-
Notifications
You must be signed in to change notification settings - Fork 183
/
67716aee-6747-4997-a37c-b96932fab4dd.txt
2165 lines (2092 loc) · 134 KB
/
67716aee-6747-4997-a37c-b96932fab4dd.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 02:00:42 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 | 1% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 |
| N/A 31C P0 117W / 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 123W / 700W | 529MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 |
| N/A 29C P0 103W / 700W | 23MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 |
| N/A 38C P0 124W / 700W | 41MiB / 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:31804ms step_avg:nanms
step:2/1530 train_loss:10.0627 train_time:31915ms step_avg:nanms
step:3/1530 train_loss:8.3455 train_time:32075ms step_avg:nanms
step:4/1530 train_loss:7.5583 train_time:32237ms step_avg:nanms
step:5/1530 train_loss:7.4941 train_time:32397ms step_avg:nanms
step:6/1530 train_loss:6.9870 train_time:32557ms step_avg:nanms
step:7/1530 train_loss:7.2244 train_time:32717ms step_avg:nanms
step:8/1530 train_loss:6.7413 train_time:32878ms step_avg:nanms
step:9/1530 train_loss:6.6268 train_time:33039ms step_avg:nanms
step:10/1530 train_loss:6.4977 train_time:33199ms step_avg:nanms
step:11/1530 train_loss:6.3988 train_time:114ms step_avg:nanms
step:12/1530 train_loss:6.3206 train_time:275ms step_avg:nanms
step:13/1530 train_loss:6.2235 train_time:435ms step_avg:144.91ms
step:14/1530 train_loss:6.2065 train_time:594ms step_avg:148.52ms
step:15/1530 train_loss:6.1935 train_time:755ms step_avg:151.03ms
step:16/1530 train_loss:6.1411 train_time:916ms step_avg:152.68ms
step:17/1530 train_loss:6.1477 train_time:1075ms step_avg:153.52ms
step:18/1530 train_loss:5.9550 train_time:1234ms step_avg:154.26ms
step:19/1530 train_loss:5.9519 train_time:1395ms step_avg:155.00ms
step:20/1530 train_loss:5.6706 train_time:1555ms step_avg:155.49ms
step:21/1530 train_loss:5.9449 train_time:1714ms step_avg:155.80ms
step:22/1530 train_loss:6.1646 train_time:1875ms step_avg:156.22ms
step:23/1530 train_loss:5.8336 train_time:2035ms step_avg:156.54ms
step:24/1530 train_loss:6.0032 train_time:2196ms step_avg:156.85ms
step:25/1530 train_loss:5.6628 train_time:2356ms step_avg:157.06ms
step:26/1530 train_loss:5.5755 train_time:2516ms step_avg:157.22ms
step:27/1530 train_loss:5.7472 train_time:2676ms step_avg:157.40ms
step:28/1530 train_loss:5.4046 train_time:2835ms step_avg:157.48ms
step:29/1530 train_loss:5.6492 train_time:2995ms step_avg:157.65ms
step:30/1530 train_loss:5.4569 train_time:3156ms step_avg:157.79ms
step:31/1530 train_loss:5.4224 train_time:3315ms step_avg:157.87ms
step:32/1530 train_loss:5.2717 train_time:3475ms step_avg:157.96ms
step:33/1530 train_loss:5.5639 train_time:3636ms step_avg:158.07ms
step:34/1530 train_loss:5.4843 train_time:3795ms step_avg:158.14ms
step:35/1530 train_loss:5.5933 train_time:3954ms step_avg:158.17ms
step:36/1530 train_loss:5.5399 train_time:4114ms step_avg:158.24ms
step:37/1530 train_loss:5.4454 train_time:4275ms step_avg:158.34ms
step:38/1530 train_loss:5.2994 train_time:4435ms step_avg:158.40ms
step:39/1530 train_loss:5.3133 train_time:4596ms step_avg:158.47ms
step:40/1530 train_loss:5.2385 train_time:4756ms step_avg:158.54ms
step:41/1530 train_loss:5.2311 train_time:4916ms step_avg:158.58ms
step:42/1530 train_loss:5.1823 train_time:5075ms step_avg:158.60ms
step:43/1530 train_loss:5.2455 train_time:5235ms step_avg:158.62ms
step:44/1530 train_loss:5.2182 train_time:5396ms step_avg:158.71ms
step:45/1530 train_loss:5.3653 train_time:5556ms step_avg:158.75ms
step:46/1530 train_loss:5.1496 train_time:5716ms step_avg:158.78ms
step:47/1530 train_loss:5.0475 train_time:5876ms step_avg:158.82ms
step:48/1530 train_loss:5.1965 train_time:6037ms step_avg:158.87ms
step:49/1530 train_loss:5.1284 train_time:6196ms step_avg:158.86ms
step:50/1530 train_loss:5.2475 train_time:6356ms step_avg:158.89ms
step:51/1530 train_loss:5.1511 train_time:6516ms step_avg:158.93ms
step:52/1530 train_loss:5.0317 train_time:6676ms step_avg:158.95ms
step:53/1530 train_loss:5.1629 train_time:6836ms step_avg:158.97ms
step:54/1530 train_loss:5.0016 train_time:6997ms step_avg:159.03ms
step:55/1530 train_loss:5.3892 train_time:7157ms step_avg:159.05ms
step:56/1530 train_loss:5.0140 train_time:7317ms step_avg:159.06ms
step:57/1530 train_loss:4.8694 train_time:7478ms step_avg:159.10ms
step:58/1530 train_loss:5.0283 train_time:7639ms step_avg:159.15ms
step:59/1530 train_loss:5.0046 train_time:7800ms step_avg:159.18ms
step:60/1530 train_loss:5.1240 train_time:7962ms step_avg:159.23ms
step:61/1530 train_loss:4.8350 train_time:8122ms step_avg:159.25ms
step:62/1530 train_loss:4.9706 train_time:8283ms step_avg:159.29ms
step:63/1530 train_loss:4.9918 train_time:8443ms step_avg:159.30ms
step:64/1530 train_loss:4.8168 train_time:8604ms step_avg:159.34ms
step:65/1530 train_loss:4.8240 train_time:8765ms step_avg:159.36ms
step:66/1530 train_loss:4.9240 train_time:8925ms step_avg:159.37ms
step:67/1530 train_loss:4.8140 train_time:9086ms step_avg:159.40ms
step:68/1530 train_loss:5.0903 train_time:9246ms step_avg:159.41ms
step:69/1530 train_loss:4.7303 train_time:9407ms step_avg:159.45ms
step:70/1530 train_loss:4.8253 train_time:9567ms step_avg:159.46ms
step:71/1530 train_loss:4.9628 train_time:9728ms step_avg:159.47ms
step:72/1530 train_loss:4.8728 train_time:9889ms step_avg:159.50ms
step:73/1530 train_loss:4.7691 train_time:10048ms step_avg:159.50ms
step:74/1530 train_loss:4.8963 train_time:10209ms step_avg:159.52ms
step:75/1530 train_loss:4.8650 train_time:10370ms step_avg:159.54ms
step:76/1530 train_loss:4.7925 train_time:10531ms step_avg:159.56ms
step:77/1530 train_loss:4.9034 train_time:10692ms step_avg:159.59ms
step:78/1530 train_loss:5.1088 train_time:10852ms step_avg:159.59ms
step:79/1530 train_loss:4.8357 train_time:11013ms step_avg:159.60ms
step:80/1530 train_loss:4.8491 train_time:11173ms step_avg:159.61ms
step:81/1530 train_loss:4.6401 train_time:11333ms step_avg:159.62ms
step:82/1530 train_loss:4.8122 train_time:11493ms step_avg:159.63ms
step:83/1530 train_loss:4.7727 train_time:11653ms step_avg:159.63ms
step:84/1530 train_loss:4.7469 train_time:11814ms step_avg:159.65ms
step:85/1530 train_loss:4.6046 train_time:11975ms step_avg:159.66ms
step:86/1530 train_loss:4.8270 train_time:12135ms step_avg:159.67ms
step:87/1530 train_loss:4.7474 train_time:12296ms step_avg:159.69ms
step:88/1530 train_loss:4.7433 train_time:12457ms step_avg:159.70ms
step:89/1530 train_loss:4.7171 train_time:12617ms step_avg:159.71ms
step:90/1530 train_loss:4.6314 train_time:12776ms step_avg:159.70ms
step:91/1530 train_loss:4.6195 train_time:12937ms step_avg:159.71ms
step:92/1530 train_loss:4.7782 train_time:13097ms step_avg:159.72ms
step:93/1530 train_loss:4.6104 train_time:13257ms step_avg:159.72ms
step:94/1530 train_loss:4.6399 train_time:13417ms step_avg:159.72ms
step:95/1530 train_loss:4.6804 train_time:13577ms step_avg:159.74ms
step:96/1530 train_loss:4.5933 train_time:13738ms step_avg:159.74ms
step:97/1530 train_loss:4.6499 train_time:13897ms step_avg:159.73ms
step:98/1530 train_loss:4.5754 train_time:14058ms step_avg:159.75ms
step:99/1530 train_loss:4.6646 train_time:14218ms step_avg:159.75ms
step:100/1530 train_loss:4.6866 train_time:14378ms step_avg:159.75ms
step:101/1530 train_loss:4.5521 train_time:14538ms step_avg:159.76ms
step:102/1530 train_loss:4.7045 train_time:14699ms step_avg:159.78ms
step:103/1530 train_loss:4.5823 train_time:14860ms step_avg:159.78ms
step:104/1530 train_loss:4.5262 train_time:15019ms step_avg:159.78ms
step:105/1530 train_loss:4.5596 train_time:15181ms step_avg:159.80ms
step:106/1530 train_loss:4.6322 train_time:15341ms step_avg:159.80ms
step:107/1530 train_loss:4.5170 train_time:15501ms step_avg:159.81ms
step:108/1530 train_loss:4.3673 train_time:15663ms step_avg:159.83ms
step:109/1530 train_loss:4.4995 train_time:15824ms step_avg:159.84ms
step:110/1530 train_loss:4.4928 train_time:15984ms step_avg:159.84ms
step:111/1530 train_loss:4.4212 train_time:16144ms step_avg:159.84ms
step:112/1530 train_loss:4.5827 train_time:16305ms step_avg:159.86ms
step:113/1530 train_loss:4.4826 train_time:16465ms step_avg:159.86ms
step:114/1530 train_loss:4.3603 train_time:16626ms step_avg:159.87ms
step:115/1530 train_loss:4.5138 train_time:16790ms step_avg:159.90ms
step:116/1530 train_loss:4.4855 train_time:16954ms step_avg:159.94ms
step:117/1530 train_loss:4.3715 train_time:17117ms step_avg:159.97ms
step:118/1530 train_loss:4.5857 train_time:17282ms step_avg:160.02ms
step:119/1530 train_loss:4.4557 train_time:17446ms step_avg:160.05ms
step:120/1530 train_loss:4.3425 train_time:17610ms step_avg:160.09ms
step:121/1530 train_loss:4.3080 train_time:17773ms step_avg:160.12ms
step:122/1530 train_loss:4.4629 train_time:17937ms step_avg:160.15ms
step:123/1530 train_loss:4.2884 train_time:18101ms step_avg:160.19ms
step:124/1530 train_loss:4.5845 train_time:18266ms step_avg:160.23ms
step:125/1530 train_loss:4.4674 train_time:18430ms step_avg:160.26ms
step:125/1530 val_loss:4.4062 train_time:18477ms step_avg:160.67ms
step:126/1530 train_loss:4.4113 train_time:18596ms step_avg:160.31ms
step:127/1530 train_loss:4.4291 train_time:18760ms step_avg:160.34ms
step:128/1530 train_loss:4.3813 train_time:18923ms step_avg:160.37ms
step:129/1530 train_loss:4.6753 train_time:19087ms step_avg:160.40ms
step:130/1530 train_loss:4.3668 train_time:19251ms step_avg:160.42ms
step:131/1530 train_loss:4.3969 train_time:19415ms step_avg:160.45ms
step:132/1530 train_loss:4.3490 train_time:19579ms step_avg:160.49ms
step:133/1530 train_loss:4.4449 train_time:19743ms step_avg:160.51ms
step:134/1530 train_loss:4.2655 train_time:19907ms step_avg:160.54ms
step:135/1530 train_loss:4.4459 train_time:20071ms step_avg:160.57ms
step:136/1530 train_loss:4.2139 train_time:20234ms step_avg:160.59ms
step:137/1530 train_loss:4.3780 train_time:20399ms step_avg:160.62ms
step:138/1530 train_loss:4.2924 train_time:20562ms step_avg:160.64ms
step:139/1530 train_loss:4.3847 train_time:20725ms step_avg:160.66ms
step:140/1530 train_loss:4.4650 train_time:20889ms step_avg:160.69ms
step:141/1530 train_loss:4.3172 train_time:21054ms step_avg:160.72ms
step:142/1530 train_loss:4.3147 train_time:21218ms step_avg:160.74ms
step:143/1530 train_loss:4.2554 train_time:21382ms step_avg:160.76ms
step:144/1530 train_loss:4.3538 train_time:21545ms step_avg:160.79ms
step:145/1530 train_loss:4.3113 train_time:21710ms step_avg:160.82ms
step:146/1530 train_loss:4.1682 train_time:21875ms step_avg:160.85ms
step:147/1530 train_loss:4.3251 train_time:22039ms step_avg:160.87ms
step:148/1530 train_loss:4.3706 train_time:22203ms step_avg:160.89ms
step:149/1530 train_loss:4.3053 train_time:22367ms step_avg:160.92ms
step:150/1530 train_loss:4.4315 train_time:22532ms step_avg:160.94ms
step:151/1530 train_loss:4.2646 train_time:22696ms step_avg:160.96ms
step:152/1530 train_loss:4.2636 train_time:22859ms step_avg:160.98ms
step:153/1530 train_loss:4.3623 train_time:23023ms step_avg:161.00ms
step:154/1530 train_loss:4.3722 train_time:23187ms step_avg:161.02ms
step:155/1530 train_loss:4.2694 train_time:23351ms step_avg:161.04ms
step:156/1530 train_loss:4.3446 train_time:23514ms step_avg:161.05ms
step:157/1530 train_loss:4.4169 train_time:23679ms step_avg:161.08ms
step:158/1530 train_loss:4.2495 train_time:23842ms step_avg:161.09ms
step:159/1530 train_loss:4.3132 train_time:24005ms step_avg:161.11ms
step:160/1530 train_loss:4.1331 train_time:24171ms step_avg:161.14ms
step:161/1530 train_loss:4.3531 train_time:24334ms step_avg:161.15ms
step:162/1530 train_loss:4.3663 train_time:24497ms step_avg:161.17ms
step:163/1530 train_loss:4.3501 train_time:24661ms step_avg:161.18ms
step:164/1530 train_loss:4.1849 train_time:24825ms step_avg:161.20ms
step:165/1530 train_loss:4.2834 train_time:24988ms step_avg:161.21ms
step:166/1530 train_loss:4.3362 train_time:25154ms step_avg:161.24ms
step:167/1530 train_loss:4.2035 train_time:25318ms step_avg:161.26ms
step:168/1530 train_loss:4.2905 train_time:25481ms step_avg:161.27ms
step:169/1530 train_loss:4.1591 train_time:25645ms step_avg:161.29ms
step:170/1530 train_loss:4.0237 train_time:25810ms step_avg:161.31ms
step:171/1530 train_loss:4.2030 train_time:25972ms step_avg:161.32ms
step:172/1530 train_loss:4.2071 train_time:26136ms step_avg:161.33ms
step:173/1530 train_loss:4.2607 train_time:26299ms step_avg:161.35ms
step:174/1530 train_loss:4.4132 train_time:26463ms step_avg:161.36ms
step:175/1530 train_loss:4.2478 train_time:26625ms step_avg:161.36ms
step:176/1530 train_loss:4.0889 train_time:26787ms step_avg:161.37ms
step:177/1530 train_loss:4.0567 train_time:26949ms step_avg:161.37ms
step:178/1530 train_loss:4.1824 train_time:27113ms step_avg:161.39ms
step:179/1530 train_loss:4.1256 train_time:27277ms step_avg:161.40ms
step:180/1530 train_loss:4.1080 train_time:27439ms step_avg:161.41ms
step:181/1530 train_loss:4.2916 train_time:27603ms step_avg:161.42ms
step:182/1530 train_loss:4.1593 train_time:27766ms step_avg:161.43ms
step:183/1530 train_loss:4.1317 train_time:27928ms step_avg:161.43ms
step:184/1530 train_loss:4.1171 train_time:28090ms step_avg:161.44ms
step:185/1530 train_loss:4.1975 train_time:28253ms step_avg:161.45ms
step:186/1530 train_loss:4.1700 train_time:28416ms step_avg:161.45ms
step:187/1530 train_loss:4.2390 train_time:28579ms step_avg:161.46ms
step:188/1530 train_loss:4.1741 train_time:28881ms step_avg:162.25ms
step:189/1530 train_loss:4.1065 train_time:29210ms step_avg:163.19ms
step:190/1530 train_loss:4.2054 train_time:29373ms step_avg:163.18ms
step:191/1530 train_loss:4.0822 train_time:29536ms step_avg:163.18ms
step:192/1530 train_loss:4.0242 train_time:29700ms step_avg:163.19ms
step:193/1530 train_loss:4.2545 train_time:29864ms step_avg:163.19ms
step:194/1530 train_loss:4.1701 train_time:30025ms step_avg:163.18ms
step:195/1530 train_loss:4.3539 train_time:30188ms step_avg:163.18ms
step:196/1530 train_loss:4.1786 train_time:30352ms step_avg:163.18ms
step:197/1530 train_loss:4.0468 train_time:30515ms step_avg:163.18ms
step:198/1530 train_loss:4.1775 train_time:30678ms step_avg:163.18ms
step:199/1530 train_loss:4.0319 train_time:30841ms step_avg:163.18ms
step:200/1530 train_loss:4.1076 train_time:31004ms step_avg:163.18ms
step:201/1530 train_loss:4.0127 train_time:31168ms step_avg:163.18ms
step:202/1530 train_loss:4.2523 train_time:31331ms step_avg:163.18ms
step:203/1530 train_loss:4.0637 train_time:31494ms step_avg:163.18ms
step:204/1530 train_loss:4.1885 train_time:31658ms step_avg:163.19ms
step:205/1530 train_loss:4.2387 train_time:31820ms step_avg:163.18ms
step:206/1530 train_loss:3.9391 train_time:31983ms step_avg:163.18ms
step:207/1530 train_loss:4.0754 train_time:32146ms step_avg:163.18ms
step:208/1530 train_loss:4.0990 train_time:32309ms step_avg:163.18ms
step:209/1530 train_loss:4.2402 train_time:32473ms step_avg:163.18ms
step:210/1530 train_loss:4.1698 train_time:32639ms step_avg:163.19ms
step:211/1530 train_loss:4.0500 train_time:32802ms step_avg:163.19ms
step:212/1530 train_loss:4.1123 train_time:32964ms step_avg:163.19ms
step:213/1530 train_loss:4.0487 train_time:33127ms step_avg:163.19ms
step:214/1530 train_loss:4.1072 train_time:33291ms step_avg:163.19ms
step:215/1530 train_loss:3.9607 train_time:33454ms step_avg:163.19ms
step:216/1530 train_loss:4.0073 train_time:33616ms step_avg:163.19ms
step:217/1530 train_loss:4.0222 train_time:33779ms step_avg:163.19ms
step:218/1530 train_loss:4.0833 train_time:33943ms step_avg:163.19ms
step:219/1530 train_loss:4.0683 train_time:34106ms step_avg:163.19ms
step:220/1530 train_loss:4.0815 train_time:34269ms step_avg:163.19ms
step:221/1530 train_loss:4.0948 train_time:34431ms step_avg:163.18ms
step:222/1530 train_loss:3.9906 train_time:34595ms step_avg:163.18ms
step:223/1530 train_loss:3.9986 train_time:34758ms step_avg:163.18ms
step:224/1530 train_loss:4.2964 train_time:34920ms step_avg:163.18ms
step:225/1530 train_loss:3.9240 train_time:35083ms step_avg:163.18ms
step:226/1530 train_loss:3.9818 train_time:35246ms step_avg:163.18ms
step:227/1530 train_loss:3.9728 train_time:35409ms step_avg:163.18ms
step:228/1530 train_loss:4.1393 train_time:35575ms step_avg:163.19ms
step:229/1530 train_loss:3.9211 train_time:35742ms step_avg:163.20ms
step:230/1530 train_loss:4.0323 train_time:35907ms step_avg:163.21ms
step:231/1530 train_loss:3.9031 train_time:36074ms step_avg:163.23ms
step:232/1530 train_loss:3.9613 train_time:36240ms step_avg:163.24ms
step:233/1530 train_loss:4.0915 train_time:36404ms step_avg:163.25ms
step:234/1530 train_loss:4.0329 train_time:36573ms step_avg:163.27ms
step:235/1530 train_loss:3.9039 train_time:36740ms step_avg:163.29ms
step:236/1530 train_loss:4.0784 train_time:36906ms step_avg:163.30ms
step:237/1530 train_loss:4.0729 train_time:37074ms step_avg:163.32ms
step:238/1530 train_loss:3.9350 train_time:37240ms step_avg:163.33ms
step:239/1530 train_loss:4.0720 train_time:37406ms step_avg:163.35ms
step:240/1530 train_loss:4.1059 train_time:37574ms step_avg:163.37ms
step:241/1530 train_loss:3.9657 train_time:37740ms step_avg:163.38ms
step:242/1530 train_loss:4.1528 train_time:37907ms step_avg:163.39ms
step:243/1530 train_loss:4.0014 train_time:38074ms step_avg:163.41ms
step:244/1530 train_loss:4.0751 train_time:38241ms step_avg:163.42ms
step:245/1530 train_loss:4.1367 train_time:38408ms step_avg:163.44ms
step:246/1530 train_loss:4.0505 train_time:38575ms step_avg:163.45ms
step:247/1530 train_loss:3.9977 train_time:38741ms step_avg:163.46ms
step:248/1530 train_loss:4.1078 train_time:38908ms step_avg:163.48ms
step:249/1530 train_loss:3.9142 train_time:39076ms step_avg:163.50ms
step:250/1530 train_loss:3.9592 train_time:39242ms step_avg:163.51ms
step:250/1530 val_loss:3.9992 train_time:39290ms step_avg:163.71ms
step:251/1530 train_loss:4.0670 train_time:39411ms step_avg:163.53ms
step:252/1530 train_loss:4.1613 train_time:39578ms step_avg:163.54ms
step:253/1530 train_loss:3.9302 train_time:39744ms step_avg:163.55ms
step:254/1530 train_loss:3.8815 train_time:39909ms step_avg:163.56ms
step:255/1530 train_loss:4.0762 train_time:40075ms step_avg:163.57ms
step:256/1530 train_loss:3.9877 train_time:40242ms step_avg:163.58ms
step:257/1530 train_loss:3.9879 train_time:40407ms step_avg:163.59ms
step:258/1530 train_loss:3.9795 train_time:40574ms step_avg:163.60ms
step:259/1530 train_loss:4.0218 train_time:40741ms step_avg:163.62ms
step:260/1530 train_loss:4.0506 train_time:40908ms step_avg:163.63ms
step:261/1530 train_loss:4.0151 train_time:41075ms step_avg:163.65ms
step:262/1530 train_loss:3.9876 train_time:41242ms step_avg:163.66ms
step:263/1530 train_loss:3.8880 train_time:41408ms step_avg:163.67ms
step:264/1530 train_loss:3.9822 train_time:41575ms step_avg:163.68ms
step:265/1530 train_loss:3.8650 train_time:41742ms step_avg:163.69ms
step:266/1530 train_loss:3.9162 train_time:41907ms step_avg:163.70ms
step:267/1530 train_loss:3.9265 train_time:42073ms step_avg:163.71ms
step:268/1530 train_loss:3.9605 train_time:42240ms step_avg:163.72ms
step:269/1530 train_loss:3.8454 train_time:42406ms step_avg:163.73ms
step:270/1530 train_loss:4.0920 train_time:42573ms step_avg:163.74ms
step:271/1530 train_loss:3.9650 train_time:42739ms step_avg:163.75ms
step:272/1530 train_loss:3.9331 train_time:42905ms step_avg:163.76ms
step:273/1530 train_loss:3.9457 train_time:43069ms step_avg:163.76ms
step:274/1530 train_loss:4.0418 train_time:43239ms step_avg:163.79ms
step:275/1530 train_loss:4.0546 train_time:43406ms step_avg:163.80ms
step:276/1530 train_loss:4.2331 train_time:43574ms step_avg:163.81ms
step:277/1530 train_loss:4.0323 train_time:43740ms step_avg:163.82ms
step:278/1530 train_loss:4.0759 train_time:43905ms step_avg:163.83ms
step:279/1530 train_loss:3.9981 train_time:44071ms step_avg:163.83ms
step:280/1530 train_loss:4.2030 train_time:44240ms step_avg:163.85ms
step:281/1530 train_loss:3.9731 train_time:44406ms step_avg:163.86ms
step:282/1530 train_loss:3.9395 train_time:44573ms step_avg:163.87ms
step:283/1530 train_loss:3.9048 train_time:44739ms step_avg:163.88ms
step:284/1530 train_loss:4.0419 train_time:44905ms step_avg:163.89ms
step:285/1530 train_loss:4.0444 train_time:45071ms step_avg:163.89ms
step:286/1530 train_loss:4.0811 train_time:45237ms step_avg:163.90ms
step:287/1530 train_loss:3.9035 train_time:45402ms step_avg:163.91ms
step:288/1530 train_loss:4.0163 train_time:45567ms step_avg:163.91ms
step:289/1530 train_loss:3.8694 train_time:45734ms step_avg:163.92ms
step:290/1530 train_loss:3.8474 train_time:45900ms step_avg:163.93ms
step:291/1530 train_loss:3.9008 train_time:46065ms step_avg:163.93ms
step:292/1530 train_loss:3.8620 train_time:46231ms step_avg:163.94ms
step:293/1530 train_loss:3.8971 train_time:46396ms step_avg:163.94ms
step:294/1530 train_loss:3.9338 train_time:46562ms step_avg:163.95ms
step:295/1530 train_loss:3.8390 train_time:46727ms step_avg:163.95ms
step:296/1530 train_loss:3.8580 train_time:46893ms step_avg:163.96ms
step:297/1530 train_loss:3.8600 train_time:47058ms step_avg:163.97ms
step:298/1530 train_loss:3.9697 train_time:47223ms step_avg:163.97ms
step:299/1530 train_loss:3.8168 train_time:47388ms step_avg:163.97ms
step:300/1530 train_loss:3.9585 train_time:47553ms step_avg:163.98ms
step:301/1530 train_loss:3.9575 train_time:47719ms step_avg:163.98ms
step:302/1530 train_loss:3.9284 train_time:47883ms step_avg:163.98ms
step:303/1530 train_loss:3.9839 train_time:48048ms step_avg:163.99ms
step:304/1530 train_loss:3.9594 train_time:48215ms step_avg:164.00ms
step:305/1530 train_loss:4.4468 train_time:48380ms step_avg:164.00ms
step:306/1530 train_loss:3.9433 train_time:48544ms step_avg:164.00ms
step:307/1530 train_loss:3.8345 train_time:48709ms step_avg:164.00ms
step:308/1530 train_loss:3.9707 train_time:48875ms step_avg:164.01ms
step:309/1530 train_loss:3.8722 train_time:49041ms step_avg:164.02ms
step:310/1530 train_loss:4.0804 train_time:49206ms step_avg:164.02ms
step:311/1530 train_loss:3.9209 train_time:49370ms step_avg:164.02ms
step:312/1530 train_loss:3.8614 train_time:49537ms step_avg:164.03ms
step:313/1530 train_loss:3.9260 train_time:49702ms step_avg:164.03ms
step:314/1530 train_loss:4.0626 train_time:49867ms step_avg:164.03ms
step:315/1530 train_loss:3.9369 train_time:50034ms step_avg:164.05ms
step:316/1530 train_loss:3.7918 train_time:50200ms step_avg:164.05ms
step:317/1530 train_loss:3.8719 train_time:50364ms step_avg:164.05ms
step:318/1530 train_loss:3.9140 train_time:50529ms step_avg:164.05ms
step:319/1530 train_loss:3.8880 train_time:50694ms step_avg:164.06ms
step:320/1530 train_loss:4.0055 train_time:50859ms step_avg:164.06ms
step:321/1530 train_loss:3.9439 train_time:51024ms step_avg:164.06ms
step:322/1530 train_loss:3.9296 train_time:51189ms step_avg:164.07ms
step:323/1530 train_loss:4.0005 train_time:51355ms step_avg:164.07ms
step:324/1530 train_loss:3.9429 train_time:51521ms step_avg:164.08ms
step:325/1530 train_loss:4.0043 train_time:51685ms step_avg:164.08ms
step:326/1530 train_loss:3.8914 train_time:51851ms step_avg:164.09ms
step:327/1530 train_loss:4.3926 train_time:52019ms step_avg:164.10ms
step:328/1530 train_loss:4.0638 train_time:52183ms step_avg:164.10ms
step:329/1530 train_loss:3.7852 train_time:52348ms step_avg:164.10ms
step:330/1530 train_loss:3.7452 train_time:52513ms step_avg:164.10ms
step:331/1530 train_loss:3.9676 train_time:52678ms step_avg:164.11ms
step:332/1530 train_loss:3.9061 train_time:52843ms step_avg:164.11ms
step:333/1530 train_loss:3.8724 train_time:53007ms step_avg:164.11ms
step:334/1530 train_loss:3.8363 train_time:53174ms step_avg:164.12ms
step:335/1530 train_loss:4.0072 train_time:53340ms step_avg:164.12ms
step:336/1530 train_loss:3.9659 train_time:53505ms step_avg:164.13ms
step:337/1530 train_loss:4.4142 train_time:53671ms step_avg:164.13ms
step:338/1530 train_loss:3.9266 train_time:53837ms step_avg:164.14ms
step:339/1530 train_loss:3.8570 train_time:54002ms step_avg:164.14ms
step:340/1530 train_loss:3.9267 train_time:54167ms step_avg:164.14ms
step:341/1530 train_loss:3.8452 train_time:54334ms step_avg:164.15ms
step:342/1530 train_loss:3.8056 train_time:54502ms step_avg:164.16ms
step:343/1530 train_loss:3.8350 train_time:54670ms step_avg:164.17ms
step:344/1530 train_loss:3.9917 train_time:54837ms step_avg:164.18ms
step:345/1530 train_loss:3.8108 train_time:55006ms step_avg:164.20ms
step:346/1530 train_loss:3.7558 train_time:55173ms step_avg:164.20ms
step:347/1530 train_loss:3.7935 train_time:55342ms step_avg:164.22ms
step:348/1530 train_loss:3.8509 train_time:55511ms step_avg:164.23ms
step:349/1530 train_loss:3.8234 train_time:55680ms step_avg:164.25ms
step:350/1530 train_loss:3.5623 train_time:55847ms step_avg:164.26ms
step:351/1530 train_loss:3.8192 train_time:56015ms step_avg:164.27ms
step:352/1530 train_loss:4.1932 train_time:56182ms step_avg:164.28ms
step:353/1530 train_loss:3.6588 train_time:56350ms step_avg:164.29ms
step:354/1530 train_loss:3.9273 train_time:56518ms step_avg:164.30ms
step:355/1530 train_loss:3.7872 train_time:56685ms step_avg:164.30ms
step:356/1530 train_loss:3.8803 train_time:56853ms step_avg:164.31ms
step:357/1530 train_loss:3.7463 train_time:57023ms step_avg:164.33ms
step:358/1530 train_loss:3.8546 train_time:57190ms step_avg:164.34ms
step:359/1530 train_loss:3.7785 train_time:57360ms step_avg:164.36ms
step:360/1530 train_loss:3.4220 train_time:57530ms step_avg:164.37ms
step:361/1530 train_loss:4.0111 train_time:57700ms step_avg:164.39ms
step:362/1530 train_loss:3.9107 train_time:57868ms step_avg:164.40ms
step:363/1530 train_loss:3.8324 train_time:58037ms step_avg:164.41ms
step:364/1530 train_loss:3.7375 train_time:58205ms step_avg:164.42ms
step:365/1530 train_loss:3.9060 train_time:58374ms step_avg:164.43ms
step:366/1530 train_loss:3.8562 train_time:58542ms step_avg:164.44ms
step:367/1530 train_loss:3.8473 train_time:58710ms step_avg:164.45ms
step:368/1530 train_loss:3.8463 train_time:58878ms step_avg:164.46ms
step:369/1530 train_loss:3.7412 train_time:59045ms step_avg:164.47ms
step:370/1530 train_loss:3.8717 train_time:59213ms step_avg:164.48ms
step:371/1530 train_loss:3.7277 train_time:59380ms step_avg:164.49ms
step:372/1530 train_loss:3.6881 train_time:59548ms step_avg:164.50ms
step:373/1530 train_loss:3.9023 train_time:59717ms step_avg:164.51ms
step:374/1530 train_loss:3.8159 train_time:59883ms step_avg:164.52ms
step:375/1530 train_loss:3.7908 train_time:60052ms step_avg:164.53ms
step:375/1530 val_loss:3.8175 train_time:60101ms step_avg:164.66ms