-
Notifications
You must be signed in to change notification settings - Fork 51
/
utils.py
163 lines (119 loc) · 4.99 KB
/
utils.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
# coding=utf-8
"""Utilities for logging and serialization"""
import os
import random
import numpy as np
import torch
import mpu
import deepspeed
def print_rank_0(message):
if torch.distributed.is_initialized():
if torch.distributed.get_rank() == 0:
print(message, flush=True)
else:
print(message, flush=True)
def print_args(args):
"""Print arguments."""
print('arguments:', flush=True)
for arg in vars(args):
dots = '.' * (29 - len(arg))
print(' {} {} {}'.format(arg, dots, getattr(args, arg)), flush=True)
def save_rank_0(args, message):
if torch.distributed.is_initialized():
if torch.distributed.get_rank() == 0:
with open(args.log_file, "a") as f:
f.write(message + "\n")
f.flush()
else:
with open(args.log_file, "a") as f:
f.write(message + "\n")
f.flush()
def set_deepspeed_activation_checkpointing(args, num_checkpoints):
deepspeed.checkpointing.configure(mpu, deepspeed_config=args.deepspeed_config, num_checkpoints=num_checkpoints)
mpu.checkpoint = deepspeed.checkpointing.checkpoint
mpu.get_cuda_rng_tracker = deepspeed.checkpointing.get_cuda_rng_tracker
mpu.model_parallel_cuda_manual_seed = deepspeed.checkpointing.model_parallel_cuda_manual_seed
def initialize_distributed(args):
"""Initialize torch.distributed."""
# Manually set the device ids.
device = args.rank % torch.cuda.device_count()
if args.local_rank is not None:
device = args.local_rank
torch.cuda.set_device(device)
# Call the init process
init_method = 'tcp://'
master_ip = os.getenv('MASTER_ADDR', 'localhost')
master_port = os.getenv('MASTER_PORT', '6000')
init_method += master_ip + ':' + master_port
deepspeed.init_distributed()
# Set the model-parallel / data-parallel communicators.
mpu.initialize_model_parallel(args.model_parallel_size)
def set_random_seed(seed):
"""Set random seed for reproducability."""
if seed is not None and seed > 0:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
mpu.model_parallel_cuda_manual_seed(seed)
def ensure_directory_exists(filename):
dirname = os.path.dirname(filename)
if not os.path.exists(dirname):
os.makedirs(dirname)
def get_checkpoint_tracker_filename(checkpoints_path):
return os.path.join(checkpoints_path, 'latest_checkpointed_iteration.txt')
def save_checkpoint(iteration, model, optimizer,
lr_scheduler, args):
"""Save a model checkpoint."""
if args.deepspeed:
save_ds_checkpoint(iteration, model, args)
torch.distributed.barrier()
# Update the latest iteration
if torch.distributed.get_rank() == 0:
tracker_filename = get_checkpoint_tracker_filename(args.save)
with open(tracker_filename, 'w') as f:
f.write(str(iteration))
torch.distributed.barrier()
def save_ds_checkpoint(iteration, model, args):
"""Save a model checkpoint."""
sd = {}
sd['iteration'] = iteration
model.save_checkpoint(args.save, str(iteration), client_state = sd, save_zero=False)
def get_checkpoint_iteration(args):
tracker_filename = get_checkpoint_tracker_filename(args.load)
if not os.path.isfile(tracker_filename):
print_rank_0('WARNING: could not find the metadata file {} '.format(tracker_filename))
print_rank_0(' will not load any checkpoints and will start from RANDOM')
return 0, False
iteration = 0
with open(tracker_filename, 'r') as f:
metastring = f.read().strip()
try:
iteration = int(metastring)
except ValueError:
print_rank_0('ERROR: Invalid metadata file {}. Exiting'.format(tracker_filename))
exit()
assert iteration > 0, 'error parsing metadata file {}'.format(tracker_filename)
return iteration, True
def load_checkpoint(args, model, optimizer=None, lr_scheduler=None):
"""Load a model checkpoint."""
iteration, success = get_checkpoint_iteration(args)
if not success:
return 0
checkpoint_name, sd = model.load_checkpoint(args.load, iteration, load_optimizer_states=args.load_optimizer_states, load_lr_scheduler_states=args.load_lr_scheduler_states, load_module_strict=(not args.no_load_strict))
if checkpoint_name is None:
if mpu.get_data_parallel_rank() == 0:
print("Unable to load checkpoint.")
return iteration
try:
iteration = sd['iteration']
except KeyError:
try: # Backward compatible with older checkpoints
iteration = sd['total_iters']
except KeyError:
print_rank_0('A metadata file exists but Unable to load iteration '
' from checkpoint {}, exiting'.format(checkpoint_name))
exit()
torch.distributed.barrier()
if mpu.get_data_parallel_rank() == 0:
print(' successfully loaded {}'.format(checkpoint_name))
return iteration