forked from tmabraham/hf_fsdp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
264 lines (206 loc) · 8.74 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
import torch
import argparse
import os
import math
import time
import random
import wandb
import transformers
import typing
from accelerate import Accelerator, FullyShardedDataParallelPlugin
from accelerate.state import AcceleratorState
from accelerate.utils import InitProcessGroupKwargs, set_seed
from datetime import timedelta
from datasets import load_dataset, concatenate_datasets
from torch.utils.data import DataLoader
from torch.distributed.fsdp.fully_sharded_data_parallel import (
FullOptimStateDictConfig,
FullStateDictConfig,
)
from transformers import (
get_cosine_schedule_with_warmup,
set_seed,
default_data_collator,
)
from tqdm import tqdm
#from together.modeling_flash_llama import LlamaForCausalLM
from transformers import LlamaForCausalLM
from llama_flash_attn_monkey_patch import replace_llama_attn_with_flash_attn
replace_llama_attn_with_flash_attn()
def main():
parser = argparse.ArgumentParser(description='Training script for the model')
parser.add_argument('--batch_size', type=int, default=1,
help='Batch size for training.')
parser.add_argument('--gradient_accumulate_every', type=int, default=1,
help='Gradient accumulation steps.')
parser.add_argument('--resume_from_checkpoint', type=str, default=None,
help='Path to resume training from a checkpoint.')
parser.add_argument('--checkpointing_steps', type=int, default=1000,
help='Steps interval for checkpointing.')
parser.add_argument('--output_dir', type=str, default="",
help='Directory to save the model and output.')
parser.add_argument('--wandb_entity', type=str, default="",
help='WandB entity.')
parser.add_argument('--wandb_project', type=str, default="",
help='WandB project.')
parser.add_argument('--wandb_name', type=str, default="",
help='WandB name.')
parser.add_argument('--wandb_id', type=str, default=None,
help='WandB ID.')
parser.add_argument('--model_name', type=str, default="meta-llama/Llama-2-7b-hf",
help='Model name or path to be loaded for training.')
parser.add_argument('--dataset_name', type=str, default="pubmed-llama-2-7b-tokenized-chunked",
help='Name or path of the dataset.')
args = parser.parse_args()
BATCH_SIZE = args.batch_size
GRADIENT_ACCUMULATE_EVERY = args.gradient_accumulate_every
RESUME_FROM_CHECKPOINT = args.resume_from_checkpoint
CHECKPOINTING_STEPS = args.checkpointing_steps
OUTPUT_DIR = args.output_dir
WANDB_ENTITY = args.wandb_entity
WANDB_PROJECT = args.wandb_project
WANDB_NAME = args.wandb_name
WANDB_ID = args.wandb_id
MODEL_NAME = args.model_name
DATASET_NAME = args.dataset_name
set_seed(42)
timeout = InitProcessGroupKwargs(timeout=timedelta(seconds=1_000_000))
accelerator = Accelerator(
gradient_accumulation_steps=GRADIENT_ACCUMULATE_EVERY,
log_with="wandb",
kwargs_handlers=[timeout],
)
# accelerator.state.fsdp_plugin.activation_checkpointing = True
accelerator.init_trackers(
project_name=WANDB_PROJECT,
init_kwargs={
"wandb": {
# "entity": WANDB_ENTITY,
"name": WANDB_NAME,
# "id": WANDB_ID,
# "resume": "must" if WANDB_ID else None,
}
},
)
accelerator.print(f"Total GPUS: {accelerator.num_processes}")
# accelerator.state.mixed_precision = "no"
# Create fresh LlamaForCausalLM model
model = LlamaForCausalLM.from_pretrained(
MODEL_NAME,
# torch_dtype=torch.bfloat16,
use_cache=False,
)
# model = model.to_bettertransformer()
model.gradient_checkpointing_enable()
model = accelerator.prepare(model)
# accelerator.state.mixed_precision = "bf16"
accelerator.print(f"FSDP model parameters per device: {model.num_parameters():,}")
accelerator.print(
f"Training a {accelerator.num_processes * model.num_parameters():,} parameter model"
)
# Dataloaders
train_dataset = load_dataset(
DATASET_NAME,
num_proc=os.cpu_count() - 1,
)["train"]
# Optional: Select random 0.1% of the dataset
# train_dataset = train_dataset.shuffle(seed=42).select(range(0, len(train_dataset), 1000))
# train_dataset = train_dataset.select(range(0, len(train_dataset) - len(train_dataset)%(BATCH_SIZE*16*8)))
# Optional: Repeat the dataset 100 times
# train_dataset = concatenate_datasets([train_dataset] * 100)
train_loader = DataLoader(
train_dataset,
collate_fn=default_data_collator,
shuffle=True,
batch_size=BATCH_SIZE,
num_workers=4,
pin_memory=True,
)
# Optimizer set up
optim = torch.optim.AdamW(model.parameters(), lr=5e-5) #, weight_decay=1e-6, betas=(0.9, 0.95))
# Determine number of training steps
max_train_steps = math.ceil(len(train_loader) / GRADIENT_ACCUMULATE_EVERY)
accelerator.print(f"Max train steps: {max_train_steps}")
# Dummy Scheduler for DeepSpeed
scheduler = get_cosine_schedule_with_warmup(
optim,
num_training_steps=max_train_steps,
num_warmup_steps=0*AcceleratorState().num_processes,
)
# prepare
optim, train_loader, scheduler = accelerator.prepare(optim, train_loader, scheduler)
# checkpoint scheduler
accelerator.register_for_checkpointing(scheduler)
# Recalculate
max_train_steps = math.ceil(len(train_loader) / GRADIENT_ACCUMULATE_EVERY)
accelerator.print(f"Max train steps recalculated: {max_train_steps}")
# Total batch size for logging
total_batch_size = (
BATCH_SIZE * accelerator.num_processes * GRADIENT_ACCUMULATE_EVERY
)
accelerator.print(f"Total batch size: {total_batch_size}")
# Resume training
progress_bar = tqdm(
range(max_train_steps), disable=not accelerator.is_local_main_process
)
completed_steps = 0
if RESUME_FROM_CHECKPOINT:
if RESUME_FROM_CHECKPOINT is not None or RESUME_FROM_CHECKPOINT != "":
accelerator.print(f"Resuming from checkpoint {RESUME_FROM_CHECKPOINT}")
accelerator.load_state(RESUME_FROM_CHECKPOINT)
path = os.path.basename(RESUME_FROM_CHECKPOINT)
training_difference = os.path.splitext(path)[0]
resume_step = int(training_difference.replace("step_", ""))
if RESUME_FROM_CHECKPOINT and resume_step is not None:
# We need to skip steps until we reach the resumed step
train_loader = accelerator.skip_first_batches(train_loader, resume_step)
completed_steps += resume_step
progress_bar.update(resume_step)
accelerator.print(f"Resuming training from step {resume_step}")
# Training
model.train()
print_fist_batch = True
for batch in train_loader:
with accelerator.accumulate(model):
if print_fist_batch:
print(batch)
print_fist_batch = False
loss = model(**batch).loss
accelerator.backward(loss)
step_loss = accelerator.reduce(loss.detach().clone(), reduction="mean").item()
accelerator.log({"loss": step_loss / GRADIENT_ACCUMULATE_EVERY}, step=completed_steps)
if accelerator.sync_gradients:
accelerator.clip_grad_norm_(model.parameters(), 1.0)
optim.step()
scheduler.step()
optim.zero_grad()
accelerator.log({"lr": optim.param_groups[0]["lr"]}, step=completed_steps)
if accelerator.sync_gradients:
progress_bar.update(1)
completed_steps += 1
if isinstance(CHECKPOINTING_STEPS, int):
if completed_steps % CHECKPOINTING_STEPS == 0:
output_dir = f"step_{completed_steps}"
if OUTPUT_DIR is not None:
output_dir = os.path.join(OUTPUT_DIR, output_dir)
accelerator.save_state(output_dir)
accelerator.print(f"Saving Finished")
if completed_steps >= max_train_steps:
break
# end training
accelerator.print(f"Training Finished")
accelerator.end_training()
# save final model
accelerator.print(f"Saving model to {OUTPUT_DIR}")
if OUTPUT_DIR is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
f"{OUTPUT_DIR}/final",
is_main_process=accelerator.is_main_process,
save_function=accelerator.save,
state_dict=accelerator.get_state_dict(model),
)
accelerator.print(f"Saving Finished")
if __name__ == "__main__":
main()