-
Notifications
You must be signed in to change notification settings - Fork 183
/
66173c47-b15b-4a24-a835-60c82f6b8283.txt
2165 lines (2092 loc) · 134 KB
/
66173c47-b15b-4a24-a835-60c82f6b8283.txt
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
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import os
import sys
with open(sys.argv[0]) as f:
code = f.read() # read the code of this file ASAP, for logging
import uuid
import glob
import time
import contextlib
from dataclasses import dataclass
import numpy as np
import torch
from torch import nn
import torch.nn.functional as F
import torch.distributed as dist
import torch._inductor.config as config
from torch.nn.parallel import DistributedDataParallel as DDP
# Use of FlexAttention contributed by @KoszarskyB
from torch.nn.attention.flex_attention import flex_attention, create_block_mask
flex_attention = torch.compile(flex_attention, dynamic=False)
create_block_mask = torch.compile(create_block_mask, dynamic=False)
# -----------------------------------------------------------------------------
# Muon optimizer
def zeropower_via_svd(G, steps=None):
U, S, V = G.svd()
return U @ V.T
@torch.compile
def zeropower_via_newtonschulz5(G, steps=10, eps=1e-7):
"""
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
zero even beyond the point where the iteration no longer converges all the way to one everywhere
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
performance at all relative to UV^T, where USV^T = G is the SVD.
"""
assert len(G.shape) == 2
a, b, c = (3.4445, -4.7750, 2.0315)
X = G.bfloat16()
X /= (X.norm() + eps) # ensure top singular value <= 1
if G.size(0) > G.size(1):
X = X.T
for _ in range(steps):
A = X @ X.T
B = b * A + c * A @ A # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
X = a * X + B @ X
if G.size(0) > G.size(1):
X = X.T
return X
zeropower_backends = dict(svd=zeropower_via_svd, newtonschulz5=zeropower_via_newtonschulz5)
class Muon(torch.optim.Optimizer):
"""
Muon - MomentUm Orthogonalized by Newton-schulz
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
the advantage that it can be stably run in bfloat16 on the GPU.
Some warnings:
- This optimizer assumes that all parameters passed in are 2D.
- It should not be used for the embedding layer, the final fully connected layer, or any {0,1}-D
parameters; those should all be optimized by a standard method (e.g., AdamW).
- To use it with 4D convolutional filters, it works well to just flatten their last 3 dimensions.
- We believe it is unlikely to work well for training with small batch size.
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
- We have not yet tried this optimizer for training scenarios larger than NanoGPT (124M).
Arguments:
lr: The learning rate used by the internal SGD.
momentum: The momentum used by the internal SGD.
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
backend: The chosen backend for the orthogonalization step. (recommended: 'newtonschulz5')
backend_steps: The number of iteration steps to use in the backend, if it is iterative.
"""
def __init__(self, params, lr=0.02, momentum=0.95, nesterov=True,
backend='newtonschulz5', backend_steps=5):
defaults = dict(lr=lr, momentum=momentum, nesterov=nesterov, backend=backend, backend_steps=backend_steps)
super().__init__(params, defaults)
def step(self):
for group in self.param_groups:
lr = group['lr']
momentum = group['momentum']
zeropower_backend = zeropower_backends[group['backend']]
# generate weight updates in distributed fashion
total_params = sum(p.numel() for p in group['params'])
updates_flat = torch.zeros(total_params, device='cuda', dtype=torch.bfloat16)
curr_idx = 0
for i, p in enumerate(group['params']):
# luckily this will perfectly distribute a transformer with multiple of 4 layers to 8 GPUs
if i % int(os.environ['WORLD_SIZE']) == int(os.environ['RANK']):
g = p.grad
assert g is not None
state = self.state[p]
if 'momentum_buffer' not in state:
state['momentum_buffer'] = torch.zeros_like(g)
buf = state['momentum_buffer']
buf.mul_(momentum).add_(g)
g = g.add(buf, alpha=momentum) if group['nesterov'] else buf
g = zeropower_backend(g, steps=group['backend_steps'])
g *= max(1, g.size(0)/g.size(1))**0.5
updates_flat[curr_idx:curr_idx+p.numel()] = g.flatten()
curr_idx += p.numel()
# sync updates across devices. we are not memory-constrained so can do this simple deserialization
dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM)
# deserialize and apply updates
curr_idx = 0
for p in group['params']:
g = updates_flat[curr_idx:curr_idx+p.numel()].view_as(p.data).type_as(p.data)
p.data.add_(g, alpha=-lr)
curr_idx += p.numel()
# -----------------------------------------------------------------------------
# PyTorch nn.Module definitions for the GPT-2 model
def norm(x):
return F.rms_norm(x, (x.size(-1),))
class CastedLinear(nn.Linear):
def __init__(self, in_features, out_features):
super().__init__(in_features, out_features, bias=False)
def forward(self, x):
return F.linear(x, self.weight.to(x.dtype))
class Rotary(torch.nn.Module):
def __init__(self, dim, base=10000):
super().__init__()
self.register_buffer('inv_freq', (1 / base) ** (torch.arange(0, dim, 2) / dim))
self.seq_len_cached = None
self.cos_cached = None
self.sin_cached = None
def forward(self, x):
seq_len = x.shape[1]
if seq_len != self.seq_len_cached:
t = torch.arange(seq_len, device=x.device)
freqs = torch.outer(t, self.inv_freq)
self.seq_len_cached = seq_len
self.cos_cached = freqs.cos()
self.sin_cached = freqs.sin()
cos, sin = self.cos_cached[None, :, None, :], self.sin_cached[None, :, None, :]
# apply_rotary_emb(x, cos, sin)
x1, x2 = x.chunk(2, dim=3)
y1 = x1 * cos + x2 * sin
y2 = x1 * (-sin) + x2 * cos
return torch.cat((y1, y2), 3).type_as(x)
class CausalSelfAttention(nn.Module):
def __init__(self, dim, n_head):
super().__init__()
assert dim % n_head == 0
self.n_head = n_head
self.c_q = CastedLinear(dim, dim)
self.c_k = CastedLinear(dim, dim)
self.c_v = CastedLinear(dim, dim)
# value residual lambda
self.lamb = nn.Parameter(torch.tensor(0.5)) # @Grad62304977
# rotary embeddings
self.rotary = Rotary(dim // n_head) # dim // n_head = head_dim
# output projection
self.c_proj = CastedLinear(dim, dim)
self.c_proj.weight.data.zero_() # zero init suggested by @Grad62304977
def forward(self, x, vi, block_mask):
B, T = x.size(0), x.size(1) # batch size, sequence length
assert B == 1, "Must use batch size = 1 for FlexAttention"
q = self.c_q(x).view(B, T, self.n_head, -1)
k = self.c_k(x).view(B, T, self.n_head, -1)
v = self.c_v(x).view(B, T, self.n_head, -1)
v = (1 - self.lamb) * v + self.lamb * vi.view_as(v) # @Grad62304977
q, k = norm(q), norm(k) # QK norm suggested by @Grad62304977
q, k = self.rotary(q), self.rotary(k)
y = flex_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), block_mask=block_mask)
y = y.transpose(1, 2).contiguous().view_as(x) # re-assemble all head outputs side by side
y = self.c_proj(y)
return y
class MLP(nn.Module):
def __init__(self, dim):
super().__init__()
self.c_fc = CastedLinear(dim, 4 * dim)
self.c_proj = CastedLinear(4 * dim, dim)
self.c_proj.weight.data.zero_() # zero init suggested by @Grad62304977
def forward(self, x):
x = self.c_fc(x)
x = F.relu(x).square() # https://arxiv.org/abs/2109.08668v2; ~1-2% better than GELU; suggested by @SKYLINEZ007 and @Grad62304977
x = self.c_proj(x)
return x
class Block(nn.Module):
def __init__(self, config):
super().__init__()
self.attn = CausalSelfAttention(config.n_embd, config.n_head)
self.mlp = MLP(config.n_embd)
self.lambdas = nn.Parameter(torch.tensor([1., 0.]))
def forward(self, x, vi, x0, block_mask):
x = self.lambdas[0] * x + self.lambdas[1] * x0
x = x + self.attn(norm(x), vi, block_mask)
x = x + self.mlp(norm(x))
return x
# -----------------------------------------------------------------------------
# The main GPT-2 model
@dataclass
class GPTConfig:
vocab_size : int = 50304
n_layer : int = 12
n_head : int = 6 # head dim 128 suggested by @Grad62304977
n_embd : int = 768
class GPT(nn.Module):
def __init__(self, config):
super().__init__()
# U-net design by @brendanh0gan
self.num_encoder_layers = config.n_layer // 2 # Half of the layers for encoder
self.num_decoder_layers = config.n_layer - self.num_encoder_layers # Remaining for decoder
# Add learnable skip connection weights for decoder layers
self.skip_weights = nn.Parameter(torch.ones(self.num_decoder_layers))
self.transformer = nn.ModuleDict(dict(
wte = nn.Embedding(config.vocab_size, config.n_embd),
# token value embeddings by @KoszarskyB - inspired by @Grad62304977's value residual learning
vte = nn.Embedding(config.vocab_size, config.n_embd*12),
h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
))
self.lm_head = CastedLinear(config.n_embd, config.vocab_size)
self.lm_head.weight.data.zero_() # @Grad62304977
def forward(self, idx, target, attn_blocksize):
docs = (idx == 50256).cumsum(0)
def document_causal_mask(b, h, q_idx, kv_idx):
causal_mask = q_idx >= kv_idx
document_mask = docs[q_idx] == docs[kv_idx]
window_mask = q_idx - kv_idx < attn_blocksize
return causal_mask & document_mask & window_mask
S = len(idx)
block_mask = create_block_mask(document_causal_mask, None, None, S, S, device="cuda", _compile=True)
# forward the GPT model itself
x = self.transformer.wte(idx[None]) # token embeddings of shape (b, t, n_embd)
x = norm(x) # @Grad62304977
x0 = x
vi = self.transformer.vte(idx[None]).chunk(12, dim=-1)
# Store outputs for U-Net skip connections
skip_connections = []
# Encoder pass - process only the first half of the blocks
for i in range(self.num_encoder_layers):
x = self.transformer.h[i](x, vi[i], x0, block_mask)
skip_connections.append(x)
# Decoder pass - process the remaining blocks with weighted skip connections
for i in range(self.num_decoder_layers):
x = x + self.skip_weights[i] * skip_connections.pop()
x = self.transformer.h[self.num_encoder_layers + i](x, vi[self.num_encoder_layers+i], x0, block_mask)
x = norm(x)
logits = self.lm_head(x)
logits = 30 * torch.tanh(logits / 30) # @Grad62304977
logits = logits.float()
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), target.view(-1))
return loss
# -----------------------------------------------------------------------------
# Our own simple Distributed Data Loader
def _peek_data_shard(filename):
# only reads the header, returns header data
with open(filename, "rb") as f:
# first read the header, which is 256 int32 integers (4 bytes each)
header = np.frombuffer(f.read(256*4), dtype=np.int32)
if header[0] != 20240520:
print("ERROR: magic number mismatch in the data .bin file!")
print("---> HINT: Are you passing in a correct file with --input_bin?")
print("---> HINT: Dataset encoding changed recently, re-run data prepro or refer again to README")
print("---> HINT: For example re-run: `python dev/data/tinyshakespeare.py`, then re-try")
exit(1)
assert header[1] == 1, "unsupported version"
ntok = header[2] # number of tokens (claimed)
return ntok # for now just return the number of tokens
def _load_data_shard(filename):
with open(filename, "rb") as f:
# first read the header, which is 256 int32 integers (4 bytes each)
header = np.frombuffer(f.read(256*4), dtype=np.int32)
assert header[0] == 20240520, "magic number mismatch in the data .bin file"
assert header[1] == 1, "unsupported version"
ntok = header[2] # number of tokens (claimed)
# the rest of it are tokens, stored as uint16
tokens = np.frombuffer(f.read(), dtype=np.uint16)
assert len(tokens) == ntok, "number of tokens read does not match header?"
return tokens
class DistributedDataLoader:
def __init__(self, filename_pattern, T, process_rank, num_processes):
self.process_rank = process_rank
self.num_processes = num_processes
self.T = T
# glob files that match the pattern
self.files = sorted(glob.glob(filename_pattern))
assert len(self.files) > 0, f"did not find any files that match the pattern {filename_pattern}"
# load and validate all data shards, count number of tokens in total
ntok_total = 0
for fname in self.files:
shard_ntok = _peek_data_shard(fname)
assert shard_ntok >= num_processes * T + 1
ntok_total += int(shard_ntok)
self.ntok_total = ntok_total
self.reset()
def reset(self):
self.current_shard = -1
self.advance()
def advance(self): # advance to next data shard
self.current_shard = (self.current_shard + 1) % len(self.files)
self.current_position = self.process_rank * self.T
self.tokens = _load_data_shard(self.files[self.current_shard])
def next_batch(self):
batch_size = self.T * self.num_processes
buf = self.tokens[self.current_position:self.current_position+self.T+1]
buf = torch.tensor(buf.astype(np.int32), dtype=torch.long)
x = buf[:-1] # inputs
y = buf[1:] # targets
# advance current position and load next shard if necessary
self.current_position += batch_size
if self.current_position + batch_size >= len(self.tokens):
self.advance()
return x.cuda(), y.cuda()
# -----------------------------------------------------------------------------
# int main
@dataclass
class Hyperparameters:
# data hyperparams
input_bin : str = 'data/fineweb10B/fineweb_train_*.bin' # input .bin to train on
input_val_bin : str = 'data/fineweb10B/fineweb_val_*.bin' # input .bin to eval validation loss on
# optimization hyperparams
batch_size : int = 8 # batch size, in sequences, across all devices
sequence_length : int = 64*1024 # sequence length, in tokens
num_iterations : int = 1530 # number of iterations to run
warmup_iters : int = 0
cooldown_iters : int = 600 # number of iterations of linear warmup/cooldown for triangular or trapezoidal schedule
weight_decay : float = 0
# evaluation and logging hyperparams
val_loss_every : int = 125 # every how many steps to evaluate val loss? 0 for only at the end
val_tokens : int = 10485760 # how many tokens of validation data? it's important to keep this fixed for consistent comparisons
save_every : int = 0 # every how many steps to save the checkpoint? 0 for only at the end
args = Hyperparameters()
# set up DDP (distributed data parallel). torchrun sets this env variable
assert torch.cuda.is_available()
dist.init_process_group(backend='nccl')
ddp_rank = int(os.environ['RANK'])
ddp_local_rank = int(os.environ['LOCAL_RANK'])
ddp_world_size = int(os.environ['WORLD_SIZE'])
device = f'cuda:{ddp_local_rank}'
torch.cuda.set_device(device)
print(f"using device: {device}")
master_process = (ddp_rank == 0) # this process will do logging, checkpointing etc.
# begin logging
logfile = None
if master_process:
run_id = str(uuid.uuid4())
logdir = 'logs/%s/' % run_id
os.makedirs(logdir, exist_ok=True)
logfile = 'logs/%s.txt' % run_id
# create the log file
with open(logfile, "w") as f:
# begin the log by printing this file (the Python code)
f.write(code)
f.write('='*100 + '\n')
def print0(s, logonly=False):
if master_process:
with open(logfile, "a") as f:
if not logonly:
print(s)
f.write(s+'\n')
# log information about the hardware/software environment this is running on
# and print the full `nvidia-smi` to file
print0(f"Running pytorch {torch.version.__version__} compiled for CUDA {torch.version.cuda}\nnvidia-smi:")
import subprocess
result = subprocess.run(['nvidia-smi'], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
print0(f'{result.stdout}', logonly=True)
print0('='*100, logonly=True)
# convenience variables
T = args.sequence_length
# calculate the number of steps to take in the val loop.
assert args.val_tokens % (T * ddp_world_size) == 0
val_steps = args.val_tokens // (T * ddp_world_size)
# calculate the steps of gradient accumulation required to attain the desired global batch size.
assert args.batch_size % (ddp_world_size) == 0
train_accumulation_steps = args.batch_size // ddp_world_size
# load tokens
train_loader = DistributedDataLoader(args.input_bin, T, ddp_rank, ddp_world_size)
val_loader = DistributedDataLoader(args.input_val_bin, T, ddp_rank, ddp_world_size)
print0(f"Training DataLoader: total number of tokens: {train_loader.ntok_total} across {len(train_loader.files)} files")
print0(f"Validation DataLoader: total number of tokens: {val_loader.ntok_total} across {len(val_loader.files)} files")
print0('='*100, logonly=True)
x, y = train_loader.next_batch()
# there are only 50257 unique GPT-2 tokens; we extend to nearest multiple of 128 for efficiency. suggested to me by @Grad62304977.
# this originates from Karpathy's experiments.
num_vocab = 50304
model = GPT(GPTConfig(vocab_size=num_vocab, n_layer=12, n_head=6, n_embd=768))
model = model.cuda().bfloat16()
for m in model.modules():
if isinstance(m, CastedLinear):
m.float()
if hasattr(config, "coordinate_descent_tuning"):
config.coordinate_descent_tuning = True # suggested by @Chillee
model = torch.compile(model)
# here we wrap model into DDP container
model = DDP(model, device_ids=[ddp_local_rank])
raw_model = model.module # always contains the "raw" unwrapped model
# init the optimizer(s)
optimizer1 = torch.optim.Adam([raw_model.transformer.wte.weight, raw_model.transformer.vte.weight], lr=0.6, betas=(0.8, 0.95), fused=True)
optimizer2 = torch.optim.Adam([raw_model.lm_head.weight], lr=0.008, betas=(0.8, 0.95), fused=True)
params = list(raw_model.transformer.h.parameters())
matrix_params = [p for p in params if p.ndim == 2]
scalar_params = [p for p in params if p.ndim < 2] + [raw_model.skip_weights]
optimizer3 = Muon(matrix_params, lr=0.05, momentum=0.95)
optimizer4 = torch.optim.Adam(scalar_params, lr=0.04, betas=(0.8, 0.95), fused=True) # note that this learning rate is neither sensitive nor tuned
optimizers = [optimizer1, optimizer2, optimizer3, optimizer4]
# learning rate decay scheduler (linear warmup and cooldown)
def get_lr(it):
assert it <= args.num_iterations
# 1) linear warmup for warmup_iters steps
if it < args.warmup_iters:
return (it+1) / args.warmup_iters
# 2) constant lr for a while
elif it < args.num_iterations - args.cooldown_iters:
return 1.0
# 3) linear cooldown
else:
decay_ratio = (args.num_iterations - it) / args.cooldown_iters
return decay_ratio
schedulers = [torch.optim.lr_scheduler.LambdaLR(opt, get_lr) for opt in optimizers]
# Start training loop
training_time_ms = 0
# start the clock
torch.cuda.synchronize()
t0 = time.time()
# begin training
for step in range(args.num_iterations + 1):
last_step = (step == args.num_iterations)
# This effectively ignores timing first 10 steps, which are slower for weird reasons.
# Alternately, and slightly more correctly in terms of benchmarking, we could do 10
# steps with dummy data first, and then re-initialize the model and reset the loader.
if step == 10:
training_time_ms = 0
t0 = time.time()
timed_steps = float('nan') if step <= 11 else (step - 10) + 1 # <= 11 to avoid bug in val
# Set the attention blocksize for the current step, in chunks of 64. By @fernbear.bsky.social
attn_blocksize = torch.tensor(64*((step/args.num_iterations * (1792 - 64) + 64)//64), dtype=torch.int, device='cuda')
# once in a while evaluate the validation dataset
if (last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0)):
# stop the clock
torch.cuda.synchronize()
training_time_ms += 1000 * (time.time() - t0)
# run validation batches
model.eval()
val_loader.reset()
val_loss = 0.0
for _ in range(val_steps):
with torch.no_grad():
x_val, y_val = val_loader.next_batch()
val_loss += model(x_val, y_val, attn_blocksize=attn_blocksize)
dist.all_reduce(val_loss, op=dist.ReduceOp.AVG)
val_loss /= val_steps
# log val loss to console and to logfile
print0(f'step:{step}/{args.num_iterations} val_loss:{val_loss:.4f} train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms/(timed_steps-1):.2f}ms')
# start the clock again
torch.cuda.synchronize()
t0 = time.time()
if master_process and (last_step or (args.save_every > 0 and step % args.save_every == 0)):
# stop the clock
torch.cuda.synchronize()
training_time_ms += 1000 * (time.time() - t0)
# save the state of the training process
log = dict(step=step, code=code, model=raw_model.state_dict(), optimizers=[opt.state_dict() for opt in optimizers])
torch.save(log, 'logs/%s/state_step%06d.pt' % (run_id, step))
# start the clock again
torch.cuda.synchronize()
t0 = time.time()
# bit confusing: we want to make sure to eval on 0th iteration
# but also after the very last iteration. so we loop for step <= num_iterations
# instead of just < num_iterations (one extra due to <=), only to do
# the validation/sampling one last time, and then we break right here as we're done.
if last_step:
break
# --------------- TRAINING SECTION BEGIN -----------------
model.train()
for i in range(1, train_accumulation_steps+1):
ctx = model.no_sync() if i < train_accumulation_steps else contextlib.nullcontext()
with ctx: # there's no need to sync gradients every accumulation step
# forward pass
loss = model(x, y, attn_blocksize=attn_blocksize)
# advance the dataset for the next batch
x, y = train_loader.next_batch()
# backward pass
loss.backward()
train_loss = loss.detach()
for p in model.parameters():
p.grad /= train_accumulation_steps
# momentum warmup for Muon
frac = min(step/300, 1)
optimizer3.param_groups[0]['momentum'] = (1 - frac) * 0.85 + frac * 0.95
# step the optimizers and schedulers
for opt, sched in zip(optimizers, schedulers):
opt.step()
sched.step()
# null the gradients
model.zero_grad(set_to_none=True)
# --------------- TRAINING SECTION END -------------------
# everything that follows now is just diagnostics, prints, logging, etc.
#dist.all_reduce(train_loss, op=dist.ReduceOp.AVG) # all-reducing the training loss would be more correct in terms of logging, but slower
approx_time = training_time_ms + 1000 * (time.time() - t0)
print0(f"step:{step+1}/{args.num_iterations} train_loss:{train_loss.item():.4f} train_time:{approx_time:.0f}ms step_avg:{approx_time/timed_steps:.2f}ms")
if master_process:
print(f"peak memory consumption: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB")
# -------------------------------------------------------------------------
# clean up nice
dist.destroy_process_group()
====================================================================================================
Running pytorch 2.6.0.dev20241203+cu124 compiled for CUDA 12.4
nvidia-smi:
Thu Dec 5 03:03:39 2024
+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 535.183.06 Driver Version: 535.183.06 CUDA Version: 12.2 |
|-----------------------------------------+----------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+======================+======================|
| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 |
| N/A 38C P0 75W / 700W | 3MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 |
| N/A 30C P0 115W / 700W | 529MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 |
| N/A 31C P0 118W / 700W | 529MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 |
| N/A 38C P0 118W / 700W | 529MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 |
| N/A 39C P0 123W / 700W | 529MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 |
| N/A 29C P0 110W / 700W | 529MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 |
| N/A 38C P0 116W / 700W | 39MiB / 81559MiB | 1% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 |
| N/A 30C P0 118W / 700W | 529MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
+---------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=======================================================================================|
+---------------------------------------------------------------------------------------+
====================================================================================================
Training DataLoader: total number of tokens: 1100000000 across 11 files
Validation DataLoader: total number of tokens: 100000000 across 1 files
====================================================================================================
step:0/1530 val_loss:10.8258 train_time:0ms step_avg:nanms
step:1/1530 train_loss:10.8258 train_time:32094ms step_avg:nanms
step:2/1530 train_loss:10.0774 train_time:32205ms step_avg:nanms
step:3/1530 train_loss:8.3414 train_time:32364ms step_avg:nanms
step:4/1530 train_loss:7.5877 train_time:32524ms step_avg:nanms
step:5/1530 train_loss:7.4877 train_time:32685ms step_avg:nanms
step:6/1530 train_loss:7.0149 train_time:32845ms step_avg:nanms
step:7/1530 train_loss:7.2335 train_time:33006ms step_avg:nanms
step:8/1530 train_loss:6.7494 train_time:33165ms step_avg:nanms
step:9/1530 train_loss:6.6424 train_time:33326ms step_avg:nanms
step:10/1530 train_loss:6.5033 train_time:33486ms step_avg:nanms
step:11/1530 train_loss:6.4433 train_time:114ms step_avg:nanms
step:12/1530 train_loss:6.3414 train_time:274ms step_avg:nanms
step:13/1530 train_loss:6.2828 train_time:435ms step_avg:144.86ms
step:14/1530 train_loss:6.2135 train_time:595ms step_avg:148.77ms
step:15/1530 train_loss:6.1967 train_time:754ms step_avg:150.90ms
step:16/1530 train_loss:6.1162 train_time:916ms step_avg:152.59ms
step:17/1530 train_loss:6.1637 train_time:1076ms step_avg:153.73ms
step:18/1530 train_loss:5.9577 train_time:1236ms step_avg:154.51ms
step:19/1530 train_loss:6.0194 train_time:1398ms step_avg:155.31ms
step:20/1530 train_loss:5.6933 train_time:1557ms step_avg:155.72ms
step:21/1530 train_loss:5.9679 train_time:1718ms step_avg:156.15ms
step:22/1530 train_loss:6.1986 train_time:1877ms step_avg:156.43ms
step:23/1530 train_loss:5.8791 train_time:2038ms step_avg:156.78ms
step:24/1530 train_loss:6.0292 train_time:2199ms step_avg:157.08ms
step:25/1530 train_loss:5.7017 train_time:2359ms step_avg:157.26ms
step:26/1530 train_loss:5.5923 train_time:2520ms step_avg:157.49ms
step:27/1530 train_loss:5.8003 train_time:2680ms step_avg:157.64ms
step:28/1530 train_loss:5.4141 train_time:2841ms step_avg:157.83ms
step:29/1530 train_loss:5.6842 train_time:3000ms step_avg:157.91ms
step:30/1530 train_loss:5.4751 train_time:3161ms step_avg:158.04ms
step:31/1530 train_loss:5.4440 train_time:3322ms step_avg:158.18ms
step:32/1530 train_loss:5.2732 train_time:3483ms step_avg:158.31ms
step:33/1530 train_loss:5.5847 train_time:3644ms step_avg:158.42ms
step:34/1530 train_loss:5.4997 train_time:3805ms step_avg:158.53ms
step:35/1530 train_loss:5.6238 train_time:3965ms step_avg:158.60ms
step:36/1530 train_loss:5.5656 train_time:4126ms step_avg:158.68ms
step:37/1530 train_loss:5.4507 train_time:4287ms step_avg:158.77ms
step:38/1530 train_loss:5.2978 train_time:4447ms step_avg:158.81ms
step:39/1530 train_loss:5.3307 train_time:4608ms step_avg:158.90ms
step:40/1530 train_loss:5.2331 train_time:4769ms step_avg:158.97ms
step:41/1530 train_loss:5.2205 train_time:4930ms step_avg:159.02ms
step:42/1530 train_loss:5.1448 train_time:5090ms step_avg:159.07ms
step:43/1530 train_loss:5.2625 train_time:5251ms step_avg:159.11ms
step:44/1530 train_loss:5.2334 train_time:5413ms step_avg:159.20ms
step:45/1530 train_loss:5.3805 train_time:5573ms step_avg:159.23ms
step:46/1530 train_loss:5.1725 train_time:5733ms step_avg:159.26ms
step:47/1530 train_loss:5.0930 train_time:5895ms step_avg:159.33ms
step:48/1530 train_loss:5.2106 train_time:6055ms step_avg:159.34ms
step:49/1530 train_loss:5.1443 train_time:6214ms step_avg:159.34ms
step:50/1530 train_loss:5.2421 train_time:6375ms step_avg:159.37ms
step:51/1530 train_loss:5.1294 train_time:6536ms step_avg:159.42ms
step:52/1530 train_loss:5.0124 train_time:6696ms step_avg:159.44ms
step:53/1530 train_loss:5.1704 train_time:6856ms step_avg:159.45ms
step:54/1530 train_loss:4.9953 train_time:7016ms step_avg:159.46ms
step:55/1530 train_loss:5.4057 train_time:7178ms step_avg:159.50ms
step:56/1530 train_loss:5.0153 train_time:7337ms step_avg:159.50ms
step:57/1530 train_loss:4.8682 train_time:7498ms step_avg:159.53ms
step:58/1530 train_loss:5.0324 train_time:7659ms step_avg:159.56ms
step:59/1530 train_loss:5.0312 train_time:7819ms step_avg:159.57ms
step:60/1530 train_loss:5.1407 train_time:7980ms step_avg:159.60ms
step:61/1530 train_loss:4.8555 train_time:8141ms step_avg:159.63ms
step:62/1530 train_loss:4.9700 train_time:8302ms step_avg:159.65ms
step:63/1530 train_loss:4.9732 train_time:8462ms step_avg:159.65ms
step:64/1530 train_loss:4.9594 train_time:8622ms step_avg:159.66ms
step:65/1530 train_loss:4.7961 train_time:8782ms step_avg:159.68ms
step:66/1530 train_loss:4.9062 train_time:8942ms step_avg:159.68ms
step:67/1530 train_loss:4.8366 train_time:9105ms step_avg:159.73ms
step:68/1530 train_loss:5.0879 train_time:9265ms step_avg:159.75ms
step:69/1530 train_loss:4.7066 train_time:9426ms step_avg:159.77ms
step:70/1530 train_loss:4.8126 train_time:9587ms step_avg:159.79ms
step:71/1530 train_loss:4.9563 train_time:9747ms step_avg:159.79ms
step:72/1530 train_loss:4.8678 train_time:9909ms step_avg:159.83ms
step:73/1530 train_loss:4.7591 train_time:10070ms step_avg:159.84ms
step:74/1530 train_loss:4.8890 train_time:10230ms step_avg:159.85ms
step:75/1530 train_loss:4.8481 train_time:10391ms step_avg:159.87ms
step:76/1530 train_loss:4.7986 train_time:10551ms step_avg:159.87ms
step:77/1530 train_loss:4.8979 train_time:10712ms step_avg:159.89ms
step:78/1530 train_loss:5.1027 train_time:10873ms step_avg:159.89ms
step:79/1530 train_loss:4.8238 train_time:11033ms step_avg:159.89ms
step:80/1530 train_loss:4.8690 train_time:11193ms step_avg:159.90ms
step:81/1530 train_loss:4.6421 train_time:11353ms step_avg:159.91ms
step:82/1530 train_loss:4.7986 train_time:11514ms step_avg:159.92ms
step:83/1530 train_loss:4.7415 train_time:11674ms step_avg:159.91ms
step:84/1530 train_loss:4.7369 train_time:11834ms step_avg:159.92ms
step:85/1530 train_loss:4.6020 train_time:11995ms step_avg:159.93ms
step:86/1530 train_loss:4.8251 train_time:12154ms step_avg:159.92ms
step:87/1530 train_loss:4.7475 train_time:12314ms step_avg:159.92ms
step:88/1530 train_loss:4.7459 train_time:12475ms step_avg:159.93ms
step:89/1530 train_loss:4.6835 train_time:12635ms step_avg:159.94ms
step:90/1530 train_loss:4.6229 train_time:12794ms step_avg:159.93ms
step:91/1530 train_loss:4.6115 train_time:12955ms step_avg:159.94ms
step:92/1530 train_loss:4.7860 train_time:13116ms step_avg:159.95ms
step:93/1530 train_loss:4.6059 train_time:13276ms step_avg:159.95ms
step:94/1530 train_loss:4.6202 train_time:13437ms step_avg:159.96ms
step:95/1530 train_loss:4.6762 train_time:13597ms step_avg:159.96ms
step:96/1530 train_loss:4.5738 train_time:13757ms step_avg:159.96ms
step:97/1530 train_loss:4.6198 train_time:13916ms step_avg:159.95ms
step:98/1530 train_loss:4.5717 train_time:14077ms step_avg:159.96ms
step:99/1530 train_loss:4.6424 train_time:14237ms step_avg:159.96ms
step:100/1530 train_loss:4.6658 train_time:14398ms step_avg:159.98ms
step:101/1530 train_loss:4.5233 train_time:14558ms step_avg:159.98ms
step:102/1530 train_loss:4.6900 train_time:14718ms step_avg:159.97ms
step:103/1530 train_loss:4.5739 train_time:14878ms step_avg:159.98ms
step:104/1530 train_loss:4.5400 train_time:15037ms step_avg:159.97ms
step:105/1530 train_loss:4.5410 train_time:15199ms step_avg:159.99ms
step:106/1530 train_loss:4.6009 train_time:15359ms step_avg:159.99ms
step:107/1530 train_loss:4.4901 train_time:15519ms step_avg:159.99ms
step:108/1530 train_loss:4.3430 train_time:15680ms step_avg:160.00ms
step:109/1530 train_loss:4.4751 train_time:15840ms step_avg:160.00ms
step:110/1530 train_loss:4.4763 train_time:16001ms step_avg:160.01ms
step:111/1530 train_loss:4.4203 train_time:16160ms step_avg:160.00ms
step:112/1530 train_loss:4.5774 train_time:16320ms step_avg:160.00ms
step:113/1530 train_loss:4.4886 train_time:16481ms step_avg:160.01ms
step:114/1530 train_loss:4.3531 train_time:16642ms step_avg:160.02ms
step:115/1530 train_loss:4.4927 train_time:16805ms step_avg:160.05ms
step:116/1530 train_loss:4.4505 train_time:16969ms step_avg:160.08ms
step:117/1530 train_loss:4.3523 train_time:17132ms step_avg:160.12ms
step:118/1530 train_loss:4.5845 train_time:17297ms step_avg:160.16ms
step:119/1530 train_loss:4.4610 train_time:17462ms step_avg:160.20ms
step:120/1530 train_loss:4.3202 train_time:17626ms step_avg:160.23ms
step:121/1530 train_loss:4.2940 train_time:17790ms step_avg:160.27ms
step:122/1530 train_loss:4.4331 train_time:17953ms step_avg:160.30ms
step:123/1530 train_loss:4.2694 train_time:18117ms step_avg:160.33ms
step:124/1530 train_loss:4.5656 train_time:18282ms step_avg:160.37ms
step:125/1530 train_loss:4.4431 train_time:18447ms step_avg:160.41ms
step:125/1530 val_loss:4.4021 train_time:18494ms step_avg:160.81ms
step:126/1530 train_loss:4.4157 train_time:18612ms step_avg:160.45ms
step:127/1530 train_loss:4.4464 train_time:18778ms step_avg:160.49ms
step:128/1530 train_loss:4.3810 train_time:18942ms step_avg:160.53ms
step:129/1530 train_loss:4.6654 train_time:19106ms step_avg:160.55ms
step:130/1530 train_loss:4.3415 train_time:19270ms step_avg:160.59ms
step:131/1530 train_loss:4.3996 train_time:19433ms step_avg:160.61ms
step:132/1530 train_loss:4.3459 train_time:19598ms step_avg:160.64ms
step:133/1530 train_loss:4.4552 train_time:19761ms step_avg:160.65ms
step:134/1530 train_loss:4.2528 train_time:19926ms step_avg:160.69ms
step:135/1530 train_loss:4.4400 train_time:20090ms step_avg:160.72ms
step:136/1530 train_loss:4.2152 train_time:20254ms step_avg:160.74ms
step:137/1530 train_loss:4.3886 train_time:20418ms step_avg:160.77ms
step:138/1530 train_loss:4.2837 train_time:20582ms step_avg:160.80ms
step:139/1530 train_loss:4.3805 train_time:20746ms step_avg:160.82ms
step:140/1530 train_loss:4.4674 train_time:20910ms step_avg:160.85ms
step:141/1530 train_loss:4.3155 train_time:21074ms step_avg:160.87ms
step:142/1530 train_loss:4.2939 train_time:21238ms step_avg:160.89ms
step:143/1530 train_loss:4.2402 train_time:21402ms step_avg:160.91ms
step:144/1530 train_loss:4.3416 train_time:21566ms step_avg:160.94ms
step:145/1530 train_loss:4.3112 train_time:21729ms step_avg:160.96ms
step:146/1530 train_loss:4.1647 train_time:21894ms step_avg:160.98ms
step:147/1530 train_loss:4.3176 train_time:22058ms step_avg:161.01ms
step:148/1530 train_loss:4.3498 train_time:22221ms step_avg:161.02ms
step:149/1530 train_loss:4.2940 train_time:22385ms step_avg:161.05ms
step:150/1530 train_loss:4.4473 train_time:22549ms step_avg:161.07ms
step:151/1530 train_loss:4.2716 train_time:22713ms step_avg:161.09ms
step:152/1530 train_loss:4.2584 train_time:22878ms step_avg:161.11ms
step:153/1530 train_loss:4.3575 train_time:23042ms step_avg:161.13ms
step:154/1530 train_loss:4.3675 train_time:23206ms step_avg:161.15ms
step:155/1530 train_loss:4.2654 train_time:23370ms step_avg:161.17ms
step:156/1530 train_loss:4.3502 train_time:23533ms step_avg:161.18ms
step:157/1530 train_loss:4.4048 train_time:23697ms step_avg:161.20ms
step:158/1530 train_loss:4.2476 train_time:23861ms step_avg:161.23ms
step:159/1530 train_loss:4.3077 train_time:24025ms step_avg:161.24ms
step:160/1530 train_loss:4.1337 train_time:24189ms step_avg:161.26ms
step:161/1530 train_loss:4.3504 train_time:24354ms step_avg:161.28ms
step:162/1530 train_loss:4.3635 train_time:24517ms step_avg:161.30ms
step:163/1530 train_loss:4.3415 train_time:24681ms step_avg:161.31ms
step:164/1530 train_loss:4.1813 train_time:24845ms step_avg:161.33ms
step:165/1530 train_loss:4.2754 train_time:25008ms step_avg:161.34ms
step:166/1530 train_loss:4.3461 train_time:25173ms step_avg:161.36ms
step:167/1530 train_loss:4.2043 train_time:25336ms step_avg:161.38ms
step:168/1530 train_loss:4.2849 train_time:25500ms step_avg:161.39ms
step:169/1530 train_loss:4.1618 train_time:25665ms step_avg:161.41ms
step:170/1530 train_loss:4.0238 train_time:25828ms step_avg:161.43ms
step:171/1530 train_loss:4.2018 train_time:25992ms step_avg:161.44ms
step:172/1530 train_loss:4.2089 train_time:26156ms step_avg:161.45ms
step:173/1530 train_loss:4.2587 train_time:26321ms step_avg:161.48ms
step:174/1530 train_loss:4.4131 train_time:26483ms step_avg:161.48ms
step:175/1530 train_loss:4.2399 train_time:26649ms step_avg:161.51ms
step:176/1530 train_loss:4.0882 train_time:26813ms step_avg:161.52ms
step:177/1530 train_loss:4.0607 train_time:26975ms step_avg:161.53ms
step:178/1530 train_loss:4.1785 train_time:27139ms step_avg:161.54ms
step:179/1530 train_loss:4.1239 train_time:27301ms step_avg:161.55ms
step:180/1530 train_loss:4.1255 train_time:27464ms step_avg:161.55ms
step:181/1530 train_loss:4.2960 train_time:27627ms step_avg:161.56ms
step:182/1530 train_loss:4.1530 train_time:27791ms step_avg:161.58ms
step:183/1530 train_loss:4.1171 train_time:27954ms step_avg:161.59ms
step:184/1530 train_loss:4.1235 train_time:28117ms step_avg:161.59ms
step:185/1530 train_loss:4.1996 train_time:28281ms step_avg:161.60ms
step:186/1530 train_loss:4.1726 train_time:28443ms step_avg:161.61ms
step:187/1530 train_loss:4.2311 train_time:28606ms step_avg:161.62ms
step:188/1530 train_loss:4.1748 train_time:28908ms step_avg:162.41ms
step:189/1530 train_loss:4.1163 train_time:29248ms step_avg:163.40ms
step:190/1530 train_loss:4.2008 train_time:29421ms step_avg:163.45ms
step:191/1530 train_loss:4.0708 train_time:29585ms step_avg:163.45ms
step:192/1530 train_loss:4.0310 train_time:29748ms step_avg:163.45ms
step:193/1530 train_loss:4.2400 train_time:29912ms step_avg:163.45ms
step:194/1530 train_loss:4.1683 train_time:30075ms step_avg:163.45ms
step:195/1530 train_loss:4.3487 train_time:30237ms step_avg:163.44ms
step:196/1530 train_loss:4.1775 train_time:30399ms step_avg:163.43ms
step:197/1530 train_loss:4.0416 train_time:30563ms step_avg:163.44ms
step:198/1530 train_loss:4.1766 train_time:30727ms step_avg:163.44ms
step:199/1530 train_loss:4.0365 train_time:30890ms step_avg:163.44ms
step:200/1530 train_loss:4.1093 train_time:31053ms step_avg:163.44ms
step:201/1530 train_loss:3.9938 train_time:31217ms step_avg:163.44ms
step:202/1530 train_loss:4.2475 train_time:31379ms step_avg:163.43ms
step:203/1530 train_loss:4.0616 train_time:31543ms step_avg:163.44ms
step:204/1530 train_loss:4.1944 train_time:31706ms step_avg:163.43ms
step:205/1530 train_loss:4.2600 train_time:31869ms step_avg:163.43ms
step:206/1530 train_loss:3.9502 train_time:32032ms step_avg:163.43ms
step:207/1530 train_loss:4.0768 train_time:32195ms step_avg:163.43ms
step:208/1530 train_loss:4.0957 train_time:32358ms step_avg:163.43ms
step:209/1530 train_loss:4.2328 train_time:32520ms step_avg:163.42ms
step:210/1530 train_loss:4.1811 train_time:32684ms step_avg:163.42ms
step:211/1530 train_loss:4.0581 train_time:32846ms step_avg:163.41ms
step:212/1530 train_loss:4.1068 train_time:33009ms step_avg:163.41ms
step:213/1530 train_loss:4.0443 train_time:33172ms step_avg:163.41ms
step:214/1530 train_loss:4.1144 train_time:33334ms step_avg:163.40ms
step:215/1530 train_loss:3.9663 train_time:33498ms step_avg:163.40ms
step:216/1530 train_loss:4.0067 train_time:33660ms step_avg:163.40ms
step:217/1530 train_loss:4.0083 train_time:33823ms step_avg:163.39ms
step:218/1530 train_loss:4.0802 train_time:33984ms step_avg:163.39ms
step:219/1530 train_loss:4.0743 train_time:34148ms step_avg:163.39ms
step:220/1530 train_loss:4.0794 train_time:34311ms step_avg:163.39ms
step:221/1530 train_loss:4.0930 train_time:34474ms step_avg:163.38ms
step:222/1530 train_loss:3.9908 train_time:34637ms step_avg:163.38ms
step:223/1530 train_loss:3.9849 train_time:34799ms step_avg:163.38ms
step:224/1530 train_loss:4.2947 train_time:34961ms step_avg:163.37ms
step:225/1530 train_loss:3.9301 train_time:35124ms step_avg:163.37ms
step:226/1530 train_loss:3.9941 train_time:35288ms step_avg:163.37ms
step:227/1530 train_loss:3.9827 train_time:35450ms step_avg:163.37ms
step:228/1530 train_loss:4.1415 train_time:35616ms step_avg:163.38ms
step:229/1530 train_loss:3.9256 train_time:35784ms step_avg:163.40ms
step:230/1530 train_loss:4.0406 train_time:35949ms step_avg:163.40ms
step:231/1530 train_loss:3.8996 train_time:36115ms step_avg:163.41ms
step:232/1530 train_loss:3.9620 train_time:36279ms step_avg:163.42ms
step:233/1530 train_loss:4.0937 train_time:36446ms step_avg:163.44ms
step:234/1530 train_loss:4.0303 train_time:36612ms step_avg:163.45ms
step:235/1530 train_loss:3.9008 train_time:36779ms step_avg:163.46ms
step:236/1530 train_loss:4.0740 train_time:36946ms step_avg:163.48ms
step:237/1530 train_loss:4.0828 train_time:37113ms step_avg:163.49ms
step:238/1530 train_loss:3.9362 train_time:37280ms step_avg:163.51ms
step:239/1530 train_loss:4.0757 train_time:37446ms step_avg:163.52ms
step:240/1530 train_loss:4.1131 train_time:37612ms step_avg:163.53ms
step:241/1530 train_loss:3.9716 train_time:37780ms step_avg:163.55ms
step:242/1530 train_loss:4.1476 train_time:37947ms step_avg:163.56ms
step:243/1530 train_loss:4.0127 train_time:38112ms step_avg:163.57ms
step:244/1530 train_loss:4.0813 train_time:38279ms step_avg:163.59ms
step:245/1530 train_loss:4.1317 train_time:38446ms step_avg:163.60ms
step:246/1530 train_loss:4.0617 train_time:38612ms step_avg:163.61ms
step:247/1530 train_loss:4.0083 train_time:38779ms step_avg:163.63ms
step:248/1530 train_loss:4.1004 train_time:38947ms step_avg:163.64ms
step:249/1530 train_loss:3.9228 train_time:39113ms step_avg:163.65ms
step:250/1530 train_loss:3.9725 train_time:39279ms step_avg:163.66ms
step:250/1530 val_loss:4.0051 train_time:39327ms step_avg:163.86ms
step:251/1530 train_loss:4.0769 train_time:39447ms step_avg:163.68ms
step:252/1530 train_loss:4.1627 train_time:39614ms step_avg:163.70ms
step:253/1530 train_loss:3.9344 train_time:39780ms step_avg:163.71ms
step:254/1530 train_loss:3.8827 train_time:39946ms step_avg:163.71ms
step:255/1530 train_loss:4.0832 train_time:40113ms step_avg:163.73ms
step:256/1530 train_loss:3.9881 train_time:40279ms step_avg:163.73ms
step:257/1530 train_loss:3.9886 train_time:40444ms step_avg:163.74ms
step:258/1530 train_loss:3.9775 train_time:40612ms step_avg:163.76ms
step:259/1530 train_loss:4.0330 train_time:40777ms step_avg:163.76ms
step:260/1530 train_loss:4.0605 train_time:40945ms step_avg:163.78ms
step:261/1530 train_loss:4.0182 train_time:41111ms step_avg:163.79ms
step:262/1530 train_loss:3.9959 train_time:41276ms step_avg:163.80ms
step:263/1530 train_loss:3.8885 train_time:41443ms step_avg:163.81ms
step:264/1530 train_loss:3.9794 train_time:41609ms step_avg:163.81ms
step:265/1530 train_loss:3.8691 train_time:41775ms step_avg:163.83ms
step:266/1530 train_loss:3.9243 train_time:41942ms step_avg:163.83ms
step:267/1530 train_loss:3.9298 train_time:42110ms step_avg:163.85ms
step:268/1530 train_loss:3.9551 train_time:42275ms step_avg:163.86ms
step:269/1530 train_loss:3.8437 train_time:42441ms step_avg:163.87ms
step:270/1530 train_loss:4.1064 train_time:42607ms step_avg:163.87ms
step:271/1530 train_loss:3.9661 train_time:42773ms step_avg:163.88ms
step:272/1530 train_loss:3.9168 train_time:42939ms step_avg:163.89ms
step:273/1530 train_loss:3.9339 train_time:43104ms step_avg:163.89ms
step:274/1530 train_loss:4.0445 train_time:43271ms step_avg:163.91ms
step:275/1530 train_loss:4.0623 train_time:43438ms step_avg:163.92ms
step:276/1530 train_loss:4.2231 train_time:43605ms step_avg:163.93ms
step:277/1530 train_loss:4.0351 train_time:43771ms step_avg:163.94ms
step:278/1530 train_loss:4.0871 train_time:43938ms step_avg:163.95ms
step:279/1530 train_loss:3.9987 train_time:44104ms step_avg:163.95ms
step:280/1530 train_loss:4.1833 train_time:44271ms step_avg:163.97ms
step:281/1530 train_loss:3.9714 train_time:44437ms step_avg:163.97ms
step:282/1530 train_loss:3.9407 train_time:44604ms step_avg:163.98ms
step:283/1530 train_loss:3.9095 train_time:44770ms step_avg:163.99ms
step:284/1530 train_loss:4.0511 train_time:44936ms step_avg:164.00ms
step:285/1530 train_loss:4.0550 train_time:45101ms step_avg:164.00ms
step:286/1530 train_loss:4.0890 train_time:45266ms step_avg:164.01ms
step:287/1530 train_loss:3.9072 train_time:45432ms step_avg:164.02ms
step:288/1530 train_loss:4.0086 train_time:45597ms step_avg:164.02ms
step:289/1530 train_loss:3.8703 train_time:45761ms step_avg:164.02ms
step:290/1530 train_loss:3.8603 train_time:45927ms step_avg:164.03ms
step:291/1530 train_loss:3.9102 train_time:46093ms step_avg:164.03ms
step:292/1530 train_loss:3.8609 train_time:46258ms step_avg:164.04ms
step:293/1530 train_loss:3.8939 train_time:46423ms step_avg:164.04ms
step:294/1530 train_loss:3.9309 train_time:46590ms step_avg:164.05ms
step:295/1530 train_loss:3.8418 train_time:46755ms step_avg:164.05ms
step:296/1530 train_loss:3.8687 train_time:46921ms step_avg:164.06ms
step:297/1530 train_loss:3.8813 train_time:47087ms step_avg:164.06ms
step:298/1530 train_loss:3.9723 train_time:47252ms step_avg:164.07ms
step:299/1530 train_loss:3.8247 train_time:47417ms step_avg:164.07ms
step:300/1530 train_loss:3.9674 train_time:47582ms step_avg:164.08ms
step:301/1530 train_loss:3.9614 train_time:47750ms step_avg:164.09ms
step:302/1530 train_loss:3.9285 train_time:47916ms step_avg:164.09ms
step:303/1530 train_loss:3.9668 train_time:48080ms step_avg:164.10ms
step:304/1530 train_loss:3.9574 train_time:48245ms step_avg:164.10ms
step:305/1530 train_loss:4.4523 train_time:48410ms step_avg:164.10ms
step:306/1530 train_loss:3.9418 train_time:48575ms step_avg:164.11ms
step:307/1530 train_loss:3.8348 train_time:48741ms step_avg:164.11ms
step:308/1530 train_loss:3.9742 train_time:48905ms step_avg:164.11ms
step:309/1530 train_loss:3.8624 train_time:49072ms step_avg:164.12ms
step:310/1530 train_loss:4.0766 train_time:49237ms step_avg:164.12ms
step:311/1530 train_loss:3.9286 train_time:49401ms step_avg:164.12ms
step:312/1530 train_loss:3.8597 train_time:49567ms step_avg:164.13ms
step:313/1530 train_loss:3.9364 train_time:49733ms step_avg:164.14ms
step:314/1530 train_loss:4.0585 train_time:49898ms step_avg:164.14ms
step:315/1530 train_loss:3.9412 train_time:50063ms step_avg:164.14ms
step:316/1530 train_loss:3.7894 train_time:50230ms step_avg:164.15ms
step:317/1530 train_loss:3.8778 train_time:50395ms step_avg:164.15ms
step:318/1530 train_loss:3.9229 train_time:50561ms step_avg:164.16ms
step:319/1530 train_loss:3.8890 train_time:50725ms step_avg:164.16ms
step:320/1530 train_loss:4.0118 train_time:50892ms step_avg:164.17ms
step:321/1530 train_loss:3.9530 train_time:51057ms step_avg:164.17ms
step:322/1530 train_loss:3.9342 train_time:51223ms step_avg:164.18ms
step:323/1530 train_loss:4.0053 train_time:51389ms step_avg:164.18ms
step:324/1530 train_loss:3.9527 train_time:51555ms step_avg:164.19ms
step:325/1530 train_loss:4.0131 train_time:51720ms step_avg:164.19ms
step:326/1530 train_loss:3.8941 train_time:51887ms step_avg:164.20ms
step:327/1530 train_loss:4.3963 train_time:52053ms step_avg:164.20ms
step:328/1530 train_loss:4.0716 train_time:52218ms step_avg:164.21ms
step:329/1530 train_loss:3.7886 train_time:52383ms step_avg:164.21ms
step:330/1530 train_loss:3.7435 train_time:52549ms step_avg:164.22ms
step:331/1530 train_loss:3.9760 train_time:52714ms step_avg:164.22ms
step:332/1530 train_loss:3.9032 train_time:52879ms step_avg:164.22ms
step:333/1530 train_loss:3.8845 train_time:53045ms step_avg:164.22ms
step:334/1530 train_loss:3.8441 train_time:53211ms step_avg:164.23ms
step:335/1530 train_loss:4.0097 train_time:53376ms step_avg:164.23ms
step:336/1530 train_loss:3.9534 train_time:53542ms step_avg:164.24ms
step:337/1530 train_loss:4.4120 train_time:53707ms step_avg:164.24ms
step:338/1530 train_loss:3.9367 train_time:53873ms step_avg:164.25ms
step:339/1530 train_loss:3.8662 train_time:54038ms step_avg:164.25ms
step:340/1530 train_loss:3.9370 train_time:54203ms step_avg:164.25ms
step:341/1530 train_loss:3.8600 train_time:54369ms step_avg:164.26ms
step:342/1530 train_loss:3.8081 train_time:54536ms step_avg:164.27ms
step:343/1530 train_loss:3.8408 train_time:54704ms step_avg:164.28ms
step:344/1530 train_loss:3.9902 train_time:54872ms step_avg:164.29ms
step:345/1530 train_loss:3.8189 train_time:55041ms step_avg:164.30ms
step:346/1530 train_loss:3.7628 train_time:55209ms step_avg:164.31ms
step:347/1530 train_loss:3.8011 train_time:55377ms step_avg:164.32ms
step:348/1530 train_loss:3.8598 train_time:55545ms step_avg:164.33ms
step:349/1530 train_loss:3.8252 train_time:55713ms step_avg:164.35ms
step:350/1530 train_loss:3.5687 train_time:55881ms step_avg:164.36ms
step:351/1530 train_loss:3.8306 train_time:56048ms step_avg:164.36ms
step:352/1530 train_loss:4.1730 train_time:56216ms step_avg:164.37ms
step:353/1530 train_loss:3.6504 train_time:56384ms step_avg:164.38ms
step:354/1530 train_loss:3.9262 train_time:56552ms step_avg:164.39ms
step:355/1530 train_loss:3.7874 train_time:56721ms step_avg:164.41ms
step:356/1530 train_loss:3.8812 train_time:56891ms step_avg:164.43ms
step:357/1530 train_loss:3.7649 train_time:57060ms step_avg:164.44ms
step:358/1530 train_loss:3.8596 train_time:57229ms step_avg:164.45ms
step:359/1530 train_loss:3.7607 train_time:57398ms step_avg:164.46ms
step:360/1530 train_loss:3.4289 train_time:57567ms step_avg:164.48ms
step:361/1530 train_loss:4.0205 train_time:57736ms step_avg:164.49ms
step:362/1530 train_loss:3.9184 train_time:57904ms step_avg:164.50ms
step:363/1530 train_loss:3.8372 train_time:58072ms step_avg:164.51ms
step:364/1530 train_loss:3.7398 train_time:58241ms step_avg:164.52ms
step:365/1530 train_loss:3.9130 train_time:58409ms step_avg:164.53ms
step:366/1530 train_loss:3.8601 train_time:58577ms step_avg:164.54ms
step:367/1530 train_loss:3.8548 train_time:58746ms step_avg:164.55ms
step:368/1530 train_loss:3.8454 train_time:58913ms step_avg:164.56ms
step:369/1530 train_loss:3.7376 train_time:59081ms step_avg:164.57ms
step:370/1530 train_loss:3.8754 train_time:59251ms step_avg:164.59ms
step:371/1530 train_loss:3.7302 train_time:59418ms step_avg:164.59ms
step:372/1530 train_loss:3.6915 train_time:59587ms step_avg:164.60ms
step:373/1530 train_loss:3.9125 train_time:59756ms step_avg:164.62ms
step:374/1530 train_loss:3.8277 train_time:59923ms step_avg:164.62ms
step:375/1530 train_loss:3.8068 train_time:60092ms step_avg:164.63ms
step:375/1530 val_loss:3.8280 train_time:60140ms step_avg:164.77ms