-
Notifications
You must be signed in to change notification settings - Fork 3
/
benchmark.py
238 lines (199 loc) · 8.51 KB
/
benchmark.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
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
import argparse
from itertools import chain
import torch
from datasets import load_dataset
from transformers import (
GPT2Tokenizer,
GPT2LMHeadModel,
GPT2Config,
default_data_collator,
)
import wandb
import json
import time
from optimizers import Adafactor
def get_model_and_tokenizer(args):
if args.model == "megabyte":
from model import Megabyte, MegabyteConfig
elif args.model == "megabyte_in_action":
from model.megabyte_in_action import Megabyte, MegabyteConfig
if args.model in ["megabyte", "megabyte_in_action"]:
PAD_ID = 257
EOS_ID = 258
V = 512
config = MegabyteConfig(
V=V,
P=8,
D_G=128,
D_L=256,
T_MAX=args.max_seq_length,
g_nheads=16,
l_nheads=4,
g_nlayers=12,
l_nlayers=6,
initializer_range=0.02,
pad_id=PAD_ID,
eos_id=EOS_ID,
)
model = Megabyte(config).to(torch.bfloat16)
tokenizer = MegabyteTokenizer(EOS_ID)
elif args.model == "gpt2":
tokenizer = GPT2Tokenizer.from_pretrained(args.model_path)
config = GPT2Config.from_pretrained(args.model_path)
model = GPT2LMHeadModel(config).to(torch.bfloat16)
else:
raise Exception(f"model {args.model} is not supported")
return model, tokenizer
def prepare_dataloader(args, tokenizer):
raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
column_names = raw_datasets["train"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
tokenized_datasets = raw_datasets.map(
lambda examples: tokenizer(examples[text_column_name]),
batched=True,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on dataset",
)
block_size = args.max_seq_length
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
if total_length >= block_size:
total_length = (total_length // block_size) * block_size
# Split by chunks of max_len.
result = {
k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
for k, t in concatenated_examples.items()
}
result["labels"] = result["input_ids"].copy()
return result
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
load_from_cache_file=not args.overwrite_cache,
desc=f"Grouping texts in chunks of {block_size}",
)
train_dataset = lm_datasets["train"]
train_dataloader = torch.utils.data.DataLoader(
train_dataset, shuffle=True, collate_fn=default_data_collator,
batch_size=args.batch_size,
generator=torch.Generator(device="cuda") if args.gpu else torch.Generator(device="cpu"),
)
eval_dataset = lm_datasets["validation"]
eval_dataloader = torch.utils.data.DataLoader(
eval_dataset, shuffle=True, collate_fn=default_data_collator,
batch_size=args.batch_size,
generator=torch.Generator(device="cuda") if args.gpu else torch.Generator(device="cpu"),
)
return train_dataloader, eval_dataloader
def train(args, model, dataloader, eval_dataloader):
print("start training")
print("args -", json.dumps(vars(args), sort_keys=True, indent=4))
if args.model == "megabyte":
model_config = model.config._asdict()
elif args.model == "gpt2":
model_config = model.config.to_dict()
else:
model_config = {}
print("model.config -", json.dumps(model_config, sort_keys=True, indent=4))
wandb.login()
run = wandb.init(
# Set the project where this run will be logged
project="megabyte-benchmark",
# Track hyperparameters and run metadata
config={
"args": args,
"model config": model_config,
"num parameters": sum([p.numel() for p in model.parameters()]),
})
def model_forward(model, ids):
if args.model == "megabyte":
output = model(ids=ids, return_loss=True, return_metrics=True)
loss = output.loss
wandb.log(output.metrics, commit=False)
elif args.model == "gpt2":
input_ids = ids[:, :-1]
labels = ids[:, 1:]
loss = model(input_ids=input_ids, labels=labels).loss
elif args.model == "MEGABYTE_pytorch":
if args.gpu:
ids = ids.to("cuda")
loss, output_norms = model(ids, return_loss=True)
wandb.log(output_norms, commit=False)
return loss
def model_eval(model):
model.eval()
eval_start_timestamp = time.time()
losses = []
for batch in eval_dataloader:
ids = batch["input_ids"]
with torch.no_grad():
loss = model_forward(model, ids)
losses.append(loss.reshape(1))
eval_loss = torch.cat(losses).mean()
model.train()
return eval_loss, time.time() - eval_start_timestamp
optimizer = Adafactor(model.parameters(), dynamic_weight_decay=True)
optimizer.zero_grad()
log_interval = 1
eval_interval = 1000
time_stone = time.time()
for step_i, batch in enumerate(dataloader, 1):
ids = batch["input_ids"]
loss = model_forward(model, ids)
loss.backward()
optimizer.step()
optimizer.zero_grad()
if step_i % eval_interval == 0:
eval_loss, spend_time = model_eval(model)
print(f"step-{step_i}, eval_loss={eval_loss}")
wandb.log({"step": step_i, "eval_loss": eval_loss, "eval_spend_time": spend_time}, commit=False)
if step_i % log_interval == 0:
print(f"step-{step_i}, loss={loss}, spend_time={time.time() - time_stone}")
wandb.log({"step": step_i, "loss": loss, "spend_time": time.time() - time_stone})
time_stone = time.time()
eval_loss, spend_time = model_eval(model)
print(f"training ends, final eval_loss={eval_loss}")
wandb.log({"eval_loss": eval_loss, "eval_spend_time": spend_time})
if args.save:
if args.model == "megabyte":
from model.megabyte_transformers import MegabyteLMHeadModel
model = MegabyteLMHeadModel.from_native_megabyte(model)
model.save_pretrained(args.save)
class MegabyteTokenizer:
def __init__(self, eos_token_id):
super().__init__()
self.eos_token_id = eos_token_id
def __call__(self, text_or_list, return_tensors="pt"):
if isinstance(text_or_list, str):
text_or_list = [text_or_list]
tokens = [bytearray(text.encode("utf-8")) for text in text_or_list]
return {"input_ids": tokens}
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--batch_size", type=int, default=32)
parser.add_argument("--max_seq_length", type=int, default=2048)
parser.add_argument("--model", choices=["gpt2", "megabyte", "megabyte_in_action"], required=True)
parser.add_argument("--model_path", default=None)
parser.add_argument("--overwrite_cache", action="store_true")
parser.add_argument("--dataset_name", required=True)
parser.add_argument("--dataset_config_name", required=True)
parser.add_argument("--gpu", action="store_true")
parser.add_argument("--save", default=None)
args = parser.parse_args()
if args.gpu:
torch.set_default_device("cuda")
print(f"model {args.model} is being created...")
model, tokenizer = get_model_and_tokenizer(args)
print(f"model {args.model} has been created.")
num_parameters = sum([p.numel() for p in model.parameters()])
print(f"num_parameters {num_parameters}")
train_dataloader, eval_dataloader = prepare_dataloader(args, tokenizer)
train(args, model, train_dataloader, eval_dataloader)
if __name__ == "__main__":
main()