Spaces:
Running
on
Zero
Running
on
Zero
Update merge.py
Browse files
merge.py
CHANGED
@@ -5,21 +5,43 @@ import shutil
|
|
5 |
import torch
|
6 |
import torch.nn.functional as F
|
7 |
from safetensors.torch import safe_open, save_file
|
|
|
8 |
|
9 |
-
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
delta = tensor2 - tensor1
|
12 |
-
# Generate the mask m^t from Bernoulli distribution
|
13 |
m = torch.from_numpy(np.random.binomial(1, p, delta.shape)).to(tensor1.dtype)
|
14 |
-
# Apply the mask to the delta to get δ̃^t
|
15 |
delta_tilde = m * delta
|
16 |
-
# Scale the masked delta by the dropout rate to get δ̂^t
|
17 |
delta_hat = delta_tilde / (1 - p)
|
18 |
return delta_hat
|
19 |
|
20 |
-
def merge_safetensors(file_path1, file_path2, p, lambda_val):
|
21 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
|
|
|
|
|
|
|
|
|
23 |
with safe_open(file_path1, framework="pt", device="cpu") as f1, safe_open(file_path2, framework="pt", device="cpu") as f2:
|
24 |
keys1 = set(f1.keys())
|
25 |
keys2 = set(f2.keys())
|
@@ -30,18 +52,31 @@ def merge_safetensors(file_path1, file_path2, p, lambda_val):
|
|
30 |
tensor2 = f2.get_tensor(key)
|
31 |
tensor1, tensor2 = resize_tensors(tensor1, tensor2)
|
32 |
merged_tensors[key] = tensor1 + lambda_val * merge_tensors(tensor1, tensor2, p)
|
33 |
-
|
34 |
|
35 |
return merged_tensors
|
36 |
|
37 |
-
class BinDataHandler
|
38 |
-
|
|
|
|
|
|
|
39 |
self.data = data
|
40 |
|
41 |
-
def get_tensor(self, key):
|
42 |
return self.data[key]
|
43 |
|
44 |
-
def read_tensors(file_path, ext):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
if ext == ".safetensors" and file_path.endswith(".safetensors"):
|
46 |
f = safe_open(file_path, framework="pt", device="cpu")
|
47 |
return f, set(f.keys())
|
@@ -51,11 +86,20 @@ def read_tensors(file_path, ext):
|
|
51 |
return f, set(data.keys())
|
52 |
return None, None
|
53 |
|
54 |
-
def resize_tensors(tensor1, tensor2):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
if len(tensor1.shape) not in [1, 2]:
|
56 |
return tensor1, tensor2
|
57 |
|
58 |
-
# Pad along the last dimension (width)
|
59 |
if tensor1.shape[-1] < tensor2.shape[-1]:
|
60 |
padding_size = tensor2.shape[-1] - tensor1.shape[-1]
|
61 |
tensor1 = F.pad(tensor1, (0, padding_size, 0, 0))
|
@@ -63,7 +107,6 @@ def resize_tensors(tensor1, tensor2):
|
|
63 |
padding_size = tensor1.shape[-1] - tensor2.shape[-1]
|
64 |
tensor2 = F.pad(tensor2, (0, padding_size, 0, 0))
|
65 |
|
66 |
-
# Pad along the first dimension (height)
|
67 |
if tensor1.shape[0] < tensor2.shape[0]:
|
68 |
padding_size = tensor2.shape[0] - tensor1.shape[0]
|
69 |
tensor1 = F.pad(tensor1, (0, 0, 0, padding_size))
|
@@ -73,18 +116,28 @@ def resize_tensors(tensor1, tensor2):
|
|
73 |
|
74 |
return tensor1, tensor2
|
75 |
|
76 |
-
def merge_folder(tensor_map, directory_path, p, lambda_val):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
keys1 = set(tensor_map.keys())
|
78 |
-
# Some repos have both bin and safetensors, choose safetensors if so
|
79 |
ext = None
|
80 |
for filename in os.listdir(directory_path):
|
81 |
-
# Default to safetensors
|
82 |
if filename.endswith(".safetensors"):
|
83 |
ext = ".safetensors"
|
84 |
if filename.endswith(".bin") and ext is None:
|
85 |
ext = ".bin"
|
86 |
if ext is None:
|
87 |
-
raise "Could not find model files"
|
88 |
|
89 |
for filename in os.listdir(directory_path):
|
90 |
file_path = os.path.join(directory_path, filename)
|
@@ -95,7 +148,7 @@ def merge_folder(tensor_map, directory_path, p, lambda_val):
|
|
95 |
if "block_sparse_moe.gate" in key:
|
96 |
tensor1 = tensor_map[key]['tensor']
|
97 |
tensor2 = f.get_tensor(key)
|
98 |
-
tensor_map[key]['tensor'] = (tensor1 + tensor2) /2.0
|
99 |
continue
|
100 |
tensor1 = tensor_map[key]['tensor']
|
101 |
tensor2 = f.get_tensor(key)
|
@@ -103,27 +156,48 @@ def merge_folder(tensor_map, directory_path, p, lambda_val):
|
|
103 |
tensor_map[key]['tensor'] = tensor1 + lambda_val * merge_tensors(tensor1, tensor2, p)
|
104 |
return tensor_map
|
105 |
|
106 |
-
def map_tensors_to_files(directory_path):
|
107 |
-
|
|
|
|
|
|
|
|
|
108 |
|
|
|
|
|
|
|
|
|
109 |
for filename in os.listdir(directory_path):
|
110 |
file_path = os.path.join(directory_path, filename)
|
111 |
f, keys = read_tensors(file_path, '.safetensors')
|
112 |
if keys:
|
113 |
for key in keys:
|
114 |
tensor = f.get_tensor(key)
|
115 |
-
tensor_map[key] = {'filename':filename, 'shape':tensor.shape, 'tensor': tensor}
|
116 |
-
|
117 |
return tensor_map
|
118 |
|
119 |
-
def copy_nontensor_files(from_path, to_path):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
for filename in os.listdir(from_path):
|
121 |
file_path = os.path.join(from_path, filename)
|
122 |
if from_path != to_path and not filename.startswith(".") and not filename.startswith("README") and not filename.endswith(".bin") and not filename.endswith(".safetensors") and not filename.endswith(".pt") and not os.path.isdir(file_path):
|
123 |
-
|
124 |
-
shutil.copyfile(file_path, to_path+'/'+filename)
|
|
|
|
|
|
|
|
|
125 |
|
126 |
-
|
|
|
|
|
|
|
127 |
metadata = {'format': 'pt'}
|
128 |
by_filename = {}
|
129 |
|
@@ -135,12 +209,14 @@ def save_tensor_map(tensor_map, output_folder):
|
|
135 |
by_filename[filename][key] = tensor
|
136 |
|
137 |
for filename in sorted(by_filename.keys()):
|
138 |
-
output_file = output_folder+'/'+filename
|
139 |
-
|
140 |
save_file(by_filename[filename], output_file, metadata=metadata)
|
141 |
|
142 |
def main():
|
143 |
-
|
|
|
|
|
144 |
parser = argparse.ArgumentParser(description='Merge two safetensor model files.')
|
145 |
parser.add_argument('base_model', type=str, help='The base model safetensor file')
|
146 |
parser.add_argument('second_model', type=str, help='The second model safetensor file')
|
@@ -162,4 +238,4 @@ def main():
|
|
162 |
save_file(merged, args.output_model)
|
163 |
|
164 |
if __name__ == '__main__':
|
165 |
-
main()
|
|
|
5 |
import torch
|
6 |
import torch.nn.functional as F
|
7 |
from safetensors.torch import safe_open, save_file
|
8 |
+
import logging
|
9 |
|
10 |
+
# Set up logging
|
11 |
+
logging.basicConfig(level=logging.INFO)
|
12 |
+
|
13 |
+
def merge_tensors(tensor1: torch.Tensor, tensor2: torch.Tensor, p: float) -> torch.Tensor:
|
14 |
+
"""
|
15 |
+
Merge two tensors using dropout and scaling.
|
16 |
+
|
17 |
+
Args:
|
18 |
+
tensor1 (torch.Tensor): The first tensor.
|
19 |
+
tensor2 (torch.Tensor): The second tensor.
|
20 |
+
p (float): Dropout probability.
|
21 |
+
|
22 |
+
Returns:
|
23 |
+
torch.Tensor: The merged tensor.
|
24 |
+
"""
|
25 |
delta = tensor2 - tensor1
|
|
|
26 |
m = torch.from_numpy(np.random.binomial(1, p, delta.shape)).to(tensor1.dtype)
|
|
|
27 |
delta_tilde = m * delta
|
|
|
28 |
delta_hat = delta_tilde / (1 - p)
|
29 |
return delta_hat
|
30 |
|
31 |
+
def merge_safetensors(file_path1: str, file_path2: str, p: float, lambda_val: float) -> dict:
|
32 |
+
"""
|
33 |
+
Merge two safetensors files.
|
34 |
+
|
35 |
+
Args:
|
36 |
+
file_path1 (str): Path to the first safetensors file.
|
37 |
+
file_path2 (str): Path to the second safetensors file.
|
38 |
+
p (float): Dropout probability.
|
39 |
+
lambda_val (float): Scaling factor.
|
40 |
|
41 |
+
Returns:
|
42 |
+
dict: A dictionary of merged tensors.
|
43 |
+
"""
|
44 |
+
merged_tensors = {}
|
45 |
with safe_open(file_path1, framework="pt", device="cpu") as f1, safe_open(file_path2, framework="pt", device="cpu") as f2:
|
46 |
keys1 = set(f1.keys())
|
47 |
keys2 = set(f2.keys())
|
|
|
52 |
tensor2 = f2.get_tensor(key)
|
53 |
tensor1, tensor2 = resize_tensors(tensor1, tensor2)
|
54 |
merged_tensors[key] = tensor1 + lambda_val * merge_tensors(tensor1, tensor2, p)
|
55 |
+
logging.info(f"Merging {key}")
|
56 |
|
57 |
return merged_tensors
|
58 |
|
59 |
+
class BinDataHandler:
|
60 |
+
"""
|
61 |
+
A handler for binary data files.
|
62 |
+
"""
|
63 |
+
def __init__(self, data: dict):
|
64 |
self.data = data
|
65 |
|
66 |
+
def get_tensor(self, key: str) -> torch.Tensor:
|
67 |
return self.data[key]
|
68 |
|
69 |
+
def read_tensors(file_path: str, ext: str) -> tuple:
|
70 |
+
"""
|
71 |
+
Read tensors from a file.
|
72 |
+
|
73 |
+
Args:
|
74 |
+
file_path (str): Path to the file.
|
75 |
+
ext (str): File extension.
|
76 |
+
|
77 |
+
Returns:
|
78 |
+
tuple: A tuple containing the file handler and the set of keys.
|
79 |
+
"""
|
80 |
if ext == ".safetensors" and file_path.endswith(".safetensors"):
|
81 |
f = safe_open(file_path, framework="pt", device="cpu")
|
82 |
return f, set(f.keys())
|
|
|
86 |
return f, set(data.keys())
|
87 |
return None, None
|
88 |
|
89 |
+
def resize_tensors(tensor1: torch.Tensor, tensor2: torch.Tensor) -> tuple:
|
90 |
+
"""
|
91 |
+
Resize tensors to ensure they have the same shape.
|
92 |
+
|
93 |
+
Args:
|
94 |
+
tensor1 (torch.Tensor): The first tensor.
|
95 |
+
tensor2 (torch.Tensor): The second tensor.
|
96 |
+
|
97 |
+
Returns:
|
98 |
+
tuple: A tuple containing the resized tensors.
|
99 |
+
"""
|
100 |
if len(tensor1.shape) not in [1, 2]:
|
101 |
return tensor1, tensor2
|
102 |
|
|
|
103 |
if tensor1.shape[-1] < tensor2.shape[-1]:
|
104 |
padding_size = tensor2.shape[-1] - tensor1.shape[-1]
|
105 |
tensor1 = F.pad(tensor1, (0, padding_size, 0, 0))
|
|
|
107 |
padding_size = tensor1.shape[-1] - tensor2.shape[-1]
|
108 |
tensor2 = F.pad(tensor2, (0, padding_size, 0, 0))
|
109 |
|
|
|
110 |
if tensor1.shape[0] < tensor2.shape[0]:
|
111 |
padding_size = tensor2.shape[0] - tensor1.shape[0]
|
112 |
tensor1 = F.pad(tensor1, (0, 0, 0, padding_size))
|
|
|
116 |
|
117 |
return tensor1, tensor2
|
118 |
|
119 |
+
def merge_folder(tensor_map: dict, directory_path: str, p: float, lambda_val: float) -> dict:
|
120 |
+
"""
|
121 |
+
Merge tensors from a directory of model files.
|
122 |
+
|
123 |
+
Args:
|
124 |
+
tensor_map (dict): A dictionary mapping tensor keys to their file paths.
|
125 |
+
directory_path (str): Path to the directory containing model files.
|
126 |
+
p (float): Dropout probability.
|
127 |
+
lambda_val (float): Scaling factor.
|
128 |
+
|
129 |
+
Returns:
|
130 |
+
dict: A dictionary of merged tensors.
|
131 |
+
"""
|
132 |
keys1 = set(tensor_map.keys())
|
|
|
133 |
ext = None
|
134 |
for filename in os.listdir(directory_path):
|
|
|
135 |
if filename.endswith(".safetensors"):
|
136 |
ext = ".safetensors"
|
137 |
if filename.endswith(".bin") and ext is None:
|
138 |
ext = ".bin"
|
139 |
if ext is None:
|
140 |
+
raise FileNotFoundError("Could not find model files")
|
141 |
|
142 |
for filename in os.listdir(directory_path):
|
143 |
file_path = os.path.join(directory_path, filename)
|
|
|
148 |
if "block_sparse_moe.gate" in key:
|
149 |
tensor1 = tensor_map[key]['tensor']
|
150 |
tensor2 = f.get_tensor(key)
|
151 |
+
tensor_map[key]['tensor'] = (tensor1 + tensor2) / 2.0
|
152 |
continue
|
153 |
tensor1 = tensor_map[key]['tensor']
|
154 |
tensor2 = f.get_tensor(key)
|
|
|
156 |
tensor_map[key]['tensor'] = tensor1 + lambda_val * merge_tensors(tensor1, tensor2, p)
|
157 |
return tensor_map
|
158 |
|
159 |
+
def map_tensors_to_files(directory_path: str) -> dict:
|
160 |
+
"""
|
161 |
+
Map tensors to their respective files in a directory.
|
162 |
+
|
163 |
+
Args:
|
164 |
+
directory_path (str): Path to the directory containing model files.
|
165 |
|
166 |
+
Returns:
|
167 |
+
dict: A dictionary mapping tensor keys to their file paths.
|
168 |
+
"""
|
169 |
+
tensor_map = {}
|
170 |
for filename in os.listdir(directory_path):
|
171 |
file_path = os.path.join(directory_path, filename)
|
172 |
f, keys = read_tensors(file_path, '.safetensors')
|
173 |
if keys:
|
174 |
for key in keys:
|
175 |
tensor = f.get_tensor(key)
|
176 |
+
tensor_map[key] = {'filename': filename, 'shape': tensor.shape, 'tensor': tensor}
|
|
|
177 |
return tensor_map
|
178 |
|
179 |
+
def copy_nontensor_files(from_path: str, to_path: str):
|
180 |
+
"""
|
181 |
+
Copy non-tensor files from one directory to another.
|
182 |
+
|
183 |
+
Args:
|
184 |
+
from_path (str): Path to the source directory.
|
185 |
+
to_path (str): Path to the destination directory.
|
186 |
+
"""
|
187 |
for filename in os.listdir(from_path):
|
188 |
file_path = os.path.join(from_path, filename)
|
189 |
if from_path != to_path and not filename.startswith(".") and not filename.startswith("README") and not filename.endswith(".bin") and not filename.endswith(".safetensors") and not filename.endswith(".pt") and not os.path.isdir(file_path):
|
190 |
+
logging.info(f"Copying {file_path} to {to_path}")
|
191 |
+
shutil.copyfile(file_path, to_path + '/' + filename)
|
192 |
+
|
193 |
+
def save_tensor_map(tensor_map: dict, output_folder: str):
|
194 |
+
"""
|
195 |
+
Save the merged tensor map to the output directory.
|
196 |
|
197 |
+
Args:
|
198 |
+
tensor_map (dict): A dictionary of merged tensors.
|
199 |
+
output_folder (str): Path to the output directory.
|
200 |
+
"""
|
201 |
metadata = {'format': 'pt'}
|
202 |
by_filename = {}
|
203 |
|
|
|
209 |
by_filename[filename][key] = tensor
|
210 |
|
211 |
for filename in sorted(by_filename.keys()):
|
212 |
+
output_file = output_folder + '/' + filename
|
213 |
+
logging.info(f"Saving: {output_file}")
|
214 |
save_file(by_filename[filename], output_file, metadata=metadata)
|
215 |
|
216 |
def main():
|
217 |
+
"""
|
218 |
+
Main function to parse command-line arguments and orchestrate the merging process.
|
219 |
+
"""
|
220 |
parser = argparse.ArgumentParser(description='Merge two safetensor model files.')
|
221 |
parser.add_argument('base_model', type=str, help='The base model safetensor file')
|
222 |
parser.add_argument('second_model', type=str, help='The second model safetensor file')
|
|
|
238 |
save_file(merged, args.output_model)
|
239 |
|
240 |
if __name__ == '__main__':
|
241 |
+
main()
|