-
Notifications
You must be signed in to change notification settings - Fork 1
/
change_mp.py
168 lines (157 loc) · 7.7 KB
/
change_mp.py
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
import sys
import os
import torch
import copy
def change_mp(cp_path: str, target_mp: int, fix_module_parameters: set=None):
checkpoint = cp_path
assert os.path.isdir(checkpoint)
iteration_file = os.path.join(checkpoint, 'latest_checkpointed_iteration.txt')
if os.path.exists(iteration_file):
with open(iteration_file) as fin:
iteration = fin.read().strip() # 兼容 best
checkpoint = os.path.join(checkpoint, str(iteration))
else:
iteration = None
filenames = os.listdir(checkpoint)
filenames = [filename for filename in filenames if filename.startswith("mp_rank_")]
filenames = sorted(filenames,
key=lambda x: int(x.split('_')[2]))
filenames = [os.path.join(checkpoint, x) for x in filenames]
if target_mp == len(filenames):
print("MP size keeps the same.")
exit(0)
if cp_path[-1] == '/':
new_checkpoint = cp_path[:-1] + f'_MP{target_mp}'
else:
new_checkpoint = cp_path + f'_MP{target_mp}'
new_cp_path = new_checkpoint
if not os.path.exists(new_checkpoint):
os.mkdir(new_checkpoint)
if iteration is not None:
with open(os.path.join(new_checkpoint, 'latest_checkpointed_iteration.txt'), 'w') as fout:
fout.write("{}\n".format(iteration))
new_checkpoint = os.path.join(new_checkpoint, str(iteration))
if not os.path.exists(new_checkpoint):
os.mkdir(new_checkpoint)
preserve_keys = [
"lr_scheduler",
"skipped_steps",
"global_steps",
"global_samples",
"dp_world_size",
"iteration",
"client_lr_scheduler",
"np_rng_state",
"random_rng_state",
"torch_rng_state",
"cuda_rng_state",
"rng_tracker_states",
"important_nonparameters",
]
if target_mp < len(filenames):
print("Decrease MP size.")
assert len(filenames) % target_mp == 0
ratio = len(filenames) // target_mp
for i in range(target_mp):
start = ratio * i
end = ratio * (i + 1)
d = torch.load(filenames[start],
map_location='cpu')
for k in d.keys():
if k != 'module':
if k in preserve_keys:
pass
elif k == "mp_world_size":
d[k] = target_mp
else:
d[k] = None
for j in range(start + 1, end):
d_new = torch.load(filenames[j],
map_location='cpu')
for k, v in d_new['module'].items():
assert len(v.shape) < 3
if k in fix_module_parameters:
d['module'][k] = v.clone()
elif len(v.shape) == 2 and 'position' not in k:
if 'query' in k:
size_1 = d['module'][k].shape[0] // 3
size_2 = v.shape[0] // 3
target = d['module'][k]
d['module'][k] = torch.cat([
target[:size_1, :], v[:size_2, :],
target[size_1:size_1 * 2, :], v[size_2:size_2 * 2, :],
target[size_1 * 2:, :], v[size_2 * 2:, :]], 0)
elif 'word' in k or 'h_to_4h' in k or 'relative' in k or "r_w_bias" in k or "r_r_bias" in k:
d['module'][k] = torch.cat([d['module'][k], v], 0)
else:
d['module'][k] = torch.cat([d['module'][k], v], 1)
elif len(v.shape) == 1 and 'query_key_value' in k:
size_1 = d['module'][k].shape[0] // 3
size_2 = v.shape[0] // 3
target = d['module'][k]
d['module'][k] = torch.cat([
target[:size_1], v[:size_2],
target[size_1:size_1 * 2], v[size_2:size_2 * 2],
target[size_1 * 2:], v[size_2 * 2:]], 0)
elif len(v.shape) == 1 and ('dense_h_to_4h' in k or "attention.relative" in k):
d['module'][k] = torch.cat([d['module'][k], v], 0)
filename = os.path.join(new_checkpoint, "mp_rank_{:02d}_model_states.pt".format(i))
torch.save(d, filename)
if target_mp > len(filenames):
print("Increase MP size.")
assert target_mp % len(filenames) == 0
ratio = target_mp // len(filenames)
for i in range(len(filenames)):
start = ratio * i
end = ratio * (i + 1)
d = torch.load(filenames[i],
map_location='cpu')
for j in range(start, end):
d_new = {}
shift = j - start
for k, v in d.items():
if k != 'module':
if k in preserve_keys:
d_new[k] = copy.deepcopy(d[k])
elif k == "mp_world_size":
d_new[k] = target_mp
else:
d_new[k] = None
d_new['module'] = {}
with torch.no_grad():
for k, v in d['module'].items():
assert len(v.shape) < 3
if fix_module_parameters and k in fix_module_parameters:
d_new['module'][k] = v.clone()
elif len(v.shape) == 2 and 'position' not in k:
if 'query' in k:
part = v.shape[0] // ratio // 3
d_new['module'][k] = torch.cat([v[shift * part:(shift + 1) * part, :].clone(),
v[(shift + ratio) * part:(shift + 1 + ratio) * part,
:].clone(),
v[(shift + 2 * ratio) * part:(shift + 1 + 2 * ratio) * part,
:].clone()], 0)
elif 'word' in k or 'h_to_4h' in k or 'relative' in k or "r_w_bias" in k or "r_r_bias" in k:
part = v.shape[0] // ratio
d_new['module'][k] = v[shift * part:(shift + 1) * part, :].clone()
else:
part = v.shape[1] // ratio
d_new['module'][k] = v[:, shift * part:(shift + 1) * part].clone()
elif len(v.shape) == 1 and ('dense_h_to_4h' in k or "attention.relative" in k):
part = v.shape[0] // ratio
d_new['module'][k] = v[shift * part:(shift + 1) * part].clone()
elif len(v.shape) == 1 and 'query_key_value' in k:
part = v.shape[0] // ratio // 3
d_new['module'][k] = torch.cat(
[v[shift * part:(shift + 1) * part].clone(),
v[(shift + ratio) * part:(shift + 1 + ratio) * part].clone(),
v[(shift + 2 * ratio) * part:(shift + 1 + 2 * ratio) * part].clone()], 0)
else:
d_new['module'][k] = v.clone()
filename = os.path.join(new_checkpoint, "mp_rank_{:02d}_model_states.pt".format(j))
torch.save(d_new, filename)
return new_cp_path
if __name__ == "__main__":
cp_path = sys.argv[1]
target_mp = int(sys.argv[2])
change_mp(cp_path, target_mp)