-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
358 lines (292 loc) · 13 KB
/
train.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
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
import argparse
from itertools import chain
import json
import time
import os
from contextlib import contextmanager
import contextlib
import torch
from datasets import load_dataset, load_from_disk
from transformers import default_data_collator
import wandb
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel
from torch.utils.data.distributed import DistributedSampler
from optimizers import Adafactor
from model.megabyte_in_action import Megabyte
from model.megabyte_transformers import MegabyteConfig, MegabyteLMHeadModel
def get_model_and_tokenizer(args):
model = MegabyteLMHeadModel.from_pretrained(args.model_config_or_pretrained_model, Megabyte)
model = model.inner_model.to(torch.bfloat16)
tokenizer = MegabyteTokenizer(
eos_token_id=model.config.eos_id,
pad_id=model.config.pad_id,
)
return model, tokenizer
def fixed_seq_length_of_datasets(
datasets,
fixed_seq_length,
tokenizer,
load_from_cache_file=False,
):
block_size = fixed_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]])
# Padding in front of tokens to align it with the group size.
if total_length % block_size != 0:
count_pad_ids = block_size - (total_length % block_size)
concatenated_examples[list(examples.keys())[0]] = count_pad_ids*[tokenizer.pad_id] + concatenated_examples[list(examples.keys())[0]]
# 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 = datasets.map(
group_texts,
batched=True,
num_proc=os.cpu_count(),
load_from_cache_file=load_from_cache_file,
desc=f"Grouping texts in chunks of {block_size}",
)
return lm_datasets
def prepare_dataloader(args, tokenizer, dp):
step_interval = args.gradient_accumulation_steps
assert args.batch_size % (dp.world_size * step_interval) == 0
per_device_train_batch_size = args.batch_size//(dp.world_size*step_interval)
if args.load_dataset_from_disk:
raw_datasets = load_from_disk(args.load_dataset_from_disk)
else:
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], add_eos_token=True),
batched=True,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on dataset",
)
lm_datasets = fixed_seq_length_of_datasets(
tokenized_datasets,
args.seq_length,
tokenizer,
load_from_cache_file=not args.overwrite_cache,
)
train_dataset = lm_datasets["train"]
# TODO: when DistributedSampler turns on shuffle, Dataloader does not work properly, fix this issue.
train_sampler = DistributedSampler(train_dataset, dp.world_size, dp.rank, shuffle=False)
train_dataloader = torch.utils.data.DataLoader(
train_dataset, shuffle=False, collate_fn=default_data_collator,
batch_size=per_device_train_batch_size,
sampler=train_sampler,
num_workers=1,
)
if args.length_expolation_eval:
eval_dataloaders = []
del tokenized_datasets["train"]
for eval_seq_length in [args.seq_length, args.seq_length*2, args.seq_length*8]:
lm_datasets = fixed_seq_length_of_datasets(
tokenized_datasets,
eval_seq_length,
tokenizer,
load_from_cache_file=not args.overwrite_cache,
)
eval_dataset = lm_datasets["validation"]
eval_dataloader = torch.utils.data.DataLoader(
eval_dataset, shuffle=True, collate_fn=default_data_collator,
batch_size=args.eval_batch_size,
generator=torch.Generator(device=dp.device),
)
eval_dataloaders.append(eval_dataloader)
return train_dataloader, eval_dataloaders
eval_dataset = lm_datasets["validation"]
eval_dataloader = torch.utils.data.DataLoader(
eval_dataset, shuffle=True, collate_fn=default_data_collator,
batch_size=args.eval_batch_size,
)
return train_dataloader, eval_dataloader
def train(args, model, train_dataloader, eval_dataloader_or_dataloaders, dp):
print("start training")
print("args -", json.dumps(vars(args), sort_keys=True, indent=4))
model_config = model.module.config._asdict()
print("model.config -", json.dumps(model_config, sort_keys=True, indent=4))
if dp.is_main_process:
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):
output = model(ids=ids, return_loss=True, return_metrics=True)
loss = output.loss
wandb.log(output.metrics, commit=False)
return loss
def model_eval(model, eval_dataloader):
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
def model_eval2(model, dataloader_or_dataloaders):
if args.length_expolation_eval:
eval_losses = []
spend_time_l = []
for i, dataloader in enumerate(dataloader_or_dataloaders):
eval_loss, spend_time = model_eval(model, dataloader)
eval_losses.append(eval_loss)
spend_time_l.append(spend_time)
wandb.log({f"eval_loss-{i}": eval_loss}, commit=False)
return eval_losses, spend_time_l
eval_loss, spend_time = model_eval(model, dataloader_or_dataloaders)
wandb.log({f"eval_loss": eval_loss}, commit=False)
return eval_loss, spend_time
optimizer = Adafactor(model.parameters(), dynamic_weight_decay=True)
optimizer.zero_grad()
time_stone = time.time()
completed_steps = 0
total_loss = 0
for i, batch in enumerate(train_dataloader, 1):
with dp.accumulate(model):
ids = batch["input_ids"].to(dp.device)
loss = model_forward(model, ids)
loss_copy = loss.detach().float()
total_loss += loss_copy
loss /= args.gradient_accumulation_steps
loss.backward()
if dp.sync_gradients:
optimizer.step()
optimizer.zero_grad()
completed_steps += 1
avg_loss = total_loss / args.gradient_accumulation_steps
spend_time = time.time() - time_stone
wandb.log({
"loss": avg_loss,
"spend_time": spend_time,
}, step=completed_steps)
total_loss = 0
time_stone = time.time()
if dp.is_main_process:
print(f"step-{completed_steps}, loss={avg_loss}, spend_time={spend_time}")
if completed_steps % args.eval_interval == 0:
eval_loss, spend_time = model_eval2(model, eval_dataloader_or_dataloaders)
dp.barrier()
if dp.is_main_process:
eval_loss, spend_time = model_eval2(model, eval_dataloader_or_dataloaders)
print(f"training ends, final eval_loss={eval_loss}")
wandb.log({}, commit=True)
if args.save:
from model.megabyte_transformers import MegabyteLMHeadModel
model = MegabyteLMHeadModel.from_native_megabyte(model)
model.save_pretrained(args.save)
dp.barrier()
class MegabyteTokenizer:
def __init__(self, eos_token_id, pad_id):
super().__init__()
self.eos_token_id = eos_token_id
self.pad_id = pad_id
def __call__(self, text_or_list, return_tensors="pt", add_eos_token=False):
if isinstance(text_or_list, str):
text_or_list = [text_or_list]
tokens = [bytearray(text.encode("utf-8")) for text in text_or_list]
if add_eos_token:
tokens = [list(x) + [self.eos_token_id] for x in tokens]
return {"input_ids": tokens}
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--model_config_or_pretrained_model", required=True)
parser.add_argument("--batch_size", type=int, default=32)
parser.add_argument("--gradient_accumulation_steps", type=int, default=1)
parser.add_argument("--eval_batch_size", type=int, default=1)
parser.add_argument("--seq_length", type=int, default=2048)
parser.add_argument("--length_expolation_eval", action="store_true")
parser.add_argument("--overwrite_cache", action="store_true")
parser.add_argument("--dataset_name", default=None)
parser.add_argument("--dataset_config_name", default=None)
parser.add_argument("--gpu", action="store_true")
parser.add_argument("--eval_interval", default=1000, type=int)
parser.add_argument("--save", default=None)
parser.add_argument("--load_dataset_from_disk", default=None)
args = parser.parse_args()
return args
class DataParallel:
def __init__(self, gradient_accumulation_steps=1):
# TODO: init process group with backend "nccl|gloo" does not work, fix this issue.
dist.init_process_group(backend="nccl")
self.world_size = dist.get_world_size()
self.rank = dist.get_rank()
self.local_rank = int(os.environ["LOCAL_RANK"])
self.sync_gradients = False
self.completed_steps = 0
self.num_forward = 0
self.gradient_accumulation_steps = gradient_accumulation_steps
self.device = f"cuda:{self.local_rank}"
self.is_main_process = (self.rank == 0)
@contextmanager
def main_process_first(self):
if not self.is_main_process:
dist.barrier()
yield
class DataParallel:
def __init__(self, gradient_accumulation_steps=1):
# TODO: init process group with backend "nccl|gloo" does not work, fix this issue.
dist.init_process_group(backend="nccl")
self.world_size = dist.get_world_size()
self.rank = dist.get_rank()
self.local_rank = int(os.environ["LOCAL_RANK"])
self.sync_gradients = False
self.completed_steps = 0
self.num_forward = 0
self.gradient_accumulation_steps = gradient_accumulation_steps
self.device = f"cuda:{self.local_rank}"
self.is_main_process = (self.rank == 0)
@contextmanager
def main_process_first(self):
if not self.is_main_process:
dist.barrier()
yield
if self.is_main_process:
dist.barrier()
@contextmanager
def accumulate(self, model):
self.num_forward += 1
self.sync_gradients = self.num_forward % self.gradient_accumulation_steps == 0
if self.sync_gradients:
context = contextlib.nullcontext
else:
context = model.no_sync
with context():
yield
def barrier(self):
dist.barrier()
def main():
args = get_args()
dp = DataParallel(gradient_accumulation_steps=args.gradient_accumulation_steps)
torch.set_default_device(f"cuda:{dp.local_rank}")
print(f"model megabyte is being created...")
model, tokenizer = get_model_and_tokenizer(args)
ddp_model = DistributedDataParallel(model)
print(f"model megabyte has been created.")
num_parameters = sum([p.numel() for p in model.parameters()])
print(f"num_parameters {num_parameters}")
with dp.main_process_first():
train_dataloader, eval_dataloader_or_dataloaders = prepare_dataloader(args, tokenizer, dp)
train(args, ddp_model, train_dataloader, eval_dataloader_or_dataloaders, dp)
if __name__ == "__main__":
torch.multiprocessing.set_start_method('spawn')
main()