ACCC1380 commited on
Commit
acfa326
1 Parent(s): 4026052

Upload lora-scripts/sd-scripts/networks/control_net_lllite_for_train.py with huggingface_hub

Browse files
lora-scripts/sd-scripts/networks/control_net_lllite_for_train.py ADDED
@@ -0,0 +1,505 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # cond_imageをU-Netのforwardで渡すバージョンのControlNet-LLLite検証用実装
2
+ # ControlNet-LLLite implementation for verification with cond_image passed in U-Net's forward
3
+
4
+ import os
5
+ import re
6
+ from typing import Optional, List, Type
7
+ import torch
8
+ from library import sdxl_original_unet
9
+ from library.utils import setup_logging
10
+ setup_logging()
11
+ import logging
12
+ logger = logging.getLogger(__name__)
13
+
14
+ # input_blocksに適用するかどうか / if True, input_blocks are not applied
15
+ SKIP_INPUT_BLOCKS = False
16
+
17
+ # output_blocksに適用するかどうか / if True, output_blocks are not applied
18
+ SKIP_OUTPUT_BLOCKS = True
19
+
20
+ # conv2dに適用するかどうか / if True, conv2d are not applied
21
+ SKIP_CONV2D = False
22
+
23
+ # transformer_blocksのみに適用するかどうか。Trueの場合、ResBlockには適用されない
24
+ # if True, only transformer_blocks are applied, and ResBlocks are not applied
25
+ TRANSFORMER_ONLY = True # if True, SKIP_CONV2D is ignored because conv2d is not used in transformer_blocks
26
+
27
+ # Trueならattn1とattn2にのみ適用し、ffなどには適用しない / if True, apply only to attn1 and attn2, not to ff etc.
28
+ ATTN1_2_ONLY = True
29
+
30
+ # Trueならattn1のQKV、attn2のQにのみ適用する、ATTN1_2_ONLY指定時のみ有効 / if True, apply only to attn1 QKV and attn2 Q, only valid when ATTN1_2_ONLY is specified
31
+ ATTN_QKV_ONLY = True
32
+
33
+ # Trueならattn1やffなどにのみ適用し、attn2などには適用しない / if True, apply only to attn1 and ff, not to attn2
34
+ # ATTN1_2_ONLYと同時にTrueにできない / cannot be True at the same time as ATTN1_2_ONLY
35
+ ATTN1_ETC_ONLY = False # True
36
+
37
+ # transformer_blocksの最大インデックス。Noneなら全てのtransformer_blocksに適用
38
+ # max index of transformer_blocks. if None, apply to all transformer_blocks
39
+ TRANSFORMER_MAX_BLOCK_INDEX = None
40
+
41
+ ORIGINAL_LINEAR = torch.nn.Linear
42
+ ORIGINAL_CONV2D = torch.nn.Conv2d
43
+
44
+
45
+ def add_lllite_modules(module: torch.nn.Module, in_dim: int, depth, cond_emb_dim, mlp_dim) -> None:
46
+ # conditioning1はconditioning imageを embedding する。timestepごとに呼ばれない
47
+ # conditioning1 embeds conditioning image. it is not called for each timestep
48
+ modules = []
49
+ modules.append(ORIGINAL_CONV2D(3, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0)) # to latent (from VAE) size
50
+ if depth == 1:
51
+ modules.append(torch.nn.ReLU(inplace=True))
52
+ modules.append(ORIGINAL_CONV2D(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0))
53
+ elif depth == 2:
54
+ modules.append(torch.nn.ReLU(inplace=True))
55
+ modules.append(ORIGINAL_CONV2D(cond_emb_dim // 2, cond_emb_dim, kernel_size=4, stride=4, padding=0))
56
+ elif depth == 3:
57
+ # kernel size 8は大きすぎるので、4にする / kernel size 8 is too large, so set it to 4
58
+ modules.append(torch.nn.ReLU(inplace=True))
59
+ modules.append(ORIGINAL_CONV2D(cond_emb_dim // 2, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0))
60
+ modules.append(torch.nn.ReLU(inplace=True))
61
+ modules.append(ORIGINAL_CONV2D(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0))
62
+
63
+ module.lllite_conditioning1 = torch.nn.Sequential(*modules)
64
+
65
+ # downで入力の次元数を削減する。LoRAにヒントを得ていることにする
66
+ # midでconditioning image embeddingと入力を結合する
67
+ # upで元の次元数に戻す
68
+ # これらはtimestepごとに呼ばれる
69
+ # reduce the number of input dimensions with down. inspired by LoRA
70
+ # combine conditioning image embedding and input with mid
71
+ # restore to the original dimension with up
72
+ # these are called for each timestep
73
+
74
+ module.lllite_down = torch.nn.Sequential(
75
+ ORIGINAL_LINEAR(in_dim, mlp_dim),
76
+ torch.nn.ReLU(inplace=True),
77
+ )
78
+ module.lllite_mid = torch.nn.Sequential(
79
+ ORIGINAL_LINEAR(mlp_dim + cond_emb_dim, mlp_dim),
80
+ torch.nn.ReLU(inplace=True),
81
+ )
82
+ module.lllite_up = torch.nn.Sequential(
83
+ ORIGINAL_LINEAR(mlp_dim, in_dim),
84
+ )
85
+
86
+ # Zero-Convにする / set to Zero-Conv
87
+ torch.nn.init.zeros_(module.lllite_up[0].weight) # zero conv
88
+
89
+
90
+ class LLLiteLinear(ORIGINAL_LINEAR):
91
+ def __init__(self, in_features: int, out_features: int, **kwargs):
92
+ super().__init__(in_features, out_features, **kwargs)
93
+ self.enabled = False
94
+
95
+ def set_lllite(self, depth, cond_emb_dim, name, mlp_dim, dropout=None, multiplier=1.0):
96
+ self.enabled = True
97
+ self.lllite_name = name
98
+ self.cond_emb_dim = cond_emb_dim
99
+ self.dropout = dropout
100
+ self.multiplier = multiplier # ignored
101
+
102
+ in_dim = self.in_features
103
+ add_lllite_modules(self, in_dim, depth, cond_emb_dim, mlp_dim)
104
+
105
+ self.cond_image = None
106
+ self.cond_emb = None
107
+
108
+ def set_cond_image(self, cond_image):
109
+ self.cond_image = cond_image
110
+ self.cond_emb = None
111
+
112
+ def forward(self, x):
113
+ if not self.enabled:
114
+ return super().forward(x)
115
+
116
+ if self.cond_emb is None:
117
+ self.cond_emb = self.lllite_conditioning1(self.cond_image)
118
+ cx = self.cond_emb
119
+
120
+ # reshape / b,c,h,w -> b,h*w,c
121
+ n, c, h, w = cx.shape
122
+ cx = cx.view(n, c, h * w).permute(0, 2, 1)
123
+
124
+ cx = torch.cat([cx, self.lllite_down(x)], dim=2)
125
+ cx = self.lllite_mid(cx)
126
+
127
+ if self.dropout is not None and self.training:
128
+ cx = torch.nn.functional.dropout(cx, p=self.dropout)
129
+
130
+ cx = self.lllite_up(cx) * self.multiplier
131
+
132
+ x = super().forward(x + cx) # ここで元のモジュールを呼び出す / call the original module here
133
+ return x
134
+
135
+
136
+ class LLLiteConv2d(ORIGINAL_CONV2D):
137
+ def __init__(self, in_channels: int, out_channels: int, kernel_size, **kwargs):
138
+ super().__init__(in_channels, out_channels, kernel_size, **kwargs)
139
+ self.enabled = False
140
+
141
+ def set_lllite(self, depth, cond_emb_dim, name, mlp_dim, dropout=None, multiplier=1.0):
142
+ self.enabled = True
143
+ self.lllite_name = name
144
+ self.cond_emb_dim = cond_emb_dim
145
+ self.dropout = dropout
146
+ self.multiplier = multiplier # ignored
147
+
148
+ in_dim = self.in_channels
149
+ add_lllite_modules(self, in_dim, depth, cond_emb_dim, mlp_dim)
150
+
151
+ self.cond_image = None
152
+ self.cond_emb = None
153
+
154
+ def set_cond_image(self, cond_image):
155
+ self.cond_image = cond_image
156
+ self.cond_emb = None
157
+
158
+ def forward(self, x): # , cond_image=None):
159
+ if not self.enabled:
160
+ return super().forward(x)
161
+
162
+ if self.cond_emb is None:
163
+ self.cond_emb = self.lllite_conditioning1(self.cond_image)
164
+ cx = self.cond_emb
165
+
166
+ cx = torch.cat([cx, self.down(x)], dim=1)
167
+ cx = self.mid(cx)
168
+
169
+ if self.dropout is not None and self.training:
170
+ cx = torch.nn.functional.dropout(cx, p=self.dropout)
171
+
172
+ cx = self.up(cx) * self.multiplier
173
+
174
+ x = super().forward(x + cx) # ここで元のモジュールを呼び出す / call the original module here
175
+ return x
176
+
177
+
178
+ class SdxlUNet2DConditionModelControlNetLLLite(sdxl_original_unet.SdxlUNet2DConditionModel):
179
+ UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel"]
180
+ UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"]
181
+ LLLITE_PREFIX = "lllite_unet"
182
+
183
+ def __init__(self, **kwargs):
184
+ super().__init__(**kwargs)
185
+
186
+ def apply_lllite(
187
+ self,
188
+ cond_emb_dim: int = 16,
189
+ mlp_dim: int = 16,
190
+ dropout: Optional[float] = None,
191
+ varbose: Optional[bool] = False,
192
+ multiplier: Optional[float] = 1.0,
193
+ ) -> None:
194
+ def apply_to_modules(
195
+ root_module: torch.nn.Module,
196
+ target_replace_modules: List[torch.nn.Module],
197
+ ) -> List[torch.nn.Module]:
198
+ prefix = "lllite_unet"
199
+
200
+ modules = []
201
+ for name, module in root_module.named_modules():
202
+ if module.__class__.__name__ in target_replace_modules:
203
+ for child_name, child_module in module.named_modules():
204
+ is_linear = child_module.__class__.__name__ == "LLLiteLinear"
205
+ is_conv2d = child_module.__class__.__name__ == "LLLiteConv2d"
206
+
207
+ if is_linear or (is_conv2d and not SKIP_CONV2D):
208
+ # block indexからdepthを計算: depthはconditioningのサイズやチャネルを計算するのに使う
209
+ # block index to depth: depth is using to calculate conditioning size and channels
210
+ block_name, index1, index2 = (name + "." + child_name).split(".")[:3]
211
+ index1 = int(index1)
212
+ if block_name == "input_blocks":
213
+ if SKIP_INPUT_BLOCKS:
214
+ continue
215
+ depth = 1 if index1 <= 2 else (2 if index1 <= 5 else 3)
216
+ elif block_name == "middle_block":
217
+ depth = 3
218
+ elif block_name == "output_blocks":
219
+ if SKIP_OUTPUT_BLOCKS:
220
+ continue
221
+ depth = 3 if index1 <= 2 else (2 if index1 <= 5 else 1)
222
+ if int(index2) >= 2:
223
+ depth -= 1
224
+ else:
225
+ raise NotImplementedError()
226
+
227
+ lllite_name = prefix + "." + name + "." + child_name
228
+ lllite_name = lllite_name.replace(".", "_")
229
+
230
+ if TRANSFORMER_MAX_BLOCK_INDEX is not None:
231
+ p = lllite_name.find("transformer_blocks")
232
+ if p >= 0:
233
+ tf_index = int(lllite_name[p:].split("_")[2])
234
+ if tf_index > TRANSFORMER_MAX_BLOCK_INDEX:
235
+ continue
236
+
237
+ # time embは適用外とする
238
+ # attn2のconditioning (CLIPからの入力) はshapeが違うので適用できない
239
+ # time emb is not applied
240
+ # attn2 conditioning (input from CLIP) cannot be applied because the shape is different
241
+ if "emb_layers" in lllite_name or (
242
+ "attn2" in lllite_name and ("to_k" in lllite_name or "to_v" in lllite_name)
243
+ ):
244
+ continue
245
+
246
+ if ATTN1_2_ONLY:
247
+ if not ("attn1" in lllite_name or "attn2" in lllite_name):
248
+ continue
249
+ if ATTN_QKV_ONLY:
250
+ if "to_out" in lllite_name:
251
+ continue
252
+
253
+ if ATTN1_ETC_ONLY:
254
+ if "proj_out" in lllite_name:
255
+ pass
256
+ elif "attn1" in lllite_name and (
257
+ "to_k" in lllite_name or "to_v" in lllite_name or "to_out" in lllite_name
258
+ ):
259
+ pass
260
+ elif "ff_net_2" in lllite_name:
261
+ pass
262
+ else:
263
+ continue
264
+
265
+ child_module.set_lllite(depth, cond_emb_dim, lllite_name, mlp_dim, dropout, multiplier)
266
+ modules.append(child_module)
267
+
268
+ return modules
269
+
270
+ target_modules = SdxlUNet2DConditionModelControlNetLLLite.UNET_TARGET_REPLACE_MODULE
271
+ if not TRANSFORMER_ONLY:
272
+ target_modules = target_modules + SdxlUNet2DConditionModelControlNetLLLite.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3
273
+
274
+ # create module instances
275
+ self.lllite_modules = apply_to_modules(self, target_modules)
276
+ logger.info(f"enable ControlNet LLLite for U-Net: {len(self.lllite_modules)} modules.")
277
+
278
+ # def prepare_optimizer_params(self):
279
+ def prepare_params(self):
280
+ train_params = []
281
+ non_train_params = []
282
+ for name, p in self.named_parameters():
283
+ if "lllite" in name:
284
+ train_params.append(p)
285
+ else:
286
+ non_train_params.append(p)
287
+ logger.info(f"count of trainable parameters: {len(train_params)}")
288
+ logger.info(f"count of non-trainable parameters: {len(non_train_params)}")
289
+
290
+ for p in non_train_params:
291
+ p.requires_grad_(False)
292
+
293
+ # without this, an error occurs in the optimizer
294
+ # RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn
295
+ non_train_params[0].requires_grad_(True)
296
+
297
+ for p in train_params:
298
+ p.requires_grad_(True)
299
+
300
+ return train_params
301
+
302
+ # def prepare_grad_etc(self):
303
+ # self.requires_grad_(True)
304
+
305
+ # def on_epoch_start(self):
306
+ # self.train()
307
+
308
+ def get_trainable_params(self):
309
+ return [p[1] for p in self.named_parameters() if "lllite" in p[0]]
310
+
311
+ def save_lllite_weights(self, file, dtype, metadata):
312
+ if metadata is not None and len(metadata) == 0:
313
+ metadata = None
314
+
315
+ org_state_dict = self.state_dict()
316
+
317
+ # copy LLLite keys from org_state_dict to state_dict with key conversion
318
+ state_dict = {}
319
+ for key in org_state_dict.keys():
320
+ # split with ".lllite"
321
+ pos = key.find(".lllite")
322
+ if pos < 0:
323
+ continue
324
+ lllite_key = SdxlUNet2DConditionModelControlNetLLLite.LLLITE_PREFIX + "." + key[:pos]
325
+ lllite_key = lllite_key.replace(".", "_") + key[pos:]
326
+ lllite_key = lllite_key.replace(".lllite_", ".")
327
+ state_dict[lllite_key] = org_state_dict[key]
328
+
329
+ if dtype is not None:
330
+ for key in list(state_dict.keys()):
331
+ v = state_dict[key]
332
+ v = v.detach().clone().to("cpu").to(dtype)
333
+ state_dict[key] = v
334
+
335
+ if os.path.splitext(file)[1] == ".safetensors":
336
+ from safetensors.torch import save_file
337
+
338
+ save_file(state_dict, file, metadata)
339
+ else:
340
+ torch.save(state_dict, file)
341
+
342
+ def load_lllite_weights(self, file, non_lllite_unet_sd=None):
343
+ r"""
344
+ LLLiteの重みを読み込まない(initされた値を使う)場合はfileにNoneを指定する。
345
+ この場合、non_lllite_unet_sdにはU-Netのstate_dictを指定する。
346
+
347
+ If you do not want to load LLLite weights (use initialized values), specify None for file.
348
+ In this case, specify the state_dict of U-Net for non_lllite_unet_sd.
349
+ """
350
+ if not file:
351
+ state_dict = self.state_dict()
352
+ for key in non_lllite_unet_sd:
353
+ if key in state_dict:
354
+ state_dict[key] = non_lllite_unet_sd[key]
355
+ info = self.load_state_dict(state_dict, False)
356
+ return info
357
+
358
+ if os.path.splitext(file)[1] == ".safetensors":
359
+ from safetensors.torch import load_file
360
+
361
+ weights_sd = load_file(file)
362
+ else:
363
+ weights_sd = torch.load(file, map_location="cpu")
364
+
365
+ # module_name = module_name.replace("_block", "@blocks")
366
+ # module_name = module_name.replace("_layer", "@layer")
367
+ # module_name = module_name.replace("to_", "to@")
368
+ # module_name = module_name.replace("time_embed", "time@embed")
369
+ # module_name = module_name.replace("label_emb", "label@emb")
370
+ # module_name = module_name.replace("skip_connection", "skip@connection")
371
+ # module_name = module_name.replace("proj_in", "proj@in")
372
+ # module_name = module_name.replace("proj_out", "proj@out")
373
+ pattern = re.compile(r"(_block|_layer|to_|time_embed|label_emb|skip_connection|proj_in|proj_out)")
374
+
375
+ # convert to lllite with U-Net state dict
376
+ state_dict = non_lllite_unet_sd.copy() if non_lllite_unet_sd is not None else {}
377
+ for key in weights_sd.keys():
378
+ # split with "."
379
+ pos = key.find(".")
380
+ if pos < 0:
381
+ continue
382
+
383
+ module_name = key[:pos]
384
+ weight_name = key[pos + 1 :] # exclude "."
385
+ module_name = module_name.replace(SdxlUNet2DConditionModelControlNetLLLite.LLLITE_PREFIX + "_", "")
386
+
387
+ # これはうまくいかない。逆変換を考えなかった設計が悪い / this does not work well. bad design because I didn't think about inverse conversion
388
+ # module_name = module_name.replace("_", ".")
389
+
390
+ # ださいけどSDXLのU-Netの "_" を "@" に変換する / ugly but convert "_" of SDXL U-Net to "@"
391
+ matches = pattern.findall(module_name)
392
+ if matches is not None:
393
+ for m in matches:
394
+ logger.info(f"{module_name} {m}")
395
+ module_name = module_name.replace(m, m.replace("_", "@"))
396
+ module_name = module_name.replace("_", ".")
397
+ module_name = module_name.replace("@", "_")
398
+
399
+ lllite_key = module_name + ".lllite_" + weight_name
400
+
401
+ state_dict[lllite_key] = weights_sd[key]
402
+
403
+ info = self.load_state_dict(state_dict, False)
404
+ return info
405
+
406
+ def forward(self, x, timesteps=None, context=None, y=None, cond_image=None, **kwargs):
407
+ for m in self.lllite_modules:
408
+ m.set_cond_image(cond_image)
409
+ return super().forward(x, timesteps, context, y, **kwargs)
410
+
411
+
412
+ def replace_unet_linear_and_conv2d():
413
+ logger.info("replace torch.nn.Linear and torch.nn.Conv2d to LLLiteLinear and LLLiteConv2d in U-Net")
414
+ sdxl_original_unet.torch.nn.Linear = LLLiteLinear
415
+ sdxl_original_unet.torch.nn.Conv2d = LLLiteConv2d
416
+
417
+
418
+ if __name__ == "__main__":
419
+ # デバッグ用 / for debug
420
+
421
+ # sdxl_original_unet.USE_REENTRANT = False
422
+ replace_unet_linear_and_conv2d()
423
+
424
+ # test shape etc
425
+ logger.info("create unet")
426
+ unet = SdxlUNet2DConditionModelControlNetLLLite()
427
+
428
+ logger.info("enable ControlNet-LLLite")
429
+ unet.apply_lllite(32, 64, None, False, 1.0)
430
+ unet.to("cuda") # .to(torch.float16)
431
+
432
+ # from safetensors.torch import load_file
433
+
434
+ # model_sd = load_file(r"E:\Work\SD\Models\sdxl\sd_xl_base_1.0_0.9vae.safetensors")
435
+ # unet_sd = {}
436
+
437
+ # # copy U-Net keys from unet_state_dict to state_dict
438
+ # prefix = "model.diffusion_model."
439
+ # for key in model_sd.keys():
440
+ # if key.startswith(prefix):
441
+ # converted_key = key[len(prefix) :]
442
+ # unet_sd[converted_key] = model_sd[key]
443
+
444
+ # info = unet.load_lllite_weights("r:/lllite_from_unet.safetensors", unet_sd)
445
+ # logger.info(info)
446
+
447
+ # logger.info(unet)
448
+
449
+ # logger.info number of parameters
450
+ params = unet.prepare_params()
451
+ logger.info(f"number of parameters {sum(p.numel() for p in params)}")
452
+ # logger.info("type any key to continue")
453
+ # input()
454
+
455
+ unet.set_use_memory_efficient_attention(True, False)
456
+ unet.set_gradient_checkpointing(True)
457
+ unet.train() # for gradient checkpointing
458
+
459
+ # # visualize
460
+ # import torchviz
461
+ # logger.info("run visualize")
462
+ # controlnet.set_control(conditioning_image)
463
+ # output = unet(x, t, ctx, y)
464
+ # logger.info("make_dot")
465
+ # image = torchviz.make_dot(output, params=dict(controlnet.named_parameters()))
466
+ # logger.info("render")
467
+ # image.format = "svg" # "png"
468
+ # image.render("NeuralNet") # すごく時間がかかるので注意 / be careful because it takes a long time
469
+ # input()
470
+
471
+ import bitsandbytes
472
+
473
+ optimizer = bitsandbytes.adam.Adam8bit(params, 1e-3)
474
+
475
+ scaler = torch.cuda.amp.GradScaler(enabled=True)
476
+
477
+ logger.info("start training")
478
+ steps = 10
479
+ batch_size = 1
480
+
481
+ sample_param = [p for p in unet.named_parameters() if ".lllite_up." in p[0]][0]
482
+ for step in range(steps):
483
+ logger.info(f"step {step}")
484
+
485
+ conditioning_image = torch.rand(batch_size, 3, 1024, 1024).cuda() * 2.0 - 1.0
486
+ x = torch.randn(batch_size, 4, 128, 128).cuda()
487
+ t = torch.randint(low=0, high=10, size=(batch_size,)).cuda()
488
+ ctx = torch.randn(batch_size, 77, 2048).cuda()
489
+ y = torch.randn(batch_size, sdxl_original_unet.ADM_IN_CHANNELS).cuda()
490
+
491
+ with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
492
+ output = unet(x, t, ctx, y, conditioning_image)
493
+ target = torch.randn_like(output)
494
+ loss = torch.nn.functional.mse_loss(output, target)
495
+
496
+ scaler.scale(loss).backward()
497
+ scaler.step(optimizer)
498
+ scaler.update()
499
+ optimizer.zero_grad(set_to_none=True)
500
+ logger.info(sample_param)
501
+
502
+ # from safetensors.torch import save_file
503
+
504
+ # logger.info("save weights")
505
+ # unet.save_lllite_weights("r:/lllite_from_unet.safetensors", torch.float16, None)