Upload lora-scripts/sd-scripts/tools/merge_models.py with huggingface_hub
Browse files
lora-scripts/sd-scripts/tools/merge_models.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from safetensors import safe_open
|
6 |
+
from safetensors.torch import load_file, save_file
|
7 |
+
from tqdm import tqdm
|
8 |
+
from library.utils import setup_logging
|
9 |
+
setup_logging()
|
10 |
+
import logging
|
11 |
+
logger = logging.getLogger(__name__)
|
12 |
+
|
13 |
+
def is_unet_key(key):
|
14 |
+
# VAE or TextEncoder, the last one is for SDXL
|
15 |
+
return not ("first_stage_model" in key or "cond_stage_model" in key or "conditioner." in key)
|
16 |
+
|
17 |
+
|
18 |
+
TEXT_ENCODER_KEY_REPLACEMENTS = [
|
19 |
+
("cond_stage_model.transformer.embeddings.", "cond_stage_model.transformer.text_model.embeddings."),
|
20 |
+
("cond_stage_model.transformer.encoder.", "cond_stage_model.transformer.text_model.encoder."),
|
21 |
+
("cond_stage_model.transformer.final_layer_norm.", "cond_stage_model.transformer.text_model.final_layer_norm."),
|
22 |
+
]
|
23 |
+
|
24 |
+
|
25 |
+
# support for models with different text encoder keys
|
26 |
+
def replace_text_encoder_key(key):
|
27 |
+
for rep_from, rep_to in TEXT_ENCODER_KEY_REPLACEMENTS:
|
28 |
+
if key.startswith(rep_from):
|
29 |
+
return True, rep_to + key[len(rep_from) :]
|
30 |
+
return False, key
|
31 |
+
|
32 |
+
|
33 |
+
def merge(args):
|
34 |
+
if args.precision == "fp16":
|
35 |
+
dtype = torch.float16
|
36 |
+
elif args.precision == "bf16":
|
37 |
+
dtype = torch.bfloat16
|
38 |
+
else:
|
39 |
+
dtype = torch.float
|
40 |
+
|
41 |
+
if args.saving_precision == "fp16":
|
42 |
+
save_dtype = torch.float16
|
43 |
+
elif args.saving_precision == "bf16":
|
44 |
+
save_dtype = torch.bfloat16
|
45 |
+
else:
|
46 |
+
save_dtype = torch.float
|
47 |
+
|
48 |
+
# check if all models are safetensors
|
49 |
+
for model in args.models:
|
50 |
+
if not model.endswith("safetensors"):
|
51 |
+
logger.info(f"Model {model} is not a safetensors model")
|
52 |
+
exit()
|
53 |
+
if not os.path.isfile(model):
|
54 |
+
logger.info(f"Model {model} does not exist")
|
55 |
+
exit()
|
56 |
+
|
57 |
+
assert args.ratios is None or len(args.models) == len(args.ratios), "ratios must be the same length as models"
|
58 |
+
|
59 |
+
# load and merge
|
60 |
+
ratio = 1.0 / len(args.models) # default
|
61 |
+
supplementary_key_ratios = {} # [key] = ratio, for keys not in all models, add later
|
62 |
+
|
63 |
+
merged_sd = None
|
64 |
+
first_model_keys = set() # check missing keys in other models
|
65 |
+
for i, model in enumerate(args.models):
|
66 |
+
if args.ratios is not None:
|
67 |
+
ratio = args.ratios[i]
|
68 |
+
|
69 |
+
if merged_sd is None:
|
70 |
+
# load first model
|
71 |
+
logger.info(f"Loading model {model}, ratio = {ratio}...")
|
72 |
+
merged_sd = {}
|
73 |
+
with safe_open(model, framework="pt", device=args.device) as f:
|
74 |
+
for key in tqdm(f.keys()):
|
75 |
+
value = f.get_tensor(key)
|
76 |
+
_, key = replace_text_encoder_key(key)
|
77 |
+
|
78 |
+
first_model_keys.add(key)
|
79 |
+
|
80 |
+
if not is_unet_key(key) and args.unet_only:
|
81 |
+
supplementary_key_ratios[key] = 1.0 # use first model's value for VAE or TextEncoder
|
82 |
+
continue
|
83 |
+
|
84 |
+
value = ratio * value.to(dtype) # first model's value * ratio
|
85 |
+
merged_sd[key] = value
|
86 |
+
|
87 |
+
logger.info(f"Model has {len(merged_sd)} keys " + ("(UNet only)" if args.unet_only else ""))
|
88 |
+
continue
|
89 |
+
|
90 |
+
# load other models
|
91 |
+
logger.info(f"Loading model {model}, ratio = {ratio}...")
|
92 |
+
|
93 |
+
with safe_open(model, framework="pt", device=args.device) as f:
|
94 |
+
model_keys = f.keys()
|
95 |
+
for key in tqdm(model_keys):
|
96 |
+
_, new_key = replace_text_encoder_key(key)
|
97 |
+
if new_key not in merged_sd:
|
98 |
+
if args.show_skipped and new_key not in first_model_keys:
|
99 |
+
logger.info(f"Skip: {new_key}")
|
100 |
+
continue
|
101 |
+
|
102 |
+
value = f.get_tensor(key)
|
103 |
+
merged_sd[new_key] = merged_sd[new_key] + ratio * value.to(dtype)
|
104 |
+
|
105 |
+
# enumerate keys not in this model
|
106 |
+
model_keys = set(model_keys)
|
107 |
+
for key in merged_sd.keys():
|
108 |
+
if key in model_keys:
|
109 |
+
continue
|
110 |
+
logger.warning(f"Key {key} not in model {model}, use first model's value")
|
111 |
+
if key in supplementary_key_ratios:
|
112 |
+
supplementary_key_ratios[key] += ratio
|
113 |
+
else:
|
114 |
+
supplementary_key_ratios[key] = ratio
|
115 |
+
|
116 |
+
# add supplementary keys' value (including VAE and TextEncoder)
|
117 |
+
if len(supplementary_key_ratios) > 0:
|
118 |
+
logger.info("add first model's value")
|
119 |
+
with safe_open(args.models[0], framework="pt", device=args.device) as f:
|
120 |
+
for key in tqdm(f.keys()):
|
121 |
+
_, new_key = replace_text_encoder_key(key)
|
122 |
+
if new_key not in supplementary_key_ratios:
|
123 |
+
continue
|
124 |
+
|
125 |
+
if is_unet_key(new_key): # not VAE or TextEncoder
|
126 |
+
logger.warning(f"Key {new_key} not in all models, ratio = {supplementary_key_ratios[new_key]}")
|
127 |
+
|
128 |
+
value = f.get_tensor(key) # original key
|
129 |
+
|
130 |
+
if new_key not in merged_sd:
|
131 |
+
merged_sd[new_key] = supplementary_key_ratios[new_key] * value.to(dtype)
|
132 |
+
else:
|
133 |
+
merged_sd[new_key] = merged_sd[new_key] + supplementary_key_ratios[new_key] * value.to(dtype)
|
134 |
+
|
135 |
+
# save
|
136 |
+
output_file = args.output
|
137 |
+
if not output_file.endswith(".safetensors"):
|
138 |
+
output_file = output_file + ".safetensors"
|
139 |
+
|
140 |
+
logger.info(f"Saving to {output_file}...")
|
141 |
+
|
142 |
+
# convert to save_dtype
|
143 |
+
for k in merged_sd.keys():
|
144 |
+
merged_sd[k] = merged_sd[k].to(save_dtype)
|
145 |
+
|
146 |
+
save_file(merged_sd, output_file)
|
147 |
+
|
148 |
+
logger.info("Done!")
|
149 |
+
|
150 |
+
|
151 |
+
if __name__ == "__main__":
|
152 |
+
parser = argparse.ArgumentParser(description="Merge models")
|
153 |
+
parser.add_argument("--models", nargs="+", type=str, help="Models to merge")
|
154 |
+
parser.add_argument("--output", type=str, help="Output model")
|
155 |
+
parser.add_argument("--ratios", nargs="+", type=float, help="Ratios of models, default is equal, total = 1.0")
|
156 |
+
parser.add_argument("--unet_only", action="store_true", help="Only merge unet")
|
157 |
+
parser.add_argument("--device", type=str, default="cpu", help="Device to use, default is cpu")
|
158 |
+
parser.add_argument(
|
159 |
+
"--precision", type=str, default="float", choices=["float", "fp16", "bf16"], help="Calculation precision, default is float"
|
160 |
+
)
|
161 |
+
parser.add_argument(
|
162 |
+
"--saving_precision",
|
163 |
+
type=str,
|
164 |
+
default="float",
|
165 |
+
choices=["float", "fp16", "bf16"],
|
166 |
+
help="Saving precision, default is float",
|
167 |
+
)
|
168 |
+
parser.add_argument("--show_skipped", action="store_true", help="Show skipped keys (keys not in first model)")
|
169 |
+
|
170 |
+
args = parser.parse_args()
|
171 |
+
merge(args)
|