

    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
        if model_management.should_use_fp16(load_device, prioritize_performance=False):
            params['dtype'] = torch.float16
        else:
            params['dtype'] = torch.float32

        self.cond_stage_model = clip(**(params))

        self.tokenizer = tokenizer(embedding_directory=embedding_directory)
        self.patcher = comfy.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):
        if self.layer_idx is not None:
            self.cond_stage_model.clip_layer(self.layer_idx)
        else:
            self.cond_stage_model.reset_clip_layer()

        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:
    def __init__(self, ckpt_path=None, device=None, config=None):
        if config is None:
            #default SD1.x/SD2.x VAE parameters
            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}
            self.first_stage_model = AutoencoderKL(ddconfig, {'target': 'torch.nn.Identity'}, 4, monitor="val/rec_loss")
        else:
            self.first_stage_model = AutoencoderKL(**(config['params']))
        self.first_stage_model = self.first_stage_model.eval()
        if ckpt_path is not None:
            sd = comfy.utils.load_torch_file(ckpt_path)
            if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
                sd = diffusers_convert.convert_vae_state_dict(sd)
            self.first_stage_model.load_state_dict(sd, strict=False)

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

    def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
        steps = samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x, tile_y, overlap)
        steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x // 2, tile_y * 2, overlap)
        steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x * 2, tile_y // 2, overlap)
        pbar = comfy.utils.ProgressBar(steps)

        decode_fn = lambda a: (self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)) + 1.0).float()
        output = torch.clamp((
            (comfy.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = 8, pbar = pbar) +
            comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = 8, pbar = pbar) +
             comfy.utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = 8, pbar = pbar))
            / 3.0) / 2.0, min=0.0, max=1.0)
        return output

    def encode_tiled_(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
        steps = pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x, tile_y, overlap)
        steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x // 2, tile_y * 2, overlap)
        steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x * 2, tile_y // 2, overlap)
        pbar = comfy.utils.ProgressBar(steps)

        encode_fn = lambda a: self.first_stage_model.encode(2. * a.to(self.vae_dtype).to(self.device) - 1.).sample().float()
        samples = comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar)
        samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar)
        samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar)
        samples /= 3.0
        return samples

    def decode(self, samples_in):
        self.first_stage_model = self.first_stage_model.to(self.device)
        try:
            memory_used = (2562 * samples_in.shape[2] * samples_in.shape[3] * 64) * 1.7
            model_management.free_memory(memory_used, self.device)
            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] * 8), round(samples_in.shape[3] * 8)), device="cpu")
            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) + 1.0) / 2.0, min=0.0, max=1.0).cpu().float()
        except model_management.OOM_EXCEPTION as e:
            print("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
            pixel_samples = self.decode_tiled_(samples_in)

        self.first_stage_model = self.first_stage_model.to(self.offload_device)
        pixel_samples = pixel_samples.cpu().movedim(1,-1)
        return pixel_samples

    def decode_tiled(self, samples, tile_x=64, tile_y=64, overlap = 16):
        self.first_stage_model = self.first_stage_model.to(self.device)
        output = self.decode_tiled_(samples, tile_x, tile_y, overlap)
        self.first_stage_model = self.first_stage_model.to(self.offload_device)
        return output.movedim(1,-1)

    def encode(self, pixel_samples):
        self.first_stage_model = self.first_stage_model.to(self.device)
        pixel_samples = pixel_samples.movedim(-1,1)
        try:
            memory_used = (2078 * pixel_samples.shape[2] * pixel_samples.shape[3]) * 1.7 #NOTE: this constant along with the one in the decode above are estimated from the mem usage for the VAE and could change.
            model_management.free_memory(memory_used, self.device)
            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], 4, round(pixel_samples.shape[2] // 8), round(pixel_samples.shape[3] // 8)), device="cpu")
            for x in range(0, pixel_samples.shape[0], batch_number):
                pixels_in = (2. * pixel_samples[x:x+batch_number] - 1.).to(self.vae_dtype).to(self.device)
                samples[x:x+batch_number] = self.first_stage_model.encode(pixels_in).sample().cpu().float()

        except model_management.OOM_EXCEPTION as e:
            print("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.")
            samples = self.encode_tiled_(pixel_samples)

        self.first_stage_model = self.first_stage_model.to(self.offload_device)
        return samples

    def encode_tiled(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
        self.first_stage_model = self.first_stage_model.to(self.device)
        pixel_samples = pixel_samples.movedim(-1,1)
        samples = self.encode_tiled_(pixel_samples, tile_x=tile_x, tile_y=tile_y, overlap=overlap)
        self.first_stage_model = self.first_stage_model.to(self.offload_device)
        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 = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
    keys = model_data.keys()
    if "style_embedding" in keys:
        model = comfy.t2i_adapter.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(comfy.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] = comfy.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 = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
    model = gligen.load_gligen(data)
    if model_management.should_use_fp16():
        model = model.half()
    return comfy.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):
    #TODO: this function is a mess and should be removed eventually
    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["use_fp16"]

    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 = comfy.utils.load_torch_file(ckpt_path)

    class EmptyClass:
        pass

    model_config = comfy.supported_models_base.BASE({})

    from . import latent_formats
    model_config.latent_format = latent_formats.SD15(scale_factor=scale_factor)
    model_config.unet_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:
        w = WeightsLoader()
        vae = VAE(config=vae_config)
        w.first_stage_model = vae.first_stage_model
        load_model_weights(w, state_dict)

    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
        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
        load_clip_weights(w, state_dict)

    return (comfy.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):
    sd = comfy.utils.load_torch_file(ckpt_path)
    sd_keys = sd.keys()
    clip = None
    clipvision = None
    vae = None
    model = None
    clip_target = None

    parameters = comfy.utils.calculate_parameters(sd, "model.diffusion_model.")
    fp16 = model_management.should_use_fp16(model_params=parameters)

    class WeightsLoader(torch.nn.Module):
        pass

    model_config = model_detection.model_config_from_unet(sd, "model.diffusion_model.", fp16)
    if model_config is None:
        raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))

    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)

    dtype = torch.float32
    if fp16:
        dtype = torch.float16

    inital_load_device = model_management.unet_inital_load_device(parameters, 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 = VAE()
        w = WeightsLoader()
        w.first_stage_model = vae.first_stage_model
        load_model_weights(w, sd)

    if output_clip:
        w = WeightsLoader()
        clip_target = model_config.clip_target()
        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)

    model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_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(unet_path): #load unet in diffusers format
    sd = comfy.utils.load_torch_file(unet_path)
    parameters = comfy.utils.calculate_parameters(sd)
    fp16 = model_management.should_use_fp16(model_params=parameters)
    if "input_blocks.0.0.weight" in sd: #ldm
        model_config = model_detection.model_config_from_unet(sd, "", fp16)
        if model_config is None:
            raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
        new_sd = sd

    else: #diffusers
        model_config = model_detection.model_config_from_diffusers_unet(sd, fp16)
        if model_config is None:
            print("ERROR UNSUPPORTED UNET", unet_path)
            return None

        diffusers_keys = comfy.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 = model_config.get_model(new_sd, "")
    model = model.to(offload_device)
    model.load_model_weights(new_sd, "")
    return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device)

def save_checkpoint(output_path, model, clip, vae, metadata=None):
    model_management.load_models_gpu([model, clip.load_model()])
    sd = model.model.state_dict_for_saving(clip.get_sd(), vae.get_sd())
    comfy.utils.save_torch_file(sd, output_path, metadata=metadata)