-
Notifications
You must be signed in to change notification settings - Fork 183
/
f0f173e3-69ee-4970-ba0b-f7f3e5d92e33.txt
2165 lines (2092 loc) · 134 KB
/
f0f173e3-69ee-4970-ba0b-f7f3e5d92e33.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 01:41:49 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 119W / 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 38C P0 122W / 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 127W / 700W | 529MiB / 81559MiB | 0% 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:31928ms step_avg:nanms
step:2/1530 train_loss:10.0748 train_time:32040ms step_avg:nanms
step:3/1530 train_loss:8.4025 train_time:32202ms step_avg:nanms
step:4/1530 train_loss:7.6038 train_time:32363ms step_avg:nanms
step:5/1530 train_loss:7.4218 train_time:32523ms step_avg:nanms
step:6/1530 train_loss:6.9788 train_time:32684ms step_avg:nanms
step:7/1530 train_loss:7.1986 train_time:32845ms step_avg:nanms
step:8/1530 train_loss:6.7363 train_time:33006ms step_avg:nanms
step:9/1530 train_loss:6.6368 train_time:33167ms step_avg:nanms
step:10/1530 train_loss:6.5188 train_time:33328ms step_avg:nanms
step:11/1530 train_loss:6.5013 train_time:114ms step_avg:nanms
step:12/1530 train_loss:6.3350 train_time:274ms step_avg:nanms
step:13/1530 train_loss:6.2568 train_time:435ms step_avg:144.87ms
step:14/1530 train_loss:6.1921 train_time:595ms step_avg:148.70ms
step:15/1530 train_loss:6.1662 train_time:754ms step_avg:150.86ms
step:16/1530 train_loss:6.0887 train_time:915ms step_avg:152.53ms
step:17/1530 train_loss:6.1591 train_time:1076ms step_avg:153.65ms
step:18/1530 train_loss:5.9709 train_time:1236ms step_avg:154.52ms
step:19/1530 train_loss:5.9771 train_time:1397ms step_avg:155.19ms
step:20/1530 train_loss:5.6572 train_time:1556ms step_avg:155.63ms
step:21/1530 train_loss:5.9969 train_time:1716ms step_avg:155.97ms
step:22/1530 train_loss:6.1916 train_time:1875ms step_avg:156.27ms
step:23/1530 train_loss:5.8676 train_time:2036ms step_avg:156.62ms
step:24/1530 train_loss:6.0291 train_time:2196ms step_avg:156.87ms
step:25/1530 train_loss:5.7017 train_time:2357ms step_avg:157.11ms
step:26/1530 train_loss:5.5967 train_time:2517ms step_avg:157.31ms
step:27/1530 train_loss:5.8027 train_time:2677ms step_avg:157.48ms
step:28/1530 train_loss:5.3959 train_time:2838ms step_avg:157.67ms
step:29/1530 train_loss:5.6790 train_time:2998ms step_avg:157.77ms
step:30/1530 train_loss:5.4678 train_time:3159ms step_avg:157.94ms
step:31/1530 train_loss:5.4373 train_time:3320ms step_avg:158.08ms
step:32/1530 train_loss:5.2885 train_time:3479ms step_avg:158.15ms
step:33/1530 train_loss:5.5968 train_time:3638ms step_avg:158.17ms
step:34/1530 train_loss:5.4940 train_time:3799ms step_avg:158.30ms
step:35/1530 train_loss:5.6297 train_time:3960ms step_avg:158.39ms
step:36/1530 train_loss:5.5495 train_time:4118ms step_avg:158.40ms
step:37/1530 train_loss:5.4513 train_time:4279ms step_avg:158.47ms
step:38/1530 train_loss:5.2996 train_time:4439ms step_avg:158.54ms
step:39/1530 train_loss:5.3374 train_time:4600ms step_avg:158.61ms
step:40/1530 train_loss:5.2579 train_time:4760ms step_avg:158.67ms
step:41/1530 train_loss:5.2393 train_time:4919ms step_avg:158.68ms
step:42/1530 train_loss:5.1672 train_time:5080ms step_avg:158.75ms
step:43/1530 train_loss:5.2557 train_time:5238ms step_avg:158.73ms
step:44/1530 train_loss:5.2382 train_time:5399ms step_avg:158.78ms
step:45/1530 train_loss:5.3829 train_time:5558ms step_avg:158.81ms
step:46/1530 train_loss:5.1805 train_time:5718ms step_avg:158.83ms
step:47/1530 train_loss:5.0834 train_time:5879ms step_avg:158.90ms
step:48/1530 train_loss:5.2177 train_time:6040ms step_avg:158.95ms
step:49/1530 train_loss:5.1482 train_time:6201ms step_avg:159.00ms
step:50/1530 train_loss:5.2600 train_time:6360ms step_avg:159.00ms
step:51/1530 train_loss:5.1386 train_time:6520ms step_avg:159.02ms
step:52/1530 train_loss:5.0207 train_time:6680ms step_avg:159.04ms
step:53/1530 train_loss:5.1730 train_time:6839ms step_avg:159.05ms
step:54/1530 train_loss:5.0218 train_time:7000ms step_avg:159.10ms
step:55/1530 train_loss:5.4176 train_time:7160ms step_avg:159.11ms
step:56/1530 train_loss:5.0406 train_time:7319ms step_avg:159.12ms
step:57/1530 train_loss:4.8917 train_time:7478ms step_avg:159.10ms
step:58/1530 train_loss:5.0435 train_time:7638ms step_avg:159.12ms
step:59/1530 train_loss:5.0173 train_time:7799ms step_avg:159.17ms
step:60/1530 train_loss:5.1406 train_time:7959ms step_avg:159.19ms
step:61/1530 train_loss:4.8572 train_time:8119ms step_avg:159.19ms
step:62/1530 train_loss:4.9753 train_time:8280ms step_avg:159.23ms
step:63/1530 train_loss:4.9788 train_time:8441ms step_avg:159.26ms
step:64/1530 train_loss:4.8849 train_time:8600ms step_avg:159.26ms
step:65/1530 train_loss:4.7885 train_time:8761ms step_avg:159.29ms
step:66/1530 train_loss:4.9236 train_time:8921ms step_avg:159.30ms
step:67/1530 train_loss:4.8203 train_time:9080ms step_avg:159.30ms
step:68/1530 train_loss:5.0880 train_time:9241ms step_avg:159.33ms
step:69/1530 train_loss:4.7288 train_time:9402ms step_avg:159.36ms
step:70/1530 train_loss:4.8570 train_time:9563ms step_avg:159.38ms
step:71/1530 train_loss:4.9845 train_time:9722ms step_avg:159.38ms
step:72/1530 train_loss:4.9076 train_time:9883ms step_avg:159.41ms
step:73/1530 train_loss:4.7911 train_time:10045ms step_avg:159.44ms
step:74/1530 train_loss:4.9320 train_time:10208ms step_avg:159.50ms
step:75/1530 train_loss:4.8881 train_time:10369ms step_avg:159.52ms
step:76/1530 train_loss:4.8015 train_time:10529ms step_avg:159.53ms
step:77/1530 train_loss:4.9267 train_time:10689ms step_avg:159.54ms
step:78/1530 train_loss:5.1224 train_time:10849ms step_avg:159.54ms
step:79/1530 train_loss:4.8204 train_time:11010ms step_avg:159.57ms
step:80/1530 train_loss:4.8805 train_time:11171ms step_avg:159.59ms
step:81/1530 train_loss:4.6614 train_time:11331ms step_avg:159.60ms
step:82/1530 train_loss:4.8269 train_time:11493ms step_avg:159.62ms
step:83/1530 train_loss:4.7762 train_time:11653ms step_avg:159.63ms
step:84/1530 train_loss:4.7706 train_time:11814ms step_avg:159.65ms
step:85/1530 train_loss:4.6315 train_time:11974ms step_avg:159.66ms
step:86/1530 train_loss:4.8443 train_time:12134ms step_avg:159.66ms
step:87/1530 train_loss:4.7398 train_time:12295ms step_avg:159.68ms
step:88/1530 train_loss:4.7511 train_time:12456ms step_avg:159.69ms
step:89/1530 train_loss:4.7116 train_time:12616ms step_avg:159.69ms
step:90/1530 train_loss:4.6469 train_time:12776ms step_avg:159.70ms
step:91/1530 train_loss:4.6465 train_time:12936ms step_avg:159.70ms
step:92/1530 train_loss:4.8138 train_time:13096ms step_avg:159.71ms
step:93/1530 train_loss:4.6347 train_time:13257ms step_avg:159.72ms
step:94/1530 train_loss:4.6552 train_time:13416ms step_avg:159.72ms
step:95/1530 train_loss:4.6947 train_time:13576ms step_avg:159.72ms
step:96/1530 train_loss:4.5831 train_time:13736ms step_avg:159.72ms
step:97/1530 train_loss:4.6530 train_time:13896ms step_avg:159.73ms
step:98/1530 train_loss:4.5926 train_time:14056ms step_avg:159.72ms
step:99/1530 train_loss:4.6806 train_time:14217ms step_avg:159.74ms
step:100/1530 train_loss:4.6866 train_time:14378ms step_avg:159.76ms
step:101/1530 train_loss:4.5388 train_time:14538ms step_avg:159.76ms
step:102/1530 train_loss:4.7194 train_time:14699ms step_avg:159.77ms
step:103/1530 train_loss:4.5978 train_time:14858ms step_avg:159.77ms
step:104/1530 train_loss:4.5343 train_time:15018ms step_avg:159.77ms
step:105/1530 train_loss:4.5649 train_time:15177ms step_avg:159.76ms
step:106/1530 train_loss:4.6624 train_time:15337ms step_avg:159.76ms
step:107/1530 train_loss:4.5268 train_time:15498ms step_avg:159.77ms
step:108/1530 train_loss:4.3652 train_time:15658ms step_avg:159.77ms
step:109/1530 train_loss:4.4969 train_time:15818ms step_avg:159.78ms
step:110/1530 train_loss:4.5012 train_time:15979ms step_avg:159.79ms
step:111/1530 train_loss:4.4325 train_time:16139ms step_avg:159.79ms
step:112/1530 train_loss:4.5948 train_time:16298ms step_avg:159.78ms
step:113/1530 train_loss:4.5027 train_time:16459ms step_avg:159.79ms
step:114/1530 train_loss:4.3743 train_time:16620ms step_avg:159.81ms
step:115/1530 train_loss:4.5167 train_time:16782ms step_avg:159.83ms
step:116/1530 train_loss:4.4789 train_time:16946ms step_avg:159.87ms
step:117/1530 train_loss:4.3793 train_time:17111ms step_avg:159.91ms
step:118/1530 train_loss:4.5945 train_time:17275ms step_avg:159.95ms
step:119/1530 train_loss:4.4693 train_time:17439ms step_avg:159.99ms
step:120/1530 train_loss:4.3537 train_time:17604ms step_avg:160.03ms
step:121/1530 train_loss:4.3071 train_time:17767ms step_avg:160.06ms
step:122/1530 train_loss:4.4571 train_time:17931ms step_avg:160.10ms
step:123/1530 train_loss:4.2954 train_time:18095ms step_avg:160.13ms
step:124/1530 train_loss:4.5970 train_time:18258ms step_avg:160.16ms
step:125/1530 train_loss:4.4726 train_time:18423ms step_avg:160.20ms
step:125/1530 val_loss:4.4195 train_time:18470ms step_avg:160.61ms
step:126/1530 train_loss:4.4293 train_time:18587ms step_avg:160.23ms
step:127/1530 train_loss:4.4514 train_time:18754ms step_avg:160.29ms
step:128/1530 train_loss:4.3723 train_time:18917ms step_avg:160.32ms
step:129/1530 train_loss:4.6860 train_time:19082ms step_avg:160.36ms
step:130/1530 train_loss:4.3737 train_time:19246ms step_avg:160.39ms
step:131/1530 train_loss:4.4071 train_time:19409ms step_avg:160.41ms
step:132/1530 train_loss:4.3541 train_time:19575ms step_avg:160.45ms
step:133/1530 train_loss:4.4507 train_time:19740ms step_avg:160.49ms
step:134/1530 train_loss:4.2756 train_time:19904ms step_avg:160.51ms
step:135/1530 train_loss:4.4625 train_time:20069ms step_avg:160.56ms
step:136/1530 train_loss:4.2158 train_time:20235ms step_avg:160.60ms
step:137/1530 train_loss:4.3849 train_time:20400ms step_avg:160.63ms
step:138/1530 train_loss:4.2872 train_time:20564ms step_avg:160.65ms
step:139/1530 train_loss:4.3914 train_time:20727ms step_avg:160.68ms
step:140/1530 train_loss:4.4834 train_time:20892ms step_avg:160.70ms
step:141/1530 train_loss:4.3233 train_time:21056ms step_avg:160.73ms
step:142/1530 train_loss:4.3323 train_time:21219ms step_avg:160.75ms
step:143/1530 train_loss:4.2668 train_time:21384ms step_avg:160.78ms
step:144/1530 train_loss:4.3690 train_time:21548ms step_avg:160.81ms
step:145/1530 train_loss:4.3287 train_time:21712ms step_avg:160.83ms
step:146/1530 train_loss:4.1777 train_time:21877ms step_avg:160.86ms
step:147/1530 train_loss:4.3381 train_time:22041ms step_avg:160.88ms
step:148/1530 train_loss:4.3753 train_time:22205ms step_avg:160.91ms
step:149/1530 train_loss:4.3128 train_time:22370ms step_avg:160.93ms
step:150/1530 train_loss:4.4425 train_time:22535ms step_avg:160.96ms
step:151/1530 train_loss:4.2797 train_time:22699ms step_avg:160.98ms
step:152/1530 train_loss:4.2811 train_time:22863ms step_avg:161.00ms
step:153/1530 train_loss:4.3650 train_time:23026ms step_avg:161.02ms
step:154/1530 train_loss:4.3762 train_time:23189ms step_avg:161.03ms
step:155/1530 train_loss:4.2800 train_time:23354ms step_avg:161.06ms
step:156/1530 train_loss:4.3711 train_time:23518ms step_avg:161.09ms
step:157/1530 train_loss:4.4249 train_time:23683ms step_avg:161.11ms
step:158/1530 train_loss:4.2570 train_time:23847ms step_avg:161.13ms
step:159/1530 train_loss:4.3296 train_time:24012ms step_avg:161.15ms
step:160/1530 train_loss:4.1445 train_time:24178ms step_avg:161.18ms
step:161/1530 train_loss:4.3517 train_time:24341ms step_avg:161.20ms
step:162/1530 train_loss:4.3556 train_time:24506ms step_avg:161.22ms
step:163/1530 train_loss:4.3481 train_time:24671ms step_avg:161.25ms
step:164/1530 train_loss:4.1849 train_time:24835ms step_avg:161.27ms
step:165/1530 train_loss:4.2879 train_time:24999ms step_avg:161.28ms
step:166/1530 train_loss:4.3579 train_time:25163ms step_avg:161.30ms
step:167/1530 train_loss:4.2215 train_time:25326ms step_avg:161.31ms
step:168/1530 train_loss:4.2862 train_time:25489ms step_avg:161.33ms
step:169/1530 train_loss:4.1655 train_time:25654ms step_avg:161.35ms
step:170/1530 train_loss:4.0246 train_time:25819ms step_avg:161.37ms
step:171/1530 train_loss:4.2145 train_time:25982ms step_avg:161.38ms
step:172/1530 train_loss:4.2161 train_time:26145ms step_avg:161.39ms
step:173/1530 train_loss:4.2862 train_time:26307ms step_avg:161.40ms
step:174/1530 train_loss:4.4302 train_time:26471ms step_avg:161.41ms
step:175/1530 train_loss:4.2462 train_time:26634ms step_avg:161.42ms
step:176/1530 train_loss:4.0912 train_time:26797ms step_avg:161.43ms
step:177/1530 train_loss:4.0624 train_time:26960ms step_avg:161.44ms
step:178/1530 train_loss:4.1920 train_time:27122ms step_avg:161.44ms
step:179/1530 train_loss:4.1380 train_time:27285ms step_avg:161.45ms
step:180/1530 train_loss:4.1234 train_time:27448ms step_avg:161.46ms
step:181/1530 train_loss:4.2996 train_time:27609ms step_avg:161.45ms
step:182/1530 train_loss:4.1492 train_time:27773ms step_avg:161.47ms
step:183/1530 train_loss:4.1360 train_time:27936ms step_avg:161.48ms
step:184/1530 train_loss:4.1271 train_time:28099ms step_avg:161.49ms
step:185/1530 train_loss:4.2143 train_time:28262ms step_avg:161.50ms
step:186/1530 train_loss:4.1824 train_time:28425ms step_avg:161.51ms
step:187/1530 train_loss:4.2451 train_time:28588ms step_avg:161.51ms
step:188/1530 train_loss:4.1761 train_time:28887ms step_avg:162.29ms
step:189/1530 train_loss:4.1190 train_time:29219ms step_avg:163.23ms
step:190/1530 train_loss:4.2178 train_time:29381ms step_avg:163.23ms
step:191/1530 train_loss:4.0878 train_time:29545ms step_avg:163.23ms
step:192/1530 train_loss:4.0376 train_time:29707ms step_avg:163.22ms
step:193/1530 train_loss:4.2535 train_time:29870ms step_avg:163.22ms
step:194/1530 train_loss:4.1904 train_time:30033ms step_avg:163.22ms
step:195/1530 train_loss:4.3617 train_time:30196ms step_avg:163.22ms
step:196/1530 train_loss:4.1834 train_time:30360ms step_avg:163.23ms
step:197/1530 train_loss:4.0456 train_time:30523ms step_avg:163.22ms
step:198/1530 train_loss:4.1750 train_time:30686ms step_avg:163.22ms
step:199/1530 train_loss:4.0303 train_time:30849ms step_avg:163.22ms
step:200/1530 train_loss:4.1135 train_time:31013ms step_avg:163.22ms
step:201/1530 train_loss:4.0157 train_time:31175ms step_avg:163.22ms
step:202/1530 train_loss:4.2647 train_time:31338ms step_avg:163.22ms
step:203/1530 train_loss:4.0765 train_time:31501ms step_avg:163.22ms
step:204/1530 train_loss:4.1967 train_time:31665ms step_avg:163.22ms
step:205/1530 train_loss:4.2529 train_time:31827ms step_avg:163.22ms
step:206/1530 train_loss:3.9463 train_time:31989ms step_avg:163.21ms
step:207/1530 train_loss:4.0821 train_time:32151ms step_avg:163.20ms
step:208/1530 train_loss:4.1108 train_time:32315ms step_avg:163.20ms
step:209/1530 train_loss:4.2444 train_time:32478ms step_avg:163.21ms
step:210/1530 train_loss:4.1766 train_time:32641ms step_avg:163.21ms
step:211/1530 train_loss:4.0667 train_time:32804ms step_avg:163.20ms
step:212/1530 train_loss:4.1309 train_time:32967ms step_avg:163.20ms
step:213/1530 train_loss:4.0567 train_time:33128ms step_avg:163.19ms
step:214/1530 train_loss:4.1228 train_time:33291ms step_avg:163.19ms
step:215/1530 train_loss:3.9798 train_time:33454ms step_avg:163.19ms
step:216/1530 train_loss:4.0019 train_time:33616ms step_avg:163.19ms
step:217/1530 train_loss:4.0170 train_time:33780ms step_avg:163.19ms
step:218/1530 train_loss:4.0845 train_time:33943ms step_avg:163.19ms
step:219/1530 train_loss:4.0734 train_time:34105ms step_avg:163.18ms
step:220/1530 train_loss:4.0872 train_time:34269ms step_avg:163.19ms
step:221/1530 train_loss:4.0943 train_time:34431ms step_avg:163.18ms
step:222/1530 train_loss:3.9986 train_time:34593ms step_avg:163.18ms
step:223/1530 train_loss:3.9946 train_time:34757ms step_avg:163.18ms
step:224/1530 train_loss:4.3091 train_time:34919ms step_avg:163.17ms
step:225/1530 train_loss:3.9214 train_time:35082ms step_avg:163.17ms
step:226/1530 train_loss:3.9870 train_time:35245ms step_avg:163.17ms
step:227/1530 train_loss:3.9727 train_time:35407ms step_avg:163.17ms
step:228/1530 train_loss:4.1415 train_time:35573ms step_avg:163.18ms
step:229/1530 train_loss:3.9291 train_time:35738ms step_avg:163.19ms
step:230/1530 train_loss:4.0447 train_time:35904ms step_avg:163.20ms
step:231/1530 train_loss:3.9057 train_time:36070ms step_avg:163.21ms
step:232/1530 train_loss:3.9635 train_time:36236ms step_avg:163.23ms
step:233/1530 train_loss:4.0946 train_time:36402ms step_avg:163.24ms
step:234/1530 train_loss:4.0333 train_time:36568ms step_avg:163.25ms
step:235/1530 train_loss:3.9100 train_time:36736ms step_avg:163.27ms
step:236/1530 train_loss:4.0822 train_time:36902ms step_avg:163.28ms
step:237/1530 train_loss:4.0780 train_time:37068ms step_avg:163.30ms
step:238/1530 train_loss:3.9464 train_time:37236ms step_avg:163.31ms
step:239/1530 train_loss:4.0869 train_time:37402ms step_avg:163.33ms
step:240/1530 train_loss:4.1218 train_time:37568ms step_avg:163.34ms
step:241/1530 train_loss:3.9617 train_time:37733ms step_avg:163.35ms
step:242/1530 train_loss:4.1463 train_time:37900ms step_avg:163.36ms
step:243/1530 train_loss:4.0057 train_time:38066ms step_avg:163.37ms
step:244/1530 train_loss:4.0910 train_time:38232ms step_avg:163.39ms
step:245/1530 train_loss:4.1385 train_time:38398ms step_avg:163.40ms
step:246/1530 train_loss:4.0595 train_time:38564ms step_avg:163.41ms
step:247/1530 train_loss:4.0123 train_time:38730ms step_avg:163.42ms
step:248/1530 train_loss:4.1102 train_time:38896ms step_avg:163.43ms
step:249/1530 train_loss:3.9160 train_time:39063ms step_avg:163.44ms
step:250/1530 train_loss:3.9702 train_time:39228ms step_avg:163.45ms
step:250/1530 val_loss:4.0043 train_time:39276ms step_avg:163.65ms
step:251/1530 train_loss:4.0718 train_time:39398ms step_avg:163.48ms
step:252/1530 train_loss:4.1626 train_time:39565ms step_avg:163.49ms
step:253/1530 train_loss:3.9262 train_time:39731ms step_avg:163.50ms
step:254/1530 train_loss:3.8804 train_time:39898ms step_avg:163.52ms
step:255/1530 train_loss:4.0823 train_time:40064ms step_avg:163.53ms
step:256/1530 train_loss:4.0013 train_time:40230ms step_avg:163.54ms
step:257/1530 train_loss:3.9928 train_time:40396ms step_avg:163.55ms
step:258/1530 train_loss:3.9920 train_time:40562ms step_avg:163.56ms
step:259/1530 train_loss:4.0331 train_time:40729ms step_avg:163.57ms
step:260/1530 train_loss:4.0549 train_time:40896ms step_avg:163.58ms
step:261/1530 train_loss:4.0209 train_time:41063ms step_avg:163.60ms
step:262/1530 train_loss:3.9956 train_time:41229ms step_avg:163.61ms
step:263/1530 train_loss:3.8935 train_time:41396ms step_avg:163.62ms
step:264/1530 train_loss:3.9864 train_time:41562ms step_avg:163.63ms
step:265/1530 train_loss:3.8649 train_time:41728ms step_avg:163.64ms
step:266/1530 train_loss:3.9123 train_time:41894ms step_avg:163.65ms
step:267/1530 train_loss:3.9313 train_time:42060ms step_avg:163.66ms
step:268/1530 train_loss:3.9568 train_time:42225ms step_avg:163.66ms
step:269/1530 train_loss:3.8546 train_time:42390ms step_avg:163.67ms
step:270/1530 train_loss:4.1011 train_time:42558ms step_avg:163.68ms
step:271/1530 train_loss:3.9735 train_time:42724ms step_avg:163.69ms
step:272/1530 train_loss:3.9308 train_time:42890ms step_avg:163.70ms
step:273/1530 train_loss:3.9456 train_time:43055ms step_avg:163.71ms
step:274/1530 train_loss:4.0448 train_time:43222ms step_avg:163.72ms
step:275/1530 train_loss:4.0676 train_time:43388ms step_avg:163.73ms
step:276/1530 train_loss:4.2340 train_time:43554ms step_avg:163.74ms
step:277/1530 train_loss:4.0433 train_time:43720ms step_avg:163.75ms
step:278/1530 train_loss:4.0906 train_time:43886ms step_avg:163.75ms
step:279/1530 train_loss:4.0044 train_time:44052ms step_avg:163.76ms
step:280/1530 train_loss:4.1916 train_time:44218ms step_avg:163.77ms
step:281/1530 train_loss:3.9763 train_time:44384ms step_avg:163.78ms
step:282/1530 train_loss:3.9434 train_time:44549ms step_avg:163.78ms
step:283/1530 train_loss:3.9118 train_time:44716ms step_avg:163.79ms
step:284/1530 train_loss:4.0520 train_time:44881ms step_avg:163.80ms
step:285/1530 train_loss:4.0684 train_time:45046ms step_avg:163.81ms
step:286/1530 train_loss:4.0968 train_time:45212ms step_avg:163.81ms
step:287/1530 train_loss:3.9060 train_time:45378ms step_avg:163.82ms
step:288/1530 train_loss:4.0098 train_time:45543ms step_avg:163.82ms
step:289/1530 train_loss:3.8750 train_time:45708ms step_avg:163.83ms
step:290/1530 train_loss:3.8564 train_time:45874ms step_avg:163.83ms
step:291/1530 train_loss:3.9010 train_time:46038ms step_avg:163.84ms
step:292/1530 train_loss:3.8653 train_time:46203ms step_avg:163.84ms
step:293/1530 train_loss:3.9012 train_time:46369ms step_avg:163.85ms
step:294/1530 train_loss:3.9402 train_time:46534ms step_avg:163.85ms
step:295/1530 train_loss:3.8418 train_time:46699ms step_avg:163.85ms
step:296/1530 train_loss:3.8562 train_time:46864ms step_avg:163.86ms
step:297/1530 train_loss:3.8605 train_time:47030ms step_avg:163.87ms
step:298/1530 train_loss:3.9719 train_time:47195ms step_avg:163.87ms
step:299/1530 train_loss:3.8239 train_time:47360ms step_avg:163.88ms
step:300/1530 train_loss:3.9696 train_time:47526ms step_avg:163.88ms
step:301/1530 train_loss:3.9578 train_time:47690ms step_avg:163.88ms
step:302/1530 train_loss:3.9358 train_time:47854ms step_avg:163.89ms
step:303/1530 train_loss:3.9792 train_time:48020ms step_avg:163.89ms
step:304/1530 train_loss:3.9679 train_time:48185ms step_avg:163.89ms
step:305/1530 train_loss:4.4568 train_time:48350ms step_avg:163.90ms
step:306/1530 train_loss:3.9340 train_time:48516ms step_avg:163.90ms
step:307/1530 train_loss:3.8377 train_time:48680ms step_avg:163.91ms
step:308/1530 train_loss:3.9856 train_time:48846ms step_avg:163.91ms
step:309/1530 train_loss:3.8741 train_time:49011ms step_avg:163.92ms
step:310/1530 train_loss:4.0890 train_time:49176ms step_avg:163.92ms
step:311/1530 train_loss:3.9264 train_time:49342ms step_avg:163.93ms
step:312/1530 train_loss:3.8662 train_time:49506ms step_avg:163.93ms
step:313/1530 train_loss:3.9300 train_time:49670ms step_avg:163.93ms
step:314/1530 train_loss:4.0592 train_time:49838ms step_avg:163.94ms
step:315/1530 train_loss:3.9407 train_time:50003ms step_avg:163.94ms
step:316/1530 train_loss:3.7951 train_time:50167ms step_avg:163.95ms
step:317/1530 train_loss:3.8756 train_time:50335ms step_avg:163.96ms
step:318/1530 train_loss:3.9204 train_time:50501ms step_avg:163.96ms
step:319/1530 train_loss:3.8952 train_time:50666ms step_avg:163.97ms
step:320/1530 train_loss:4.0216 train_time:50833ms step_avg:163.98ms
step:321/1530 train_loss:3.9563 train_time:50998ms step_avg:163.98ms
step:322/1530 train_loss:3.9285 train_time:51163ms step_avg:163.99ms
step:323/1530 train_loss:4.0040 train_time:51329ms step_avg:163.99ms
step:324/1530 train_loss:3.9473 train_time:51495ms step_avg:164.00ms
step:325/1530 train_loss:4.0163 train_time:51661ms step_avg:164.00ms
step:326/1530 train_loss:3.8922 train_time:51826ms step_avg:164.01ms
step:327/1530 train_loss:4.3883 train_time:51992ms step_avg:164.01ms
step:328/1530 train_loss:4.0750 train_time:52156ms step_avg:164.01ms
step:329/1530 train_loss:3.7997 train_time:52322ms step_avg:164.02ms
step:330/1530 train_loss:3.7481 train_time:52486ms step_avg:164.02ms
step:331/1530 train_loss:3.9771 train_time:52650ms step_avg:164.02ms
step:332/1530 train_loss:3.9238 train_time:52816ms step_avg:164.02ms
step:333/1530 train_loss:3.8864 train_time:52981ms step_avg:164.03ms
step:334/1530 train_loss:3.8452 train_time:53146ms step_avg:164.03ms
step:335/1530 train_loss:4.0073 train_time:53310ms step_avg:164.03ms
step:336/1530 train_loss:3.9570 train_time:53476ms step_avg:164.04ms
step:337/1530 train_loss:4.4280 train_time:53642ms step_avg:164.04ms
step:338/1530 train_loss:3.9325 train_time:53806ms step_avg:164.04ms
step:339/1530 train_loss:3.8673 train_time:53972ms step_avg:164.05ms
step:340/1530 train_loss:3.9301 train_time:54138ms step_avg:164.05ms
step:341/1530 train_loss:3.8524 train_time:54304ms step_avg:164.06ms
step:342/1530 train_loss:3.8123 train_time:54472ms step_avg:164.07ms
step:343/1530 train_loss:3.8372 train_time:54640ms step_avg:164.08ms
step:344/1530 train_loss:3.9925 train_time:54808ms step_avg:164.09ms
step:345/1530 train_loss:3.8161 train_time:54978ms step_avg:164.11ms
step:346/1530 train_loss:3.7670 train_time:55146ms step_avg:164.12ms
step:347/1530 train_loss:3.7964 train_time:55316ms step_avg:164.14ms
step:348/1530 train_loss:3.8600 train_time:55483ms step_avg:164.15ms
step:349/1530 train_loss:3.8385 train_time:55651ms step_avg:164.16ms
step:350/1530 train_loss:3.5672 train_time:55820ms step_avg:164.18ms
step:351/1530 train_loss:3.8267 train_time:55988ms step_avg:164.19ms
step:352/1530 train_loss:4.1859 train_time:56157ms step_avg:164.20ms
step:353/1530 train_loss:3.6626 train_time:56325ms step_avg:164.21ms
step:354/1530 train_loss:3.9269 train_time:56492ms step_avg:164.22ms
step:355/1530 train_loss:3.7886 train_time:56660ms step_avg:164.23ms
step:356/1530 train_loss:3.8856 train_time:56828ms step_avg:164.24ms
step:357/1530 train_loss:3.7623 train_time:56997ms step_avg:164.26ms
step:358/1530 train_loss:3.8657 train_time:57165ms step_avg:164.27ms
step:359/1530 train_loss:3.7890 train_time:57335ms step_avg:164.28ms
step:360/1530 train_loss:3.4328 train_time:57505ms step_avg:164.30ms
step:361/1530 train_loss:4.0161 train_time:57674ms step_avg:164.31ms
step:362/1530 train_loss:3.9273 train_time:57842ms step_avg:164.32ms
step:363/1530 train_loss:3.8375 train_time:58009ms step_avg:164.33ms
step:364/1530 train_loss:3.7468 train_time:58178ms step_avg:164.34ms
step:365/1530 train_loss:3.9180 train_time:58346ms step_avg:164.35ms
step:366/1530 train_loss:3.8564 train_time:58514ms step_avg:164.37ms
step:367/1530 train_loss:3.8583 train_time:58682ms step_avg:164.37ms
step:368/1530 train_loss:3.8482 train_time:58849ms step_avg:164.38ms
step:369/1530 train_loss:3.7446 train_time:59018ms step_avg:164.40ms
step:370/1530 train_loss:3.8813 train_time:59185ms step_avg:164.40ms
step:371/1530 train_loss:3.7305 train_time:59353ms step_avg:164.41ms
step:372/1530 train_loss:3.6959 train_time:59522ms step_avg:164.43ms
step:373/1530 train_loss:3.9157 train_time:59689ms step_avg:164.43ms
step:374/1530 train_loss:3.8273 train_time:59856ms step_avg:164.44ms
step:375/1530 train_loss:3.8019 train_time:60024ms step_avg:164.45ms
step:375/1530 val_loss:3.8286 train_time:60072ms step_avg:164.58ms