-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
179 lines (158 loc) · 7.26 KB
/
main.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
169
170
171
172
173
174
175
176
177
178
179
import os
import shutil
from parse import parse_args
from tqdm import tqdm
import torch
from data_tool import partition_data
from peft import (
prepare_model_for_kbit_training,
set_peft_model_state_dict,
get_peft_model_state_dict,
)
import time
import datetime
from fed_utils import (
Evaluator,
evaluate,
FedAvg,
client_selection,
GenerateClient,
FedNova,
initialize_server_and_client_control_variate,
load_variate,
ScaffoldAggregation,
)
from data_tool import DataTokenizer
from model_utils import PeftHelper, ModelHelper
# offline
os.environ['HF_DATASETS_OFFLINE'] = '1'
os.environ['TRANSFORMERS_OFFLINE'] = '1'
def main(args):
if int(os.environ.get("LOCAL_RANK", 0)) == 0:
print(f"Federated Learning PEFine-Tuning for LLM:\n")
for arg in vars(args):
print(f"{arg}: {getattr(args, arg)}")
assert args.global_model, "Please specify a --global_model, e.g. --global_model='decapoda-research/llama-7b-hf'"
assert os.path.exists(args.data_path), "Please generate the data files for each client"
gradient_accumulation_steps = args.local_batch_size // args.local_micro_batch_size
device_map = "auto"
world_size = int(os.environ.get("WORLD_SIZE", 1))
ddp = world_size != 1
if ddp:
device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}
gradient_accumulation_steps = gradient_accumulation_steps // world_size
# set up the global model & toknizer
model_helper = ModelHelper(global_model_name=args.model, global_model_path=args.global_model, device_map=device_map)
model, tokenizer = model_helper.get_model()
# since we load the model in 8-bit, so we need to prepare it for training
model = prepare_model_for_kbit_training(model)
# setup peft method
peft_helper = PeftHelper(model_name=args.model, peft_method=args.peft_method)
model, config = peft_helper.get_peft_model_for_training(args=args, model=model)
model.print_trainable_parameters()
data_tokenizer = DataTokenizer(args, tokenizer)
if not ddp and torch.cuda.device_count() > 1:
model.is_parallelizable = True
model.model_parallel = True
if args.useScaffold:
# initialize server control variate and client control variate.
dir_name = args.scaffold_dir
initialize_server_and_client_control_variate(model, args.num_clients, dir_name)
# if you want to resume training from checkpoint
# set these parameters
start_round = 0
if(args.resume_from_checkpoint):
# parameter_path = './lora-shepherd-7b/20news-dirichlet_label_uni-1-10/4/adapter_model.bin'
peft_weights = torch.load(args.parameter_path)
set_peft_model_state_dict(model, peft_weights,"default")
start_round = args.start_round
print("The process of federated instruction-tuning has started..")
previously_selected_clients_set = set()
# last_client_id = None
local_dataset_len_dict = dict()
output_dir = args.output_dir
# T_max = args.num_communication_rounds // 4
# two lines below are for evaluating the model after each round's training
evaluator = Evaluator(args)
evaluator.tokenizer = tokenizer
training_start_time = time.time()
for epoch in tqdm(range(start_round, args.num_communication_rounds)):
if args.useScaffold:
filename = os.path.join(dir_name, "server_c")
server_c = load_variate(filename)
else:
server_c = None
print("\nConducting the client selection")
selected_clients_set = client_selection(args.num_clients, args.client_selection_frac, args.client_selection_strategy,
other_info=epoch)
local_learning_rate = args.local_learning_rate
# local_learning_rate = cosine_annealing_warm_restart_LR(T_max, epoch, args.local_learning_rate)
print("learning rate of current communication: " + str(local_learning_rate))
for client_id in selected_clients_set:
if args.useScaffold:
filename = os.path.join(dir_name, "client"+str(client_id))
client_c = load_variate(filename)
else:
client_c = None
client = GenerateClient(args, client_id, model, output_dir, client_c, server_c)
print("\nPreparing the local dataset and trainer for Client_{}".format(client_id))
client.load_raw_load()
client.preprare_local_dataset(data_tokenizer.generate_and_tokenize_prompt, args.local_val_set_size)
client.build_local_trainer(tokenizer,
args.local_micro_batch_size,
gradient_accumulation_steps,
args.local_num_epochs,
local_learning_rate,
args.group_by_length,
ddp)
print("Initiating the local training of Client_{}".format(client_id))
client.initiate_local_training()
print("Local training starts ... ")
client.train()
print("\nTerminating the local training of Client_{}".format(client_id))
model, local_dataset_len_dict, previously_selected_clients_set = client.terminate_local_training(
epoch, local_dataset_len_dict, previously_selected_clients_set)
del client
print("Collecting the weights of clients and performing aggregation")
if args.useFedNova:
model = FedNova(model,
selected_clients_set,
output_dir,
local_dataset_len_dict,
args.local_batch_size,
epoch,
)
elif args.useScaffold:
model = ScaffoldAggregation(model,
selected_clients_set,
output_dir,
local_dataset_len_dict,
epoch,
server_c,
args.scaffold_dir,
args.num_clients,
)
else:
model = FedAvg(model,
selected_clients_set,
output_dir,
local_dataset_len_dict,
epoch,
)
# torch.save(get_peft_model_state_dict(model), os.path.join(output_dir, str(epoch), "adapter_model.bin"))
# save checkpoints every 5 rounds
# if epoch % 5 == 0:
torch.save(get_peft_model_state_dict(model), os.path.join(output_dir, "aggregated_model_{}.bin".format(epoch)))
# delete the clients's weights to save storage space, optional
shutil.rmtree(os.path.join(output_dir, str(epoch)))
config.save_pretrained(output_dir)
# if (epoch+1) % 5 == 0:
evaluate(epoch, evaluator, model, args.dataset)
print("END OF COMMUNICATION: " + str(epoch))
training_over_time = time.time()
training_time = int(round((training_over_time - training_start_time)))
print("Total training time: " + str(datetime.timedelta(seconds = training_time)))
if __name__ == "__main__":
args = parse_args()
# partition_data(args)
main(args)