ACCC1380 commited on
Commit
3285eaa
1 Parent(s): aa80585

Upload lora-scripts/sd-scripts/library/sdxl_model_util.py with huggingface_hub

Browse files
lora-scripts/sd-scripts/library/sdxl_model_util.py ADDED
@@ -0,0 +1,577 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from accelerate import init_empty_weights
3
+ from accelerate.utils.modeling import set_module_tensor_to_device
4
+ from safetensors.torch import load_file, save_file
5
+ from transformers import CLIPTextModel, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer
6
+ from typing import List
7
+ from diffusers import AutoencoderKL, EulerDiscreteScheduler, UNet2DConditionModel
8
+ from library import model_util
9
+ from library import sdxl_original_unet
10
+ from .utils import setup_logging
11
+ setup_logging()
12
+ import logging
13
+ logger = logging.getLogger(__name__)
14
+
15
+ VAE_SCALE_FACTOR = 0.13025
16
+ MODEL_VERSION_SDXL_BASE_V1_0 = "sdxl_base_v1-0"
17
+
18
+ # Diffusersの設定を読み込むための参照モデル
19
+ DIFFUSERS_REF_MODEL_ID_SDXL = "stabilityai/stable-diffusion-xl-base-1.0"
20
+
21
+ DIFFUSERS_SDXL_UNET_CONFIG = {
22
+ "act_fn": "silu",
23
+ "addition_embed_type": "text_time",
24
+ "addition_embed_type_num_heads": 64,
25
+ "addition_time_embed_dim": 256,
26
+ "attention_head_dim": [5, 10, 20],
27
+ "block_out_channels": [320, 640, 1280],
28
+ "center_input_sample": False,
29
+ "class_embed_type": None,
30
+ "class_embeddings_concat": False,
31
+ "conv_in_kernel": 3,
32
+ "conv_out_kernel": 3,
33
+ "cross_attention_dim": 2048,
34
+ "cross_attention_norm": None,
35
+ "down_block_types": ["DownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D"],
36
+ "downsample_padding": 1,
37
+ "dual_cross_attention": False,
38
+ "encoder_hid_dim": None,
39
+ "encoder_hid_dim_type": None,
40
+ "flip_sin_to_cos": True,
41
+ "freq_shift": 0,
42
+ "in_channels": 4,
43
+ "layers_per_block": 2,
44
+ "mid_block_only_cross_attention": None,
45
+ "mid_block_scale_factor": 1,
46
+ "mid_block_type": "UNetMidBlock2DCrossAttn",
47
+ "norm_eps": 1e-05,
48
+ "norm_num_groups": 32,
49
+ "num_attention_heads": None,
50
+ "num_class_embeds": None,
51
+ "only_cross_attention": False,
52
+ "out_channels": 4,
53
+ "projection_class_embeddings_input_dim": 2816,
54
+ "resnet_out_scale_factor": 1.0,
55
+ "resnet_skip_time_act": False,
56
+ "resnet_time_scale_shift": "default",
57
+ "sample_size": 128,
58
+ "time_cond_proj_dim": None,
59
+ "time_embedding_act_fn": None,
60
+ "time_embedding_dim": None,
61
+ "time_embedding_type": "positional",
62
+ "timestep_post_act": None,
63
+ "transformer_layers_per_block": [1, 2, 10],
64
+ "up_block_types": ["CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"],
65
+ "upcast_attention": False,
66
+ "use_linear_projection": True,
67
+ }
68
+
69
+
70
+ def convert_sdxl_text_encoder_2_checkpoint(checkpoint, max_length):
71
+ SDXL_KEY_PREFIX = "conditioner.embedders.1.model."
72
+
73
+ # SD2のと、基本的には同じ。logit_scaleを後で使うので、それを追加で返す
74
+ # logit_scaleはcheckpointの保存時に使用する
75
+ def convert_key(key):
76
+ # common conversion
77
+ key = key.replace(SDXL_KEY_PREFIX + "transformer.", "text_model.encoder.")
78
+ key = key.replace(SDXL_KEY_PREFIX, "text_model.")
79
+
80
+ if "resblocks" in key:
81
+ # resblocks conversion
82
+ key = key.replace(".resblocks.", ".layers.")
83
+ if ".ln_" in key:
84
+ key = key.replace(".ln_", ".layer_norm")
85
+ elif ".mlp." in key:
86
+ key = key.replace(".c_fc.", ".fc1.")
87
+ key = key.replace(".c_proj.", ".fc2.")
88
+ elif ".attn.out_proj" in key:
89
+ key = key.replace(".attn.out_proj.", ".self_attn.out_proj.")
90
+ elif ".attn.in_proj" in key:
91
+ key = None # 特殊なので後で処理する
92
+ else:
93
+ raise ValueError(f"unexpected key in SD: {key}")
94
+ elif ".positional_embedding" in key:
95
+ key = key.replace(".positional_embedding", ".embeddings.position_embedding.weight")
96
+ elif ".text_projection" in key:
97
+ key = key.replace("text_model.text_projection", "text_projection.weight")
98
+ elif ".logit_scale" in key:
99
+ key = None # 後で処理する
100
+ elif ".token_embedding" in key:
101
+ key = key.replace(".token_embedding.weight", ".embeddings.token_embedding.weight")
102
+ elif ".ln_final" in key:
103
+ key = key.replace(".ln_final", ".final_layer_norm")
104
+ # ckpt from comfy has this key: text_model.encoder.text_model.embeddings.position_ids
105
+ elif ".embeddings.position_ids" in key:
106
+ key = None # remove this key: position_ids is not used in newer transformers
107
+ return key
108
+
109
+ keys = list(checkpoint.keys())
110
+ new_sd = {}
111
+ for key in keys:
112
+ new_key = convert_key(key)
113
+ if new_key is None:
114
+ continue
115
+ new_sd[new_key] = checkpoint[key]
116
+
117
+ # attnの変換
118
+ for key in keys:
119
+ if ".resblocks" in key and ".attn.in_proj_" in key:
120
+ # 三つに分割
121
+ values = torch.chunk(checkpoint[key], 3)
122
+
123
+ key_suffix = ".weight" if "weight" in key else ".bias"
124
+ key_pfx = key.replace(SDXL_KEY_PREFIX + "transformer.resblocks.", "text_model.encoder.layers.")
125
+ key_pfx = key_pfx.replace("_weight", "")
126
+ key_pfx = key_pfx.replace("_bias", "")
127
+ key_pfx = key_pfx.replace(".attn.in_proj", ".self_attn.")
128
+ new_sd[key_pfx + "q_proj" + key_suffix] = values[0]
129
+ new_sd[key_pfx + "k_proj" + key_suffix] = values[1]
130
+ new_sd[key_pfx + "v_proj" + key_suffix] = values[2]
131
+
132
+ # logit_scale はDiffusersには含まれないが、保存時に戻したいので別途返す
133
+ logit_scale = checkpoint.get(SDXL_KEY_PREFIX + "logit_scale", None)
134
+
135
+ # temporary workaround for text_projection.weight.weight for Playground-v2
136
+ if "text_projection.weight.weight" in new_sd:
137
+ logger.info("convert_sdxl_text_encoder_2_checkpoint: convert text_projection.weight.weight to text_projection.weight")
138
+ new_sd["text_projection.weight"] = new_sd["text_projection.weight.weight"]
139
+ del new_sd["text_projection.weight.weight"]
140
+
141
+ return new_sd, logit_scale
142
+
143
+
144
+ # load state_dict without allocating new tensors
145
+ def _load_state_dict_on_device(model, state_dict, device, dtype=None):
146
+ # dtype will use fp32 as default
147
+ missing_keys = list(model.state_dict().keys() - state_dict.keys())
148
+ unexpected_keys = list(state_dict.keys() - model.state_dict().keys())
149
+
150
+ # similar to model.load_state_dict()
151
+ if not missing_keys and not unexpected_keys:
152
+ for k in list(state_dict.keys()):
153
+ set_module_tensor_to_device(model, k, device, value=state_dict.pop(k), dtype=dtype)
154
+ return "<All keys matched successfully>"
155
+
156
+ # error_msgs
157
+ error_msgs: List[str] = []
158
+ if missing_keys:
159
+ error_msgs.insert(0, "Missing key(s) in state_dict: {}. ".format(", ".join('"{}"'.format(k) for k in missing_keys)))
160
+ if unexpected_keys:
161
+ error_msgs.insert(0, "Unexpected key(s) in state_dict: {}. ".format(", ".join('"{}"'.format(k) for k in unexpected_keys)))
162
+
163
+ raise RuntimeError("Error(s) in loading state_dict for {}:\n\t{}".format(model.__class__.__name__, "\n\t".join(error_msgs)))
164
+
165
+
166
+ def load_models_from_sdxl_checkpoint(model_version, ckpt_path, map_location, dtype=None):
167
+ # model_version is reserved for future use
168
+ # dtype is used for full_fp16/bf16 integration. Text Encoder will remain fp32, because it runs on CPU when caching
169
+
170
+ # Load the state dict
171
+ if model_util.is_safetensors(ckpt_path):
172
+ checkpoint = None
173
+ try:
174
+ state_dict = load_file(ckpt_path, device=map_location)
175
+ except:
176
+ state_dict = load_file(ckpt_path) # prevent device invalid Error
177
+ epoch = None
178
+ global_step = None
179
+ else:
180
+ checkpoint = torch.load(ckpt_path, map_location=map_location)
181
+ if "state_dict" in checkpoint:
182
+ state_dict = checkpoint["state_dict"]
183
+ epoch = checkpoint.get("epoch", 0)
184
+ global_step = checkpoint.get("global_step", 0)
185
+ else:
186
+ state_dict = checkpoint
187
+ epoch = 0
188
+ global_step = 0
189
+ checkpoint = None
190
+
191
+ # U-Net
192
+ logger.info("building U-Net")
193
+ with init_empty_weights():
194
+ unet = sdxl_original_unet.SdxlUNet2DConditionModel()
195
+
196
+ logger.info("loading U-Net from checkpoint")
197
+ unet_sd = {}
198
+ for k in list(state_dict.keys()):
199
+ if k.startswith("model.diffusion_model."):
200
+ unet_sd[k.replace("model.diffusion_model.", "")] = state_dict.pop(k)
201
+ info = _load_state_dict_on_device(unet, unet_sd, device=map_location, dtype=dtype)
202
+ logger.info(f"U-Net: {info}")
203
+
204
+ # Text Encoders
205
+ logger.info("building text encoders")
206
+
207
+ # Text Encoder 1 is same to Stability AI's SDXL
208
+ text_model1_cfg = CLIPTextConfig(
209
+ vocab_size=49408,
210
+ hidden_size=768,
211
+ intermediate_size=3072,
212
+ num_hidden_layers=12,
213
+ num_attention_heads=12,
214
+ max_position_embeddings=77,
215
+ hidden_act="quick_gelu",
216
+ layer_norm_eps=1e-05,
217
+ dropout=0.0,
218
+ attention_dropout=0.0,
219
+ initializer_range=0.02,
220
+ initializer_factor=1.0,
221
+ pad_token_id=1,
222
+ bos_token_id=0,
223
+ eos_token_id=2,
224
+ model_type="clip_text_model",
225
+ projection_dim=768,
226
+ # torch_dtype="float32",
227
+ # transformers_version="4.25.0.dev0",
228
+ )
229
+ with init_empty_weights():
230
+ text_model1 = CLIPTextModel._from_config(text_model1_cfg)
231
+
232
+ # Text Encoder 2 is different from Stability AI's SDXL. SDXL uses open clip, but we use the model from HuggingFace.
233
+ # Note: Tokenizer from HuggingFace is different from SDXL. We must use open clip's tokenizer.
234
+ text_model2_cfg = CLIPTextConfig(
235
+ vocab_size=49408,
236
+ hidden_size=1280,
237
+ intermediate_size=5120,
238
+ num_hidden_layers=32,
239
+ num_attention_heads=20,
240
+ max_position_embeddings=77,
241
+ hidden_act="gelu",
242
+ layer_norm_eps=1e-05,
243
+ dropout=0.0,
244
+ attention_dropout=0.0,
245
+ initializer_range=0.02,
246
+ initializer_factor=1.0,
247
+ pad_token_id=1,
248
+ bos_token_id=0,
249
+ eos_token_id=2,
250
+ model_type="clip_text_model",
251
+ projection_dim=1280,
252
+ # torch_dtype="float32",
253
+ # transformers_version="4.25.0.dev0",
254
+ )
255
+ with init_empty_weights():
256
+ text_model2 = CLIPTextModelWithProjection(text_model2_cfg)
257
+
258
+ logger.info("loading text encoders from checkpoint")
259
+ te1_sd = {}
260
+ te2_sd = {}
261
+ for k in list(state_dict.keys()):
262
+ if k.startswith("conditioner.embedders.0.transformer."):
263
+ te1_sd[k.replace("conditioner.embedders.0.transformer.", "")] = state_dict.pop(k)
264
+ elif k.startswith("conditioner.embedders.1.model."):
265
+ te2_sd[k] = state_dict.pop(k)
266
+
267
+ # 最新の transformers では position_ids を含むとエラーになるので削除 / remove position_ids for latest transformers
268
+ if "text_model.embeddings.position_ids" in te1_sd:
269
+ te1_sd.pop("text_model.embeddings.position_ids")
270
+
271
+ info1 = _load_state_dict_on_device(text_model1, te1_sd, device=map_location) # remain fp32
272
+ logger.info(f"text encoder 1: {info1}")
273
+
274
+ converted_sd, logit_scale = convert_sdxl_text_encoder_2_checkpoint(te2_sd, max_length=77)
275
+ info2 = _load_state_dict_on_device(text_model2, converted_sd, device=map_location) # remain fp32
276
+ logger.info(f"text encoder 2: {info2}")
277
+
278
+ # prepare vae
279
+ logger.info("building VAE")
280
+ vae_config = model_util.create_vae_diffusers_config()
281
+ with init_empty_weights():
282
+ vae = AutoencoderKL(**vae_config)
283
+
284
+ logger.info("loading VAE from checkpoint")
285
+ converted_vae_checkpoint = model_util.convert_ldm_vae_checkpoint(state_dict, vae_config)
286
+ info = _load_state_dict_on_device(vae, converted_vae_checkpoint, device=map_location, dtype=dtype)
287
+ logger.info(f"VAE: {info}")
288
+
289
+ ckpt_info = (epoch, global_step) if epoch is not None else None
290
+ return text_model1, text_model2, vae, unet, logit_scale, ckpt_info
291
+
292
+
293
+ def make_unet_conversion_map():
294
+ unet_conversion_map_layer = []
295
+
296
+ for i in range(3): # num_blocks is 3 in sdxl
297
+ # loop over downblocks/upblocks
298
+ for j in range(2):
299
+ # loop over resnets/attentions for downblocks
300
+ hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
301
+ sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
302
+ unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
303
+
304
+ if i < 3:
305
+ # no attention layers in down_blocks.3
306
+ hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
307
+ sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."
308
+ unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
309
+
310
+ for j in range(3):
311
+ # loop over resnets/attentions for upblocks
312
+ hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
313
+ sd_up_res_prefix = f"output_blocks.{3*i + j}.0."
314
+ unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
315
+
316
+ # if i > 0: commentout for sdxl
317
+ # no attention layers in up_blocks.0
318
+ hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
319
+ sd_up_atn_prefix = f"output_blocks.{3*i + j}.1."
320
+ unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
321
+
322
+ if i < 3:
323
+ # no downsample in down_blocks.3
324
+ hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
325
+ sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."
326
+ unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
327
+
328
+ # no upsample in up_blocks.3
329
+ hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
330
+ sd_upsample_prefix = f"output_blocks.{3*i + 2}.{2}." # change for sdxl
331
+ unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
332
+
333
+ hf_mid_atn_prefix = "mid_block.attentions.0."
334
+ sd_mid_atn_prefix = "middle_block.1."
335
+ unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
336
+
337
+ for j in range(2):
338
+ hf_mid_res_prefix = f"mid_block.resnets.{j}."
339
+ sd_mid_res_prefix = f"middle_block.{2*j}."
340
+ unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
341
+
342
+ unet_conversion_map_resnet = [
343
+ # (stable-diffusion, HF Diffusers)
344
+ ("in_layers.0.", "norm1."),
345
+ ("in_layers.2.", "conv1."),
346
+ ("out_layers.0.", "norm2."),
347
+ ("out_layers.3.", "conv2."),
348
+ ("emb_layers.1.", "time_emb_proj."),
349
+ ("skip_connection.", "conv_shortcut."),
350
+ ]
351
+
352
+ unet_conversion_map = []
353
+ for sd, hf in unet_conversion_map_layer:
354
+ if "resnets" in hf:
355
+ for sd_res, hf_res in unet_conversion_map_resnet:
356
+ unet_conversion_map.append((sd + sd_res, hf + hf_res))
357
+ else:
358
+ unet_conversion_map.append((sd, hf))
359
+
360
+ for j in range(2):
361
+ hf_time_embed_prefix = f"time_embedding.linear_{j+1}."
362
+ sd_time_embed_prefix = f"time_embed.{j*2}."
363
+ unet_conversion_map.append((sd_time_embed_prefix, hf_time_embed_prefix))
364
+
365
+ for j in range(2):
366
+ hf_label_embed_prefix = f"add_embedding.linear_{j+1}."
367
+ sd_label_embed_prefix = f"label_emb.0.{j*2}."
368
+ unet_conversion_map.append((sd_label_embed_prefix, hf_label_embed_prefix))
369
+
370
+ unet_conversion_map.append(("input_blocks.0.0.", "conv_in."))
371
+ unet_conversion_map.append(("out.0.", "conv_norm_out."))
372
+ unet_conversion_map.append(("out.2.", "conv_out."))
373
+
374
+ return unet_conversion_map
375
+
376
+
377
+ def convert_diffusers_unet_state_dict_to_sdxl(du_sd):
378
+ unet_conversion_map = make_unet_conversion_map()
379
+
380
+ conversion_map = {hf: sd for sd, hf in unet_conversion_map}
381
+ return convert_unet_state_dict(du_sd, conversion_map)
382
+
383
+
384
+ def convert_unet_state_dict(src_sd, conversion_map):
385
+ converted_sd = {}
386
+ for src_key, value in src_sd.items():
387
+ # さすがに全部回すのは時間がかかるので右から要素を削りつつprefixを探す
388
+ src_key_fragments = src_key.split(".")[:-1] # remove weight/bias
389
+ while len(src_key_fragments) > 0:
390
+ src_key_prefix = ".".join(src_key_fragments) + "."
391
+ if src_key_prefix in conversion_map:
392
+ converted_prefix = conversion_map[src_key_prefix]
393
+ converted_key = converted_prefix + src_key[len(src_key_prefix) :]
394
+ converted_sd[converted_key] = value
395
+ break
396
+ src_key_fragments.pop(-1)
397
+ assert len(src_key_fragments) > 0, f"key {src_key} not found in conversion map"
398
+
399
+ return converted_sd
400
+
401
+
402
+ def convert_sdxl_unet_state_dict_to_diffusers(sd):
403
+ unet_conversion_map = make_unet_conversion_map()
404
+
405
+ conversion_dict = {sd: hf for sd, hf in unet_conversion_map}
406
+ return convert_unet_state_dict(sd, conversion_dict)
407
+
408
+
409
+ def convert_text_encoder_2_state_dict_to_sdxl(checkpoint, logit_scale):
410
+ def convert_key(key):
411
+ # position_idsの除去
412
+ if ".position_ids" in key:
413
+ return None
414
+
415
+ # common
416
+ key = key.replace("text_model.encoder.", "transformer.")
417
+ key = key.replace("text_model.", "")
418
+ if "layers" in key:
419
+ # resblocks conversion
420
+ key = key.replace(".layers.", ".resblocks.")
421
+ if ".layer_norm" in key:
422
+ key = key.replace(".layer_norm", ".ln_")
423
+ elif ".mlp." in key:
424
+ key = key.replace(".fc1.", ".c_fc.")
425
+ key = key.replace(".fc2.", ".c_proj.")
426
+ elif ".self_attn.out_proj" in key:
427
+ key = key.replace(".self_attn.out_proj.", ".attn.out_proj.")
428
+ elif ".self_attn." in key:
429
+ key = None # 特殊なので後で処理する
430
+ else:
431
+ raise ValueError(f"unexpected key in DiffUsers model: {key}")
432
+ elif ".position_embedding" in key:
433
+ key = key.replace("embeddings.position_embedding.weight", "positional_embedding")
434
+ elif ".token_embedding" in key:
435
+ key = key.replace("embeddings.token_embedding.weight", "token_embedding.weight")
436
+ elif "text_projection" in key: # no dot in key
437
+ key = key.replace("text_projection.weight", "text_projection")
438
+ elif "final_layer_norm" in key:
439
+ key = key.replace("final_layer_norm", "ln_final")
440
+ return key
441
+
442
+ keys = list(checkpoint.keys())
443
+ new_sd = {}
444
+ for key in keys:
445
+ new_key = convert_key(key)
446
+ if new_key is None:
447
+ continue
448
+ new_sd[new_key] = checkpoint[key]
449
+
450
+ # attnの変換
451
+ for key in keys:
452
+ if "layers" in key and "q_proj" in key:
453
+ # 三つを結合
454
+ key_q = key
455
+ key_k = key.replace("q_proj", "k_proj")
456
+ key_v = key.replace("q_proj", "v_proj")
457
+
458
+ value_q = checkpoint[key_q]
459
+ value_k = checkpoint[key_k]
460
+ value_v = checkpoint[key_v]
461
+ value = torch.cat([value_q, value_k, value_v])
462
+
463
+ new_key = key.replace("text_model.encoder.layers.", "transformer.resblocks.")
464
+ new_key = new_key.replace(".self_attn.q_proj.", ".attn.in_proj_")
465
+ new_sd[new_key] = value
466
+
467
+ if logit_scale is not None:
468
+ new_sd["logit_scale"] = logit_scale
469
+
470
+ return new_sd
471
+
472
+
473
+ def save_stable_diffusion_checkpoint(
474
+ output_file,
475
+ text_encoder1,
476
+ text_encoder2,
477
+ unet,
478
+ epochs,
479
+ steps,
480
+ ckpt_info,
481
+ vae,
482
+ logit_scale,
483
+ metadata,
484
+ save_dtype=None,
485
+ ):
486
+ state_dict = {}
487
+
488
+ def update_sd(prefix, sd):
489
+ for k, v in sd.items():
490
+ key = prefix + k
491
+ if save_dtype is not None:
492
+ v = v.detach().clone().to("cpu").to(save_dtype)
493
+ state_dict[key] = v
494
+
495
+ # Convert the UNet model
496
+ update_sd("model.diffusion_model.", unet.state_dict())
497
+
498
+ # Convert the text encoders
499
+ update_sd("conditioner.embedders.0.transformer.", text_encoder1.state_dict())
500
+
501
+ text_enc2_dict = convert_text_encoder_2_state_dict_to_sdxl(text_encoder2.state_dict(), logit_scale)
502
+ update_sd("conditioner.embedders.1.model.", text_enc2_dict)
503
+
504
+ # Convert the VAE
505
+ vae_dict = model_util.convert_vae_state_dict(vae.state_dict())
506
+ update_sd("first_stage_model.", vae_dict)
507
+
508
+ # Put together new checkpoint
509
+ key_count = len(state_dict.keys())
510
+ new_ckpt = {"state_dict": state_dict}
511
+
512
+ # epoch and global_step are sometimes not int
513
+ if ckpt_info is not None:
514
+ epochs += ckpt_info[0]
515
+ steps += ckpt_info[1]
516
+
517
+ new_ckpt["epoch"] = epochs
518
+ new_ckpt["global_step"] = steps
519
+
520
+ if model_util.is_safetensors(output_file):
521
+ save_file(state_dict, output_file, metadata)
522
+ else:
523
+ torch.save(new_ckpt, output_file)
524
+
525
+ return key_count
526
+
527
+
528
+ def save_diffusers_checkpoint(
529
+ output_dir, text_encoder1, text_encoder2, unet, pretrained_model_name_or_path, vae=None, use_safetensors=False, save_dtype=None
530
+ ):
531
+ from diffusers import StableDiffusionXLPipeline
532
+
533
+ # convert U-Net
534
+ unet_sd = unet.state_dict()
535
+ du_unet_sd = convert_sdxl_unet_state_dict_to_diffusers(unet_sd)
536
+
537
+ diffusers_unet = UNet2DConditionModel(**DIFFUSERS_SDXL_UNET_CONFIG)
538
+ if save_dtype is not None:
539
+ diffusers_unet.to(save_dtype)
540
+ diffusers_unet.load_state_dict(du_unet_sd)
541
+
542
+ # create pipeline to save
543
+ if pretrained_model_name_or_path is None:
544
+ pretrained_model_name_or_path = DIFFUSERS_REF_MODEL_ID_SDXL
545
+
546
+ scheduler = EulerDiscreteScheduler.from_pretrained(pretrained_model_name_or_path, subfolder="scheduler")
547
+ tokenizer1 = CLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder="tokenizer")
548
+ tokenizer2 = CLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder="tokenizer_2")
549
+ if vae is None:
550
+ vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder="vae")
551
+
552
+ # prevent local path from being saved
553
+ def remove_name_or_path(model):
554
+ if hasattr(model, "config"):
555
+ model.config._name_or_path = None
556
+ model.config._name_or_path = None
557
+
558
+ remove_name_or_path(diffusers_unet)
559
+ remove_name_or_path(text_encoder1)
560
+ remove_name_or_path(text_encoder2)
561
+ remove_name_or_path(scheduler)
562
+ remove_name_or_path(tokenizer1)
563
+ remove_name_or_path(tokenizer2)
564
+ remove_name_or_path(vae)
565
+
566
+ pipeline = StableDiffusionXLPipeline(
567
+ unet=diffusers_unet,
568
+ text_encoder=text_encoder1,
569
+ text_encoder_2=text_encoder2,
570
+ vae=vae,
571
+ scheduler=scheduler,
572
+ tokenizer=tokenizer1,
573
+ tokenizer_2=tokenizer2,
574
+ )
575
+ if save_dtype is not None:
576
+ pipeline.to(None, save_dtype)
577
+ pipeline.save_pretrained(output_dir, safe_serialization=use_safetensors)