forked from Infini-AI-Lab/Sequoia
-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
214 lines (195 loc) · 10.1 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
import argparse
from collections import OrderedDict
from datetime import datetime
import glob
from loguru import logger
import numpy as np
import os
import random
import socket
import accelerate
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM, TrainingArguments
from transformers import RepetitionPenaltyLogitsProcessor
from trl import SFTTrainer
import src.utils as utils
def _get_model_state_dict(model, adapter_only=False):
logger.info('in _get_model_state_dict')
# state_dict = model.draft_model.state_dict(prefix='draft_model.')
state_dict = {k:v for k,v in model.state_dict().items() if 'draft_model' in k}
return state_dict
# This is a very hacky way to only save the trainable parameters.
accelerate.utils.fsdp_utils._get_model_state_dict = _get_model_state_dict
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument('--run_name', type=str, default=None)
parser.add_argument('--project_name', type=str, default='train_specdec')
parser.add_argument('--draft_type', choices=['standalone', 'eagle'], default='standalone',
help='The type of draft model to use.')
parser.add_argument('--target_model_path', type=str, default=None)
parser.add_argument('--draft_model_path', type=str, default=None)
parser.add_argument('--draft_ft_path', type=str, default=None)
parser.add_argument('--draft_random_init', action='store_true', default=False,
help='Whether to randomly initialize the draft model (instead of using pretrained weights).')
parser.add_argument('--freeze_embeds_and_lm_head', action='store_true', default=False,
help='Whether to freeze the standalone models embeddings and LM head.')
parser.add_argument('--convert_standalone', action='store_true', default=False,
help='Whether to train an eagle model by converting a standalone model).')
parser.add_argument('--train_data_files', type=str, default='/data/refined_web/train*.filtered.json',
help='Regex or comma separated list of json(l) files (potentially compressed) for training set.')
parser.add_argument('--valid_data_files', type=str, default='/data/data-pool/wiki_valid_single.jsonl',
help='Regex or comma separated list of json(l) files (potentially compressed) for validation set.')
parser.add_argument('--streaming', action='store_true')
parser.add_argument('--shuffle_buffer', type=int, default=5000,
help='Size of shuffle buffer. Only used when streaming is True.')
parser.add_argument('--seq_length', type=int, default=4096)
parser.add_argument('--min_completion_tokens', type=int, default=3)
parser.add_argument('--max_steps', type=int, default=10000)
parser.add_argument('--batch_size', type=int, default=4)
parser.add_argument('--eval_batch_size', type=int, default=None)
parser.add_argument('--gradient_accumulation_steps', type=int, default=1)
parser.add_argument('--learning_rate', type=float, default=1e-4)
parser.add_argument('--lr_scheduler_type', type=str, default='cosine')
parser.add_argument('--num_warmup_steps', type=int, default=100)
parser.add_argument('--weight_decay', type=float, default=0.05)
parser.add_argument('--fp16', action='store_true', default=False)
parser.add_argument('--bf16', action='store_true', default=False)
parser.add_argument('--no_gradient_checkpointing', action='store_false', default=False)
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--output_dir', type=str, default='./checkpoints')
parser.add_argument('--hf_model_cache', type=str, default='/data/.hf_cache/transformers/hub')
parser.add_argument('--resume_from_checkpoint', type=str, default=None)
parser.add_argument('--log_freq', default=10, type=int)
parser.add_argument('--eval_freq', default=100, type=int)
parser.add_argument('--save_freq', default=200, type=int)
parser.add_argument('--deepspeed_config', default=None, type=str)
parser.add_argument('--do_eval_before_training', action='store_true')
parser.add_argument('--no_special_tokens', action='store_true')
parser.add_argument('--no_flash_attn_2', action='store_true')
parser.add_argument('--fix_chat_format', action='store_true')
parser.add_argument('--pack', action='store_true', help='Whether to pack the input data.')
parser.add_argument('--lookahead', type=int, default=4)
parser.add_argument('--coefficient', type=float, default=1/16)
parser.add_argument('--distill_loss_weight', type=float, default=0.0)
parser.add_argument('--ce_loss_weight', type=float, default=1.0)
parser.add_argument('--l1_loss_weight', type=float, default=0.0)
parser.add_argument('--add_noise', action='store_true', default=False,
help='Whether to add noise to target model hidden states during Eagle training.')
parser.add_argument('--mem_profile_path', type=str, default=None)
return parser.parse_args()
if __name__ == '__main__':
logger.info(socket.gethostname())
args = get_args()
if args.mem_profile_path:
torch.cuda.memory._record_memory_history(max_entries=100_000)
if args.run_name is None:
args.run_name = f'{utils.date_str()}_{args.draft_type}_{args.output_dir.replace("/", "_")}'
os.environ['WANDB_PROJECT'] = args.project_name
os.environ['WANDB_CODE_DIR'] = os.path.dirname(os.path.abspath(__file__))
os.environ['WANDB_DIR'] = '/home/avner/wandb'
os.environ['WANDB__SERVICE_WAIT'] = '300'
# Load tokenizer
tokenizer = utils.get_tokenizer(args.draft_model_path, args.no_special_tokens)
# Load datasets
train_dataset = utils.get_dataset(
args.train_data_files,
tokenizer,
args.seq_length,
min_completion_tokens=args.min_completion_tokens,
is_train=True,
streaming=args.streaming,
shuffle_buffer=args.shuffle_buffer,
seed=args.seed,
pack=args.pack,
fix_chat_format=args.fix_chat_format,
)
valid_dataset = utils.get_dataset(
args.valid_data_files,
tokenizer,
args.seq_length,
min_completion_tokens=args.min_completion_tokens,
is_train=False,
streaming=args.streaming,
shuffle_buffer=args.shuffle_buffer,
seed=args.seed,
pack=args.pack,
fix_chat_format=args.fix_chat_format,
)
# Load training arguments
training_args = TrainingArguments(
output_dir=args.output_dir,
dataloader_drop_last=True,
evaluation_strategy='steps',
max_steps=args.max_steps,
eval_steps=args.eval_freq,
save_steps=args.save_freq,
logging_steps=args.log_freq,
per_device_train_batch_size=args.batch_size,
per_device_eval_batch_size=args.eval_batch_size or args.batch_size,
learning_rate=args.learning_rate,
lr_scheduler_type=args.lr_scheduler_type,
warmup_steps=args.num_warmup_steps,
gradient_accumulation_steps=args.gradient_accumulation_steps,
# TODO: enable gradient checkpoint on models?
# gradient_checkpointing=not args.no_gradient_checkpointing,
fp16=args.fp16,
bf16=args.bf16,
weight_decay=args.weight_decay,
run_name=args.run_name,
report_to='wandb',
ddp_find_unused_parameters=False,
deepspeed=args.deepspeed_config,
)
# Eagle code requires exact path to model folder.
# target_model_path = (args.target_model_path if args.draft_type != 'eagle' else
# utils.get_full_model_path(args.target_model_path, hf_model_cache=args.hf_model_cache))
# draft_model_path = (args.draft_model_path if args.draft_type != 'eagle' else
# utils.get_full_model_path(args.draft_model_path, hf_model_cache=args.hf_model_cache))
target_model_path = args.target_model_path
draft_model_path = args.draft_model_path
if args.draft_type == 'standalone':
model = utils.load_standalone_model(
target_model_path, draft_model_path, args.seq_length,
no_flash_attn_2=args.no_flash_attn_2,
ce_loss_weight=args.ce_loss_weight, distill_loss_weight=args.distill_loss_weight,
lookahead=args.lookahead, freeze_embeds_and_lm_head=args.freeze_embeds_and_lm_head,
)
elif args.draft_type == 'eagle':
model = utils.load_eagle_model(
target_model_path, draft_model_path, args.seq_length,
no_flash_attn_2=args.no_flash_attn_2,
ce_loss_weight=args.ce_loss_weight, distill_loss_weight=args.distill_loss_weight, l1_loss_weight=args.l1_loss_weight,
lookahead=args.lookahead, add_noise=args.add_noise, draft_random_init=args.draft_random_init,
convert_standalone=args.convert_standalone,
)
if args.draft_ft_path:
model = utils.load_draft_ft_ckpt(model, args.draft_ft_path)
# Prepare trainer
# In order for the keys in the metrics dictionary returned by compute_metrics to be correct,
# we must be consistent with the way the metrics are extracted by the Trainer in this code:
# https://github.com/huggingface/transformers/blob/965cf677695dd363285831afca8cf479cf0c600c/src/transformers/trainer.py#L3614
output_keys = model.output_keys()
output_keys.remove('loss')
compute_metrics_fn = lambda eval_preds: utils.compute_metrics(eval_preds, output_keys)
trainer = SFTTrainer(
model=model,
tokenizer=tokenizer,
args=training_args,
train_dataset=train_dataset,
eval_dataset=valid_dataset,
max_seq_length=args.seq_length,
data_collator=utils.DataCollatorWithPadding(args.seq_length) if not args.pack else None,
compute_metrics=compute_metrics_fn,
packing=True,
)
# train
if args.resume_from_checkpoint:
trainer.train(args.resume_from_checkpoint)
else:
trainer.train()
if args.mem_profile_path:
torch.cuda.memory._dump_snapshot(args.mem_profile_path)
torch.cuda.memory._record_memory_history(enabled=None)
print('Done training!')