forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathstate_dict_utils.py
More file actions
324 lines (272 loc) 路 12.7 KB
/
Copy pathstate_dict_utils.py
File metadata and controls
324 lines (272 loc) 路 12.7 KB
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
# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
State dict utilities: utility methods for converting state dicts easily
"""
import enum
from .logging import get_logger
logger = get_logger(__name__)
class StateDictType(enum.Enum):
"""
The mode to use when converting state dicts.
"""
DIFFUSERS_OLD = "diffusers_old"
KOHYA_SS = "kohya_ss"
PEFT = "peft"
DIFFUSERS = "diffusers"
# We need to define a proper mapping for Unet since it uses different output keys than text encoder
# e.g. to_q_lora -> q_proj / to_q
UNET_TO_DIFFUSERS = {
".to_out_lora.up": ".to_out.0.lora_B",
".to_out_lora.down": ".to_out.0.lora_A",
".to_q_lora.down": ".to_q.lora_A",
".to_q_lora.up": ".to_q.lora_B",
".to_k_lora.down": ".to_k.lora_A",
".to_k_lora.up": ".to_k.lora_B",
".to_v_lora.down": ".to_v.lora_A",
".to_v_lora.up": ".to_v.lora_B",
".lora.up": ".lora_B",
".lora.down": ".lora_A",
}
DIFFUSERS_TO_PEFT = {
".q_proj.lora_linear_layer.up": ".q_proj.lora_B",
".q_proj.lora_linear_layer.down": ".q_proj.lora_A",
".k_proj.lora_linear_layer.up": ".k_proj.lora_B",
".k_proj.lora_linear_layer.down": ".k_proj.lora_A",
".v_proj.lora_linear_layer.up": ".v_proj.lora_B",
".v_proj.lora_linear_layer.down": ".v_proj.lora_A",
".out_proj.lora_linear_layer.up": ".out_proj.lora_B",
".out_proj.lora_linear_layer.down": ".out_proj.lora_A",
".lora_linear_layer.up": ".lora_B",
".lora_linear_layer.down": ".lora_A",
}
DIFFUSERS_OLD_TO_PEFT = {
".to_q_lora.up": ".q_proj.lora_B",
".to_q_lora.down": ".q_proj.lora_A",
".to_k_lora.up": ".k_proj.lora_B",
".to_k_lora.down": ".k_proj.lora_A",
".to_v_lora.up": ".v_proj.lora_B",
".to_v_lora.down": ".v_proj.lora_A",
".to_out_lora.up": ".out_proj.lora_B",
".to_out_lora.down": ".out_proj.lora_A",
".lora_linear_layer.up": ".lora_B",
".lora_linear_layer.down": ".lora_A",
}
PEFT_TO_DIFFUSERS = {
".q_proj.lora_B": ".q_proj.lora_linear_layer.up",
".q_proj.lora_A": ".q_proj.lora_linear_layer.down",
".k_proj.lora_B": ".k_proj.lora_linear_layer.up",
".k_proj.lora_A": ".k_proj.lora_linear_layer.down",
".v_proj.lora_B": ".v_proj.lora_linear_layer.up",
".v_proj.lora_A": ".v_proj.lora_linear_layer.down",
".out_proj.lora_B": ".out_proj.lora_linear_layer.up",
".out_proj.lora_A": ".out_proj.lora_linear_layer.down",
"to_k.lora_A": "to_k.lora.down",
"to_k.lora_B": "to_k.lora.up",
"to_q.lora_A": "to_q.lora.down",
"to_q.lora_B": "to_q.lora.up",
"to_v.lora_A": "to_v.lora.down",
"to_v.lora_B": "to_v.lora.up",
"to_out.0.lora_A": "to_out.0.lora.down",
"to_out.0.lora_B": "to_out.0.lora.up",
}
DIFFUSERS_OLD_TO_DIFFUSERS = {
".to_q_lora.up": ".q_proj.lora_linear_layer.up",
".to_q_lora.down": ".q_proj.lora_linear_layer.down",
".to_k_lora.up": ".k_proj.lora_linear_layer.up",
".to_k_lora.down": ".k_proj.lora_linear_layer.down",
".to_v_lora.up": ".v_proj.lora_linear_layer.up",
".to_v_lora.down": ".v_proj.lora_linear_layer.down",
".to_out_lora.up": ".out_proj.lora_linear_layer.up",
".to_out_lora.down": ".out_proj.lora_linear_layer.down",
}
PEFT_TO_KOHYA_SS = {
"lora_A": "lora_down",
"lora_B": "lora_up",
# This is not a comprehensive dict as kohya format requires replacing `.` with `_` in keys,
# adding prefixes and adding alpha values
# Check `convert_state_dict_to_kohya` for more
}
PEFT_STATE_DICT_MAPPINGS = {
StateDictType.DIFFUSERS_OLD: DIFFUSERS_OLD_TO_PEFT,
StateDictType.DIFFUSERS: DIFFUSERS_TO_PEFT,
}
DIFFUSERS_STATE_DICT_MAPPINGS = {
StateDictType.DIFFUSERS_OLD: DIFFUSERS_OLD_TO_DIFFUSERS,
StateDictType.PEFT: PEFT_TO_DIFFUSERS,
}
KOHYA_STATE_DICT_MAPPINGS = {StateDictType.PEFT: PEFT_TO_KOHYA_SS}
KEYS_TO_ALWAYS_REPLACE = {
".processor.": ".",
}
def convert_state_dict(state_dict, mapping):
r"""
Simply iterates over the state dict and replaces the patterns in `mapping` with the corresponding values.
Args:
state_dict (`dict[str, torch.Tensor]`):
The state dict to convert.
mapping (`dict[str, str]`):
The mapping to use for conversion, the mapping should be a dictionary with the following structure:
- key: the pattern to replace
- value: the pattern to replace with
Returns:
converted_state_dict (`dict`)
The converted state dict.
"""
converted_state_dict = {}
for k, v in state_dict.items():
# First, filter out the keys that we always want to replace
for pattern in KEYS_TO_ALWAYS_REPLACE.keys():
if pattern in k:
new_pattern = KEYS_TO_ALWAYS_REPLACE[pattern]
k = k.replace(pattern, new_pattern)
for pattern in mapping.keys():
if pattern in k:
new_pattern = mapping[pattern]
k = k.replace(pattern, new_pattern)
break
converted_state_dict[k] = v
return converted_state_dict
def convert_state_dict_to_peft(state_dict, original_type=None, **kwargs):
r"""
Converts a state dict to the PEFT format The state dict can be from previous diffusers format (`OLD_DIFFUSERS`), or
new diffusers format (`DIFFUSERS`). The method only supports the conversion from diffusers old/new to PEFT for now.
Args:
state_dict (`dict[str, torch.Tensor]`):
The state dict to convert.
original_type (`StateDictType`, *optional*):
The original type of the state dict, if not provided, the method will try to infer it automatically.
"""
if original_type is None:
# Old diffusers to PEFT
if any("to_out_lora" in k for k in state_dict.keys()):
original_type = StateDictType.DIFFUSERS_OLD
elif any("lora_linear_layer" in k for k in state_dict.keys()):
original_type = StateDictType.DIFFUSERS
else:
raise ValueError("Could not automatically infer state dict type")
if original_type not in PEFT_STATE_DICT_MAPPINGS.keys():
raise ValueError(f"Original type {original_type} is not supported")
mapping = PEFT_STATE_DICT_MAPPINGS[original_type]
return convert_state_dict(state_dict, mapping)
def convert_state_dict_to_diffusers(state_dict, original_type=None, **kwargs):
r"""
Converts a state dict to new diffusers format. The state dict can be from previous diffusers format
(`OLD_DIFFUSERS`), or PEFT format (`PEFT`) or new diffusers format (`DIFFUSERS`). In the last case the method will
return the state dict as is.
The method only supports the conversion from diffusers old, PEFT to diffusers new for now.
Args:
state_dict (`dict[str, torch.Tensor]`):
The state dict to convert.
original_type (`StateDictType`, *optional*):
The original type of the state dict, if not provided, the method will try to infer it automatically.
kwargs (`dict`, *args*):
Additional arguments to pass to the method.
- **adapter_name**: For example, in case of PEFT, some keys will be pre-pended
with the adapter name, therefore needs a special handling. By default PEFT also takes care of that in
`get_peft_model_state_dict` method:
https://github.com/huggingface/peft/blob/ba0477f2985b1ba311b83459d29895c809404e99/src/peft/utils/save_and_load.py#L92
but we add it here in case we don't want to rely on that method.
"""
peft_adapter_name = kwargs.pop("adapter_name", None)
if peft_adapter_name is not None:
peft_adapter_name = "." + peft_adapter_name
else:
peft_adapter_name = ""
if original_type is None:
# Old diffusers to PEFT
if any("to_out_lora" in k for k in state_dict.keys()):
original_type = StateDictType.DIFFUSERS_OLD
elif any(f".lora_A{peft_adapter_name}.weight" in k for k in state_dict.keys()):
original_type = StateDictType.PEFT
elif any("lora_linear_layer" in k for k in state_dict.keys()):
# nothing to do
return state_dict
else:
raise ValueError("Could not automatically infer state dict type")
if original_type not in DIFFUSERS_STATE_DICT_MAPPINGS.keys():
raise ValueError(f"Original type {original_type} is not supported")
mapping = DIFFUSERS_STATE_DICT_MAPPINGS[original_type]
return convert_state_dict(state_dict, mapping)
def convert_unet_state_dict_to_peft(state_dict):
r"""
Converts a state dict from UNet format to diffusers format - i.e. by removing some keys
"""
mapping = UNET_TO_DIFFUSERS
return convert_state_dict(state_dict, mapping)
def convert_all_state_dict_to_peft(state_dict):
r"""
Attempts to first `convert_state_dict_to_peft`, and if it doesn't detect `lora_linear_layer`
for a valid `DIFFUSERS` LoRA for example, attempts to exclusively convert the Unet `convert_unet_state_dict_to_peft`
"""
try:
peft_dict = convert_state_dict_to_peft(state_dict)
except Exception as e:
if str(e) == "Could not automatically infer state dict type":
peft_dict = convert_unet_state_dict_to_peft(state_dict)
else:
raise
if not any("lora_A" in key or "lora_B" in key for key in peft_dict.keys()):
raise ValueError("Your LoRA was not converted to PEFT")
return peft_dict
def convert_state_dict_to_kohya(state_dict, original_type=None, **kwargs):
r"""
Converts a `PEFT` state dict to `Kohya` format that can be used in AUTOMATIC1111, ComfyUI, SD.Next, InvokeAI, etc.
The method only supports the conversion from PEFT to Kohya for now.
Args:
state_dict (`dict[str, torch.Tensor]`):
The state dict to convert.
original_type (`StateDictType`, *optional*):
The original type of the state dict, if not provided, the method will try to infer it automatically.
kwargs (`dict`, *args*):
Additional arguments to pass to the method.
- **adapter_name**: For example, in case of PEFT, some keys will be pre-pended
with the adapter name, therefore needs a special handling. By default PEFT also takes care of that in
`get_peft_model_state_dict` method:
https://github.com/huggingface/peft/blob/ba0477f2985b1ba311b83459d29895c809404e99/src/peft/utils/save_and_load.py#L92
but we add it here in case we don't want to rely on that method.
"""
try:
import torch
except ImportError:
logger.error("Converting PEFT state dicts to Kohya requires torch to be installed.")
raise
peft_adapter_name = kwargs.pop("adapter_name", None)
if peft_adapter_name is not None:
peft_adapter_name = "." + peft_adapter_name
else:
peft_adapter_name = ""
if original_type is None:
if any(f".lora_A{peft_adapter_name}.weight" in k for k in state_dict.keys()):
original_type = StateDictType.PEFT
if original_type not in KOHYA_STATE_DICT_MAPPINGS.keys():
raise ValueError(f"Original type {original_type} is not supported")
# Use the convert_state_dict function with the appropriate mapping
kohya_ss_partial_state_dict = convert_state_dict(state_dict, KOHYA_STATE_DICT_MAPPINGS[StateDictType.PEFT])
kohya_ss_state_dict = {}
# Additional logic for replacing header, alpha parameters `.` with `_` in all keys
for kohya_key, weight in kohya_ss_partial_state_dict.items():
if "text_encoder_2." in kohya_key:
kohya_key = kohya_key.replace("text_encoder_2.", "lora_te2.")
elif "text_encoder." in kohya_key:
kohya_key = kohya_key.replace("text_encoder.", "lora_te1.")
elif "unet" in kohya_key:
kohya_key = kohya_key.replace("unet", "lora_unet")
kohya_key = kohya_key.replace(".", "_", kohya_key.count(".") - 2)
kohya_key = kohya_key.replace(peft_adapter_name, "") # Kohya doesn't take names
kohya_ss_state_dict[kohya_key] = weight
if "lora_down" in kohya_key:
alpha_key = f'{kohya_key.split(".")[0]}.alpha'
kohya_ss_state_dict[alpha_key] = torch.tensor(len(weight))
return kohya_ss_state_dict