-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
275 lines (228 loc) · 7.73 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
import argparse
import math
import os
import time
import nnt
import torch
from torch.optim.lr_scheduler import LRScheduler
from torch.utils.data import DataLoader
from tart import vocab
from tart.config import cfg, cfg_util
from tart.data import create_dataset
from tart.generation import generate_text
from tart.model.builder import build_model
from tart.model.weights import get_wd_params
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
def data_iterable(dataloader):
while True:
yield from dataloader
class LRSchedule(LRScheduler):
def __init__(self, optimizer, warmup_iter, max_iter, warmup_frac, final_frac):
self.warmup_iter = warmup_iter
self.warmup_frac = warmup_frac
self.final_frac = final_frac
self.max_iter = max_iter
super().__init__(optimizer)
def get_lr(self) -> list[float]:
i = self.last_epoch
if i < self.warmup_iter:
mult = (i / self.warmup_iter) * (1.0 - self.warmup_frac) + self.warmup_frac
else:
mult = (i - self.warmup_iter) / (self.max_iter - self.warmup_iter) * (
self.final_frac - 1.0
) + 1.0
return [lr * mult for lr in self.base_lrs]
def compute_loss(model, input, dtype=torch.bfloat16):
x = input[:, :-1].contiguous()
target = input[:, 1:].contiguous()
with torch.amp.autocast(device_type="cuda", dtype=dtype):
pred, (losses, correct) = model(x, target)
acc = correct.sum() / target.numel()
nnt.log({f"accuracy/{'train' if model.training else 'eval'}": acc.item()})
return pred, losses
def do_eval(model, val_loader, device, dtype):
val_it = 0
val_loss = 0
t = time.perf_counter()
max_val_iter = cfg.val_examples // cfg.optim.batch_size
model.eval()
with torch.no_grad():
for batch in val_loader:
batch = batch.to(device)
_, losses = compute_loss(model, batch, dtype)
val_loss += sum(l.item() for l in losses.values())
val_it += 1
if val_it >= max_val_iter:
break
nt = time.perf_counter()
val_time = nt - t
nnt.log(
{
"loss/val": val_loss / val_it,
"time/validation": val_time,
}
)
t = nt
model.train()
def train():
device = "cuda"
dtype = torch.bfloat16
max_iter = cfg.optim.total_steps
start_iter = 0
nnt.seed_all(2)
nnt.path_mgr.mkdirs(cfg.out_dir)
cfg_util.save_yaml(os.path.join(cfg.out_dir, "config.yaml"))
model = build_model().to(device)
print(model)
print(f"Model parameters: {sum(p.numel() for p in model.parameters()) / 1e6:.2f}M")
all_params = set(model.parameters())
wd_params = set(get_wd_params(model))
param_groups = [
{"params": list(wd_params), "weight_decay": cfg.optim.weight_decay},
{"params": list(all_params - wd_params), "weight_decay": 0},
]
optimizer = torch.optim.AdamW(
param_groups,
lr=cfg.optim.lr,
betas=(cfg.optim.beta1, cfg.optim.beta2),
eps=cfg.optim.eps,
)
lr_scheduler = LRSchedule(
optimizer,
warmup_iter=cfg.optim.warmup_steps,
max_iter=cfg.optim.total_steps,
warmup_frac=0.05,
final_frac=0.1,
)
train_ds = create_dataset(cfg.data.root, "train", cfg.model.seq_len + 1)
train_nbatch = len(train_ds)
val_ds = create_dataset(cfg.data.root, "validation", cfg.model.seq_len + 1)
train_loader = DataLoader(
train_ds,
num_workers=cfg.data.num_workers,
batch_size=cfg.optim.batch_size,
shuffle=True,
pin_memory=True,
drop_last=True,
)
val_loader = DataLoader(
val_ds,
num_workers=cfg.data.num_workers,
batch_size=cfg.optim.batch_size,
shuffle=True,
pin_memory=True,
drop_last=False,
persistent_workers=True,
)
nnt.log.add_writers(
nnt.TensorboardWriter(os.path.join(cfg.out_dir, "tb")),
nnt.MetricPrinter(cfg.log_period, max_iter),
)
checkpointer = nnt.Checkpointer(
model,
cfg.out_dir,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
)
if checkpointer.has_checkpoint():
ckpt = checkpointer.load(checkpointer.get_checkpoint_file())
start_iter = ckpt["iteration"] + 1
checkpointer = nnt.PeriodicCheckpointer(
checkpointer,
period=cfg.checkpoint_period,
max_iter=max_iter,
)
if cfg.compile:
model = torch.compile(model)
nnt.log.iter = start_iter
t = time.perf_counter()
data_iter = iter(data_iterable(train_loader))
tokens_per_batch = cfg.model.seq_len * cfg.optim.batch_size
ntok = 0
nbatch = 0
for iteration in range(start_iter, max_iter):
nt = time.perf_counter()
data_time = nt - t
grad_accum_steps = cfg.optim.grad_accum_steps
optimizer.zero_grad()
loss_accum = {}
for _ in range(grad_accum_steps):
nbatch += 1
batch = next(data_iter).to(device)
pred, losses = compute_loss(model, batch, dtype)
loss = sum(losses.values())
loss.backward()
for k, v in losses.items():
loss_accum[k] = v / grad_accum_steps + loss_accum.get(k, 0)
if len(loss_accum) > 1:
for k, v in loss_accum.items():
nnt.log({f"loss/{k}": v.item()})
ntok += tokens_per_batch * grad_accum_steps
nnt.log(
{
"loss/total": sum(loss_accum.values()).item(),
"Mtoken": ntok / 1e6,
"epoch": nbatch // train_nbatch,
}
)
torch.nn.utils.clip_grad_norm_(model.parameters(), cfg.optim.grad_norm_clip)
optimizer.step()
if iteration > 0 and iteration % cfg.val_period == 0:
plen = 128
vp = pred[0, :plen, :].argmax(-1).cpu()
vt = batch[0, 1 : plen + 1].cpu()
print("pred>", vocab.decode(vp.tolist()))
print("targ>", vocab.decode(vt.tolist()))
model.eval()
with torch.amp.autocast(device_type="cuda", dtype=dtype):
out = generate_text(
model,
prepend_eod=True,
max_new_tokens=256,
temperature=1.0,
top_p=0.0,
)
print(" gen>", out)
model.train()
nt = time.perf_counter()
iter_time = nt - t
t = nt
nnt.log(lr=lr_scheduler.get_last_lr()[0])
if iteration - start_iter > 2:
nnt.log(
{
"time/data": data_time,
"time/iter": iter_time,
}
)
if iteration > 0 and iteration % cfg.val_period == 0:
do_eval(model, val_loader, device, dtype)
if iteration > 0 and iteration % cfg.log_hists_period == 0:
for name, p in model.named_parameters():
nnt.log({f"parameters/{name}": p.data.to("cpu")})
lr_scheduler.step()
nnt.log.step()
checkpointer.step(iteration)
def main(args):
if args.cfg:
cfg_util.merge_yaml(args.cfg)
if args.opts:
cfg_util.merge_dotlist(args.opts)
print(cfg_util.to_yaml())
train()
def get_argument_parser():
parser = argparse.ArgumentParser()
parser.add_argument(
"--cfg", "-c", default="", metavar="FILE", help="path to config file"
)
parser.add_argument("--name", "-n", default="", help="Experiment name")
parser.add_argument(
"opts",
default=None,
nargs=argparse.REMAINDER,
)
return parser
if __name__ == "__main__":
args = get_argument_parser().parse_args()
main(args)