Hugging Face's logo

anyMODE
/
python-script-dump

Model card Files
xet
Community
python-script-dump
File size: 13,162 Bytes 
54c3b42
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
#!/usr/bin/env python

import argparse
from safetensors import safe_open
from safetensors.torch import save_file
from tqdm import tqdm

# This is the set of layers that will be removed from the file.
# Using a set provides fast lookups.
KEYS_TO_REMOVE = {
    "model.diffusion_model.double_blocks.0.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.0.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.0.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.0.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.1.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.1.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.1.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.1.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.10.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.10.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.10.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.10.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.11.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.11.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.11.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.11.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.12.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.12.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.12.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.12.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.13.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.13.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.13.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.13.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.14.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.14.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.14.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.14.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.15.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.15.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.15.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.15.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.16.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.16.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.16.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.16.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.17.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.17.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.17.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.17.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.18.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.18.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.18.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.18.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.2.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.2.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.2.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.2.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.3.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.3.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.3.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.3.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.4.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.4.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.4.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.4.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.5.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.5.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.5.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.5.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.6.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.6.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.6.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.6.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.7.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.7.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.7.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.7.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.8.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.8.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.8.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.8.txt_mod.lin.weight",
    "model.diffusion_model.double_blocks.9.img_mod.lin.bias",
    "model.diffusion_model.double_blocks.9.img_mod.lin.weight",
    "model.diffusion_model.double_blocks.9.txt_mod.lin.bias",
    "model.diffusion_model.double_blocks.9.txt_mod.lin.weight",
    "model.diffusion_model.final_layer.adaLN_modulation.1.bias",
    "model.diffusion_model.final_layer.adaLN_modulation.1.weight",
    # "model.diffusion_model.guidance_in.in_layer.bias",
    # "model.diffusion_model.guidance_in.in_layer.weight",
    # "model.diffusion_model.guidance_in.out_layer.bias",
    # "model.diffusion_model.guidance_in.out_layer.weight",
    "model.diffusion_model.single_blocks.0.modulation.lin.bias",
    "model.diffusion_model.single_blocks.0.modulation.lin.weight",
    "model.diffusion_model.single_blocks.1.modulation.lin.bias",
    "model.diffusion_model.single_blocks.1.modulation.lin.weight",
    "model.diffusion_model.single_blocks.10.modulation.lin.bias",
    "model.diffusion_model.single_blocks.10.modulation.lin.weight",
    "model.diffusion_model.single_blocks.11.modulation.lin.bias",
    "model.diffusion_model.single_blocks.11.modulation.lin.weight",
    "model.diffusion_model.single_blocks.12.modulation.lin.bias",
    "model.diffusion_model.single_blocks.12.modulation.lin.weight",
    "model.diffusion_model.single_blocks.13.modulation.lin.bias",
    "model.diffusion_model.single_blocks.13.modulation.lin.weight",
    "model.diffusion_model.single_blocks.14.modulation.lin.bias",
    "model.diffusion_model.single_blocks.14.modulation.lin.weight",
    "model.diffusion_model.single_blocks.15.modulation.lin.bias",
    "model.diffusion_model.single_blocks.15.modulation.lin.weight",
    "model.diffusion_model.single_blocks.16.modulation.lin.bias",
    "model.diffusion_model.single_blocks.16.modulation.lin.weight",
    "model.diffusion_model.single_blocks.17.modulation.lin.bias",
    "model.diffusion_model.single_blocks.17.modulation.lin.weight",
    "model.diffusion_model.single_blocks.18.modulation.lin.bias",
    "model.diffusion_model.single_blocks.18.modulation.lin.weight",
    "model.diffusion_model.single_blocks.19.modulation.lin.bias",
    "model.diffusion_model.single_blocks.19.modulation.lin.weight",
    "model.diffusion_model.single_blocks.2.modulation.lin.bias",
    "model.diffusion_model.single_blocks.2.modulation.lin.weight",
    "model.diffusion_model.single_blocks.20.modulation.lin.bias",
    "model.diffusion_model.single_blocks.20.modulation.lin.weight",
    "model.diffusion_model.single_blocks.21.modulation.lin.bias",
    "model.diffusion_model.single_blocks.21.modulation.lin.weight",
    "model.diffusion_model.single_blocks.22.modulation.lin.bias",
    "model.diffusion_model.single_blocks.22.modulation.lin.weight",
    "model.diffusion_model.single_blocks.23.modulation.lin.bias",
    "model.diffusion_model.single_blocks.23.modulation.lin.weight",
    "model.diffusion_model.single_blocks.24.modulation.lin.bias",
    "model.diffusion_model.single_blocks.24.modulation.lin.weight",
    "model.diffusion_model.single_blocks.25.modulation.lin.bias",
    "model.diffusion_model.single_blocks.25.modulation.lin.weight",
    "model.diffusion_model.single_blocks.26.modulation.lin.bias",
    "model.diffusion_model.single_blocks.26.modulation.lin.weight",
    "model.diffusion_model.single_blocks.27.modulation.lin.bias",
    "model.diffusion_model.single_blocks.27.modulation.lin.weight",
    "model.diffusion_model.single_blocks.28.modulation.lin.bias",
    "model.diffusion_model.single_blocks.28.modulation.lin.weight",
    "model.diffusion_model.single_blocks.29.modulation.lin.bias",
    "model.diffusion_model.single_blocks.29.modulation.lin.weight",
    "model.diffusion_model.single_blocks.3.modulation.lin.bias",
    "model.diffusion_model.single_blocks.3.modulation.lin.weight",
    "model.diffusion_model.single_blocks.30.modulation.lin.bias",
    "model.diffusion_model.single_blocks.30.modulation.lin.weight",
    "model.diffusion_model.single_blocks.31.modulation.lin.bias",
    "model.diffusion_model.single_blocks.31.modulation.lin.weight",
    "model.diffusion_model.single_blocks.32.modulation.lin.bias",
    "model.diffusion_model.single_blocks.32.modulation.lin.weight",
    "model.diffusion_model.single_blocks.33.modulation.lin.bias",
    "model.diffusion_model.single_blocks.33.modulation.lin.weight",
    "model.diffusion_model.single_blocks.34.modulation.lin.bias",
    "model.diffusion_model.single_blocks.34.modulation.lin.weight",
    "model.diffusion_model.single_blocks.35.modulation.lin.bias",
    "model.diffusion_model.single_blocks.35.modulation.lin.weight",
    "model.diffusion_model.single_blocks.36.modulation.lin.bias",
    "model.diffusion_model.single_blocks.36.modulation.lin.weight",
    "model.diffusion_model.single_blocks.37.modulation.lin.bias",
    "model.diffusion_model.single_blocks.37.modulation.lin.weight",
    "model.diffusion_model.single_blocks.4.modulation.lin.bias",
    "model.diffusion_model.single_blocks.4.modulation.lin.weight",
    "model.diffusion_model.single_blocks.5.modulation.lin.bias",
    "model.diffusion_model.single_blocks.5.modulation.lin.weight",
    "model.diffusion_model.single_blocks.6.modulation.lin.bias",
    "model.diffusion_model.single_blocks.6.modulation.lin.weight",
    "model.diffusion_model.single_blocks.7.modulation.lin.bias",
    "model.diffusion_model.single_blocks.7.modulation.lin.weight",
    "model.diffusion_model.single_blocks.8.modulation.lin.bias",
    "model.diffusion_model.single_blocks.8.modulation.lin.weight",
    "model.diffusion_model.single_blocks.9.modulation.lin.bias",
    "model.diffusion_model.single_blocks.9.modulation.lin.weight",
    # "model.diffusion_model.time_in.in_layer.bias",
    # "model.diffusion_model.time_in.in_layer.weight",
    # "model.diffusion_model.time_in.out_layer.bias",
    # "model.diffusion_model.time_in.out_layer.weight",
    # "model.diffusion_model.vector_in.in_layer.bias",
    # "model.diffusion_model.vector_in.in_layer.weight",
    # "model.diffusion_model.vector_in.out_layer.bias",
    # "model.diffusion_model.vector_in.out_layer.weight",
}

def prune_safetensors_file(input_path: str, output_path: str, keys_to_remove: set):
    """
    Reads a safetensors file, removes specified layers, and saves to a new file.
    """
    tensors_to_keep = {}
    original_count = 0

    print(f"▶️  Reading from: {input_path}")
    try:
        # Open the source file for reading
        with safe_open(input_path, framework="pt", device="cpu") as f:
            all_keys = f.keys()
            original_count = len(all_keys)

            # Iterate through all tensors, keeping only the ones not in the removal list
            for key in tqdm(all_keys, desc="Filtering Tensors"):
                if 'modulation' not in key :
                    tensors_to_keep[key] = f.get_tensor(key)

        removed_count = original_count - len(tensors_to_keep)
        
        print("\n--- Summary ---")
        print(f"Original tensor count: {original_count}")
        print(f"Tensors removed: {removed_count}")
        print(f"New tensor count: {len(tensors_to_keep)}")

        if removed_count == 0 and len(keys_to_remove) > 0:
            print("\n⚠️  Warning: None of the specified layers to remove were found in the input file.")

        # Save the filtered tensors to the new file
        print(f"\n▶️  Saving {len(tensors_to_keep)} tensors to: {output_path}")
        save_file(tensors_to_keep, output_path)

        print("\n✅ Pruning complete!")

    except FileNotFoundError:
        print(f"\n❌ Error: The file '{input_path}' was not found.")
    except Exception as e:
        print(f"\n❌ An error occurred: {e}")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(
        description="Removes a predefined list of layers from a .safetensors file and saves it as a new file.",
        formatter_class=argparse.RawTextHelpFormatter
    )

    parser.add_argument(
        "input_file",
        type=str,
        help="Path to the source .safetensors file."
    )
    parser.add_argument(
        "output_file",
        type=str,
        help="Path to save the new, pruned .safetensors file."
    )

    args = parser.parse_args()
    
    prune_safetensors_file(args.input_file, args.output_file, KEYS_TO_REMOVE)