-
Notifications
You must be signed in to change notification settings - Fork 137
/
1621af10-aa0c-42af-bf54-8a773c63a2af.txt
3780 lines (3712 loc) · 247 KB
/
1621af10-aa0c-42af-bf54-8a773c63a2af.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
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
# -----------------------------------------------------------------------------
# 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: torch.Tensor, steps: int = 10):
"""
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}' \sim 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()
I = torch.eye(min(G.size(0), G.size(1)), dtype=X.dtype, device=X.device)
X.div_(X.norm() + 1e-7) # ensure top singular value <= 1
if G.size(0) > G.size(1):
X = X.T
for _ in range(steps):
A = X @ X.T
S = A @ (b * I + c * A)
torch.diagonal(S).add_(a)
X = S @ 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)
if group['nesterov']:
g = g.add(buf, alpha=momentum)
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
class Rotary(torch.nn.Module):
def __init__(self, dim, base=10000):
super().__init__()
self.inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / 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:
self.seq_len_cached = seq_len
t = torch.arange(seq_len, device=x.device).type_as(self.inv_freq)
freqs = torch.outer(t, self.inv_freq).to(x.device)
self.cos_cached = freqs.cos().bfloat16()
self.sin_cached = freqs.sin().bfloat16()
return self.cos_cached[None, :, None, :], self.sin_cached[None, :, None, :]
def apply_rotary_emb(x, cos, sin):
assert x.ndim == 4 # multihead attention
d = x.shape[3]//2
x1 = x[..., :d]
x2 = x[..., d:]
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, config):
super().__init__()
self.n_head = config.n_head
self.n_embd = config.n_embd
self.head_dim = self.n_embd // self.n_head
assert self.n_embd % self.n_head == 0
self.c_q = nn.Linear(self.n_embd, self.n_embd, bias=False)
self.c_k = nn.Linear(self.n_embd, self.n_embd, bias=False)
self.c_v = nn.Linear(self.n_embd, self.n_embd, bias=False)
# output projection
self.c_proj = nn.Linear(self.n_embd, self.n_embd, bias=False)
self.c_proj.weight.data.zero_() # zero init suggested by @Grad62304977
self.rotary = Rotary(self.head_dim)
self.lambdas = nn.Parameter(torch.tensor([0.5, 0.5]))
def forward(self, x, v1=None):
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
q = self.c_q(x).view(B, T, self.n_head, self.head_dim)
k = self.c_k(x).view(B, T, self.n_head, self.head_dim)
v = self.c_v(x).view(B, T, self.n_head, self.head_dim)
if v1 is None:
v1 = v # This happens if we are in the first block. v needs to be accessed by subsequent blocks
v = self.lambdas[0] * v + self.lambdas[1] * v1.view_as(v) # @Grad62304977
cos, sin = self.rotary(q)
q, k = F.rms_norm(q, (q.size(-1),)), F.rms_norm(k, (k.size(-1),)) # QK norm suggested by @Grad62304977
q, k = apply_rotary_emb(q, cos, sin), apply_rotary_emb(k, cos, sin)
y = F.scaled_dot_product_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), is_causal=True)
y = y.transpose(1, 2).contiguous().view_as(x) # re-assemble all head outputs side by side
y = self.c_proj(y)
return y, v1
class MLP(nn.Module):
def __init__(self, config):
super().__init__()
self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=False)
self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=False)
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)
self.mlp = MLP(config)
self.lambdas = nn.Parameter(torch.tensor([1., 0.]))
def forward(self, x, v1, x0):
x = self.lambdas[0] * x + self.lambdas[1] * x0
x1, v1 = self.attn(F.rms_norm(x, (x.size(-1),)), v1)
x = x + x1
x = x + self.mlp(F.rms_norm(x, (x.size(-1),)))
return x, v1
# -----------------------------------------------------------------------------
# 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__()
self.config = config
self.transformer = nn.ModuleDict(dict(
wte = nn.Embedding(config.vocab_size, config.n_embd),
h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
))
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
self.lm_head.weight.data.zero_() # @Grad62304977
def forward(self, idx, targets=None, return_logits=True):
# forward the GPT model itself
x = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
x = F.rms_norm(x, (x.size(-1),)) # @Grad62304977
x0 = x
v1 = None
for block in self.transformer.h:
x, v1 = block(x, v1, x0)
x = F.rms_norm(x, (x.size(-1),))
if targets is not None:
# if we are given some desired targets also calculate the loss
logits = self.lm_head(x)
logits = 30 * torch.tanh(logits / 30) # @Grad62304977
logits = logits.float() # use tf32/fp32 for logits
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
else:
# inference-time mini-optimization: only forward the lm_head on the very last position
logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim
logits = 30 * torch.tanh(logits / 30) # @Grad62304977
logits = logits.float() # use tf32/fp32 for logits
loss = None
# there are performance reasons why not returning logits is prudent, if not needed
if not return_logits:
logits = None
return logits, 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, B, T, process_rank, num_processes):
self.process_rank = process_rank
self.num_processes = num_processes
self.B = B
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 * B * T + 1
ntok_total += int(shard_ntok)
self.ntok_total = ntok_total
# kick things off
self.reset()
def reset(self):
self.current_shard = 0
self.current_position = self.process_rank * self.B * self.T
self.tokens = _load_data_shard(self.files[self.current_shard])
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.B * self.T
self.tokens = _load_data_shard(self.files[self.current_shard])
def next_batch(self):
B = self.B
T = self.T
buf = self.tokens[self.current_position : self.current_position+B*T+1]
buf = torch.tensor(buf.astype(np.int32), dtype=torch.long)
x = (buf[:-1]).view(B, T) # inputs
y = (buf[1:]).view(B, T) # targets
# advance current position and load next shard if necessary
self.current_position += B * T * self.num_processes
if self.current_position + (B * T * self.num_processes + 1) > 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*64 # batch size, in sequences, across all devices
device_batch_size : int = 64 # batch size, in sequences, per device
sequence_length : int = 1024 # sequence length, in tokens
num_iterations : int = 3125 # number of iterations to run
warmup_iters : int = 0
warmdown_iters : int = 914 # number of iterations of linear warmup/warmdown 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.
# convenience variables
B, T = args.device_batch_size, args.sequence_length
# calculate the number of steps to take in the val loop.
assert args.val_tokens % (B * T * ddp_world_size) == 0
val_steps = args.val_tokens // (B * T * ddp_world_size)
# calculate the steps of gradient accumulation required to attain the desired global batch size.
assert args.batch_size % (B * ddp_world_size) == 0
train_accumulation_steps = args.batch_size // (B * ddp_world_size)
# load tokens
train_loader = DistributedDataLoader(args.input_bin, B, T, ddp_rank, ddp_world_size)
val_loader = DistributedDataLoader(args.input_val_bin, B, T, ddp_rank, ddp_world_size)
if master_process:
print(f"Training DataLoader: total number of tokens: {train_loader.ntok_total} across {len(train_loader.files)} files")
print(f"Validation DataLoader: total number of tokens: {val_loader.ntok_total} across {len(val_loader.files)} files")
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.
if master_process:
print("Building model...")
num_vocab = 50304
model = GPT(GPTConfig(vocab_size=num_vocab, n_layer=12, n_head=6, n_embd=768))
model = model.to(device)
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
ctx = torch.amp.autocast(device_type='cuda', dtype=torch.bfloat16)
if master_process:
print("Model built.")
# CUDNN attention is ~4ms faster than Flash, but doesn't get selected by default in PyTorch 2.5.1
from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp
enable_cudnn_sdp(True)
enable_flash_sdp(False)
enable_mem_efficient_sdp(False)
enable_math_sdp(False)
# init the optimizer(s)
optimizer1 = torch.optim.Adam([raw_model.transformer.wte.weight], lr=0.3, betas=(0.9, 0.95), fused=True)
optimizer2 = torch.optim.Adam([raw_model.lm_head.weight], lr=0.002, betas=(0.9, 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]
optimizer3 = Muon(matrix_params, lr=0.02, momentum=0.95)
optimizer4 = torch.optim.Adam(scalar_params, lr=0.02, betas=(0.9, 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 warmdown)
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.warmdown_iters:
return 1.0
# 3) linear warmdown
else:
decay_ratio = (args.num_iterations - it) / args.warmdown_iters
return decay_ratio
schedulers = [torch.optim.lr_scheduler.LambdaLR(opt, get_lr) for opt in optimizers]
# begin logging
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('='*100 + '\n')
f.write(code)
f.write('='*100 + '\n')
# log information about the hardware/software environment this is running on
# and print the full `nvidia-smi` to file
f.write(f"Running pytorch {torch.version.__version__} compiled for CUDA {torch.version.cuda}\nnvidia-smi:\n")
import subprocess
result = subprocess.run(['nvidia-smi'], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
f.write(f'{result.stdout}\n')
f.write('='*100 + '\n')
training_time_ms = 0
# start the clock
torch.cuda.synchronize()
t0 = time.time()
# begin training
train_loader.reset()
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
# 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):
x_val, y_val = val_loader.next_batch()
with ctx: # of course, we'd like to use no_grad() here too, but that creates a torch.compile error for some reason
_, loss = model(x_val, y_val, return_logits=False)
val_loss += loss.detach()
del loss
dist.all_reduce(val_loss, op=dist.ReduceOp.AVG)
val_loss /= val_steps
# log val loss to console and to logfile
if master_process:
print(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')
with open(logfile, "a") as f:
f.write(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\n')
# 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):
# forward pass
with ctx:
_, loss = model(x, y, return_logits=False)
train_loss = loss.detach()
# advance the dataset for the next batch
x, y = train_loader.next_batch()
# backward pass
if i < train_accumulation_steps:
with model.no_sync(): # there's no need to sync gradients every accumulation step
loss.backward()
else:
loss.backward() # just sync on the last step
for p in model.parameters():
p.grad /= train_accumulation_steps
# momentum warmup for Muon
frac = min(step/500, 1)
frac2 = 1 - max(step - args.num_iterations + 500, 0) / 500
optimizer3.param_groups[0]['momentum'] = (1 - frac) * 0.85 + frac * 0.95
optimizer1.param_groups[0]['betas'] = (
(1 - frac2) * 0.80 + frac2 * 0.90,
(1 - frac2) * 0.85 + frac2 * 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
if master_process:
approx_time = training_time_ms + 1000 * (time.time() - t0)
print(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")
with open(logfile, "a") as f:
f.write(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\n")
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 2.5.1+cu124 compiled for CUDA 12.4
nvidia-smi:
Sat Nov 9 14:38:00 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.35.03 Driver Version: 560.35.03 CUDA Version: 12.6 |
|-----------------------------------------+------------------------+----------------------+
| 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 Off | 00000000:18:00.0 Off | 0 |
| N/A 30C P0 115W / 700W | 5304MiB / 81559MiB | 1% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 1 NVIDIA H100 80GB HBM3 Off | 00000000:2A:00.0 Off | 0 |
| N/A 32C P0 116W / 700W | 5352MiB / 81559MiB | 7% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 2 NVIDIA H100 80GB HBM3 Off | 00000000:3A:00.0 Off | 0 |
| N/A 33C P0 121W / 700W | 5352MiB / 81559MiB | 1% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 3 NVIDIA H100 80GB HBM3 Off | 00000000:5D:00.0 Off | 0 |
| N/A 31C P0 117W / 700W | 5352MiB / 81559MiB | 6% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 4 NVIDIA H100 80GB HBM3 Off | 00000000:84:00.0 Off | 0 |
| N/A 31C P0 121W / 700W | 5352MiB / 81559MiB | 6% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 5 NVIDIA H100 80GB HBM3 Off | 00000000:8B:00.0 Off | 0 |
| N/A 33C P0 119W / 700W | 5352MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 6 NVIDIA H100 80GB HBM3 Off | 00000000:91:00.0 Off | 0 |
| N/A 32C P0 123W / 700W | 5352MiB / 81559MiB | 9% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 7 NVIDIA H100 80GB HBM3 Off | 00000000:E4:00.0 Off | 0 |
| N/A 31C P0 120W / 700W | 5112MiB / 81559MiB | 6% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 951 C /usr/bin/python3 0MiB |
| 1 N/A N/A 952 C /usr/bin/python3 0MiB |
| 2 N/A N/A 953 C /usr/bin/python3 0MiB |
| 3 N/A N/A 954 C /usr/bin/python3 0MiB |
| 4 N/A N/A 955 C /usr/bin/python3 0MiB |
| 5 N/A N/A 956 C /usr/bin/python3 0MiB |
| 6 N/A N/A 957 C /usr/bin/python3 0MiB |
| 7 N/A N/A 958 C /usr/bin/python3 0MiB |
+-----------------------------------------------------------------------------------------+
====================================================================================================
step:0/3125 val_loss:10.8258 train_time:396ms step_avg:nanms
step:1/3125 train_loss:10.8258 train_time:54012ms step_avg:nanms
step:2/3125 train_loss:10.4261 train_time:54118ms step_avg:nanms
step:3/3125 train_loss:9.9420 train_time:54264ms step_avg:nanms
step:4/3125 train_loss:9.0165 train_time:54414ms step_avg:nanms
step:5/3125 train_loss:8.0278 train_time:54566ms step_avg:nanms
step:6/3125 train_loss:7.5126 train_time:54716ms step_avg:nanms
step:7/3125 train_loss:7.0222 train_time:54866ms step_avg:nanms
step:8/3125 train_loss:7.2630 train_time:55021ms step_avg:nanms
step:9/3125 train_loss:6.9143 train_time:55178ms step_avg:nanms
step:10/3125 train_loss:6.8157 train_time:55331ms step_avg:nanms
step:11/3125 train_loss:6.6897 train_time:104ms step_avg:nanms
step:12/3125 train_loss:6.6411 train_time:254ms step_avg:nanms
step:13/3125 train_loss:6.4859 train_time:405ms step_avg:134.99ms
step:14/3125 train_loss:6.4686 train_time:556ms step_avg:139.06ms
step:15/3125 train_loss:6.4627 train_time:711ms step_avg:142.14ms
step:16/3125 train_loss:6.4082 train_time:866ms step_avg:144.39ms
step:17/3125 train_loss:6.4091 train_time:1019ms step_avg:145.60ms
step:18/3125 train_loss:6.4478 train_time:1172ms step_avg:146.49ms
step:19/3125 train_loss:6.2866 train_time:1324ms step_avg:147.08ms
step:20/3125 train_loss:6.3076 train_time:1474ms step_avg:147.43ms
step:21/3125 train_loss:6.0141 train_time:1627ms step_avg:147.94ms
step:22/3125 train_loss:6.3312 train_time:1780ms step_avg:148.34ms
step:23/3125 train_loss:6.5705 train_time:1937ms step_avg:149.00ms
step:24/3125 train_loss:6.2327 train_time:2090ms step_avg:149.29ms
step:25/3125 train_loss:6.3876 train_time:2242ms step_avg:149.48ms
step:26/3125 train_loss:6.0977 train_time:2394ms step_avg:149.64ms
step:27/3125 train_loss:6.0111 train_time:2546ms step_avg:149.78ms
step:28/3125 train_loss:6.1932 train_time:2698ms step_avg:149.90ms
step:29/3125 train_loss:5.8512 train_time:2852ms step_avg:150.11ms
step:30/3125 train_loss:6.1101 train_time:3007ms step_avg:150.35ms
step:31/3125 train_loss:5.9462 train_time:3160ms step_avg:150.47ms
step:32/3125 train_loss:5.9173 train_time:3313ms step_avg:150.60ms
step:33/3125 train_loss:5.7513 train_time:3467ms step_avg:150.73ms
step:34/3125 train_loss:6.0540 train_time:3618ms step_avg:150.74ms
step:35/3125 train_loss:5.9704 train_time:3773ms step_avg:150.91ms
step:36/3125 train_loss:6.1164 train_time:3926ms step_avg:151.02ms
step:37/3125 train_loss:6.0245 train_time:4078ms step_avg:151.03ms
step:38/3125 train_loss:5.9221 train_time:4233ms step_avg:151.16ms
step:39/3125 train_loss:5.8179 train_time:4385ms step_avg:151.21ms
step:40/3125 train_loss:5.8284 train_time:4537ms step_avg:151.24ms
step:41/3125 train_loss:5.7434 train_time:4691ms step_avg:151.31ms
step:42/3125 train_loss:5.7435 train_time:4845ms step_avg:151.40ms
step:43/3125 train_loss:5.6391 train_time:4997ms step_avg:151.43ms
step:44/3125 train_loss:5.7260 train_time:5151ms step_avg:151.49ms
step:45/3125 train_loss:5.7058 train_time:5302ms step_avg:151.49ms
step:46/3125 train_loss:5.8419 train_time:5455ms step_avg:151.53ms
step:47/3125 train_loss:5.6445 train_time:5609ms step_avg:151.58ms
step:48/3125 train_loss:5.5161 train_time:5762ms step_avg:151.62ms
step:49/3125 train_loss:5.7064 train_time:5916ms step_avg:151.69ms
step:50/3125 train_loss:5.5812 train_time:6070ms step_avg:151.75ms
step:51/3125 train_loss:5.7231 train_time:6224ms step_avg:151.81ms
step:52/3125 train_loss:5.5884 train_time:6377ms step_avg:151.83ms
step:53/3125 train_loss:5.4458 train_time:6532ms step_avg:151.90ms
step:54/3125 train_loss:5.5656 train_time:6684ms step_avg:151.91ms
step:55/3125 train_loss:5.4369 train_time:6836ms step_avg:151.92ms
step:56/3125 train_loss:5.7777 train_time:6990ms step_avg:151.96ms
step:57/3125 train_loss:5.4376 train_time:7144ms step_avg:151.99ms
step:58/3125 train_loss:5.3100 train_time:7297ms step_avg:152.02ms
step:59/3125 train_loss:5.4322 train_time:7450ms step_avg:152.04ms
step:60/3125 train_loss:5.4089 train_time:7602ms step_avg:152.03ms
step:61/3125 train_loss:5.5060 train_time:7754ms step_avg:152.04ms
step:62/3125 train_loss:5.2616 train_time:7908ms step_avg:152.07ms
step:63/3125 train_loss:5.3716 train_time:8060ms step_avg:152.07ms
step:64/3125 train_loss:5.3461 train_time:8212ms step_avg:152.08ms
step:65/3125 train_loss:5.1526 train_time:8365ms step_avg:152.09ms
step:66/3125 train_loss:5.1552 train_time:8517ms step_avg:152.08ms
step:67/3125 train_loss:5.3056 train_time:8671ms step_avg:152.12ms
step:68/3125 train_loss:5.1777 train_time:8825ms step_avg:152.15ms
step:69/3125 train_loss:5.4080 train_time:8977ms step_avg:152.15ms
step:70/3125 train_loss:5.0760 train_time:9130ms step_avg:152.17ms
step:71/3125 train_loss:5.1377 train_time:9284ms step_avg:152.20ms
step:72/3125 train_loss:5.3021 train_time:9438ms step_avg:152.22ms
step:73/3125 train_loss:5.2423 train_time:9591ms step_avg:152.23ms
step:74/3125 train_loss:5.1207 train_time:9743ms step_avg:152.23ms
step:75/3125 train_loss:5.2410 train_time:9895ms step_avg:152.23ms
step:76/3125 train_loss:5.2206 train_time:10050ms step_avg:152.27ms
step:77/3125 train_loss:5.1605 train_time:10202ms step_avg:152.28ms
step:78/3125 train_loss:5.2586 train_time:10356ms step_avg:152.30ms
step:79/3125 train_loss:5.3542 train_time:10510ms step_avg:152.31ms
step:80/3125 train_loss:5.1061 train_time:10664ms step_avg:152.34ms
step:81/3125 train_loss:5.1902 train_time:10817ms step_avg:152.35ms
step:82/3125 train_loss:4.9574 train_time:10972ms step_avg:152.38ms
step:83/3125 train_loss:5.1413 train_time:11124ms step_avg:152.38ms
step:84/3125 train_loss:5.0892 train_time:11276ms step_avg:152.37ms
step:85/3125 train_loss:5.0735 train_time:11430ms step_avg:152.40ms
step:86/3125 train_loss:4.9389 train_time:11582ms step_avg:152.40ms
step:87/3125 train_loss:5.1425 train_time:11737ms step_avg:152.42ms
step:88/3125 train_loss:5.0428 train_time:11890ms step_avg:152.43ms
step:89/3125 train_loss:5.0948 train_time:12042ms step_avg:152.44ms
step:90/3125 train_loss:5.0627 train_time:12195ms step_avg:152.43ms
step:91/3125 train_loss:4.9775 train_time:12349ms step_avg:152.46ms
step:92/3125 train_loss:4.9784 train_time:12502ms step_avg:152.46ms
step:93/3125 train_loss:5.1006 train_time:12656ms step_avg:152.48ms
step:94/3125 train_loss:4.9287 train_time:12809ms step_avg:152.48ms
step:95/3125 train_loss:4.9311 train_time:12961ms step_avg:152.49ms
step:96/3125 train_loss:4.9759 train_time:13114ms step_avg:152.48ms
step:97/3125 train_loss:4.8791 train_time:13267ms step_avg:152.50ms
step:98/3125 train_loss:4.9504 train_time:13420ms step_avg:152.50ms
step:99/3125 train_loss:4.8856 train_time:13576ms step_avg:152.53ms
step:100/3125 train_loss:4.9870 train_time:13729ms step_avg:152.55ms
step:101/3125 train_loss:4.9617 train_time:13881ms step_avg:152.54ms
step:102/3125 train_loss:4.8356 train_time:14035ms step_avg:152.55ms
step:103/3125 train_loss:4.9724 train_time:14188ms step_avg:152.56ms
step:104/3125 train_loss:4.9054 train_time:14340ms step_avg:152.55ms
step:105/3125 train_loss:4.7951 train_time:14496ms step_avg:152.59ms
step:106/3125 train_loss:4.8399 train_time:14649ms step_avg:152.60ms
step:107/3125 train_loss:4.9966 train_time:14801ms step_avg:152.59ms
step:108/3125 train_loss:4.8096 train_time:14954ms step_avg:152.60ms
step:109/3125 train_loss:4.6138 train_time:15108ms step_avg:152.61ms
step:110/3125 train_loss:4.7772 train_time:15261ms step_avg:152.61ms
step:111/3125 train_loss:4.7652 train_time:15415ms step_avg:152.62ms
step:112/3125 train_loss:4.7062 train_time:15568ms step_avg:152.63ms
step:113/3125 train_loss:4.8524 train_time:15720ms step_avg:152.62ms
step:114/3125 train_loss:4.7537 train_time:15874ms step_avg:152.64ms
step:115/3125 train_loss:4.6237 train_time:16029ms step_avg:152.65ms
step:116/3125 train_loss:4.7756 train_time:16181ms step_avg:152.65ms
step:117/3125 train_loss:4.7004 train_time:16335ms step_avg:152.66ms
step:118/3125 train_loss:4.6319 train_time:16489ms step_avg:152.67ms
step:119/3125 train_loss:4.8120 train_time:16642ms step_avg:152.68ms
step:120/3125 train_loss:4.7115 train_time:16795ms step_avg:152.68ms
step:121/3125 train_loss:4.6102 train_time:16948ms step_avg:152.69ms
step:122/3125 train_loss:4.5419 train_time:17101ms step_avg:152.69ms
step:123/3125 train_loss:4.6831 train_time:17254ms step_avg:152.69ms
step:124/3125 train_loss:4.5294 train_time:17408ms step_avg:152.70ms
step:125/3125 train_loss:4.8178 train_time:17560ms step_avg:152.70ms
step:125/3125 val_loss:4.6392 train_time:17612ms step_avg:153.15ms
step:126/3125 train_loss:4.6867 train_time:17724ms step_avg:152.80ms
step:127/3125 train_loss:4.6394 train_time:17876ms step_avg:152.79ms
step:128/3125 train_loss:4.6799 train_time:18027ms step_avg:152.77ms
step:129/3125 train_loss:4.5889 train_time:18178ms step_avg:152.76ms
step:130/3125 train_loss:4.8742 train_time:18328ms step_avg:152.74ms
step:131/3125 train_loss:4.5761 train_time:18479ms step_avg:152.72ms
step:132/3125 train_loss:4.6035 train_time:18635ms step_avg:152.75ms
step:133/3125 train_loss:4.5456 train_time:18792ms step_avg:152.78ms
step:134/3125 train_loss:4.6294 train_time:18946ms step_avg:152.79ms
step:135/3125 train_loss:4.4692 train_time:19099ms step_avg:152.79ms
step:136/3125 train_loss:4.6201 train_time:19249ms step_avg:152.77ms
step:137/3125 train_loss:4.3998 train_time:19401ms step_avg:152.76ms
step:138/3125 train_loss:4.5667 train_time:19553ms step_avg:152.76ms
step:139/3125 train_loss:4.4729 train_time:19706ms step_avg:152.76ms
step:140/3125 train_loss:4.5560 train_time:19861ms step_avg:152.78ms
step:141/3125 train_loss:4.6299 train_time:20014ms step_avg:152.78ms
step:142/3125 train_loss:4.4966 train_time:20167ms step_avg:152.78ms
step:143/3125 train_loss:4.4827 train_time:20320ms step_avg:152.78ms
step:144/3125 train_loss:4.3968 train_time:20471ms step_avg:152.77ms
step:145/3125 train_loss:4.5205 train_time:20624ms step_avg:152.77ms
step:146/3125 train_loss:4.4669 train_time:20778ms step_avg:152.78ms
step:147/3125 train_loss:4.3538 train_time:20930ms step_avg:152.78ms
step:148/3125 train_loss:4.4781 train_time:21084ms step_avg:152.78ms
step:149/3125 train_loss:4.5100 train_time:21236ms step_avg:152.78ms
step:150/3125 train_loss:4.4599 train_time:21389ms step_avg:152.78ms
step:151/3125 train_loss:4.5691 train_time:21540ms step_avg:152.77ms
step:152/3125 train_loss:4.4222 train_time:21694ms step_avg:152.78ms
step:153/3125 train_loss:4.4187 train_time:21846ms step_avg:152.77ms
step:154/3125 train_loss:4.4950 train_time:22000ms step_avg:152.78ms
step:155/3125 train_loss:4.4815 train_time:22153ms step_avg:152.78ms
step:156/3125 train_loss:4.4156 train_time:22304ms step_avg:152.77ms
step:157/3125 train_loss:4.4667 train_time:22457ms step_avg:152.77ms
step:158/3125 train_loss:4.5437 train_time:22608ms step_avg:152.76ms
step:159/3125 train_loss:4.3693 train_time:22762ms step_avg:152.76ms
step:160/3125 train_loss:4.4364 train_time:22914ms step_avg:152.76ms
step:161/3125 train_loss:4.2521 train_time:23067ms step_avg:152.76ms
step:162/3125 train_loss:4.4643 train_time:23221ms step_avg:152.77ms
step:163/3125 train_loss:4.4756 train_time:23374ms step_avg:152.77ms
step:164/3125 train_loss:4.4570 train_time:23525ms step_avg:152.76ms
step:165/3125 train_loss:4.3213 train_time:23679ms step_avg:152.77ms
step:166/3125 train_loss:4.3986 train_time:23831ms step_avg:152.76ms
step:167/3125 train_loss:4.4666 train_time:23984ms step_avg:152.76ms
step:168/3125 train_loss:4.3012 train_time:24137ms step_avg:152.77ms
step:169/3125 train_loss:4.3786 train_time:24290ms step_avg:152.77ms
step:170/3125 train_loss:4.2761 train_time:24442ms step_avg:152.77ms
step:171/3125 train_loss:4.1504 train_time:24597ms step_avg:152.78ms
step:172/3125 train_loss:4.3073 train_time:24749ms step_avg:152.77ms
step:173/3125 train_loss:4.3188 train_time:24902ms step_avg:152.77ms
step:174/3125 train_loss:4.3681 train_time:25055ms step_avg:152.77ms
step:175/3125 train_loss:4.5277 train_time:25207ms step_avg:152.77ms
step:176/3125 train_loss:4.3584 train_time:25362ms step_avg:152.78ms
step:177/3125 train_loss:4.2212 train_time:25515ms step_avg:152.78ms
step:178/3125 train_loss:4.1816 train_time:25667ms step_avg:152.78ms
step:179/3125 train_loss:4.2823 train_time:25821ms step_avg:152.79ms
step:180/3125 train_loss:4.2387 train_time:25974ms step_avg:152.79ms
step:181/3125 train_loss:4.2159 train_time:26126ms step_avg:152.78ms
step:182/3125 train_loss:4.3881 train_time:26279ms step_avg:152.79ms
step:183/3125 train_loss:4.2616 train_time:26432ms step_avg:152.79ms
step:184/3125 train_loss:4.2409 train_time:26584ms step_avg:152.78ms
step:185/3125 train_loss:4.2292 train_time:26737ms step_avg:152.78ms
step:186/3125 train_loss:4.3161 train_time:26890ms step_avg:152.78ms
step:187/3125 train_loss:4.2897 train_time:27043ms step_avg:152.78ms
step:188/3125 train_loss:4.3509 train_time:27198ms step_avg:152.80ms
step:189/3125 train_loss:4.2839 train_time:27463ms step_avg:153.42ms
step:190/3125 train_loss:4.2111 train_time:27751ms step_avg:154.17ms
step:191/3125 train_loss:4.3091 train_time:27902ms step_avg:154.15ms
step:192/3125 train_loss:4.1810 train_time:28053ms step_avg:154.13ms
step:193/3125 train_loss:4.1217 train_time:28203ms step_avg:154.11ms
step:194/3125 train_loss:4.3402 train_time:28355ms step_avg:154.10ms
step:195/3125 train_loss:4.2551 train_time:28505ms step_avg:154.08ms
step:196/3125 train_loss:4.4594 train_time:28662ms step_avg:154.10ms
step:197/3125 train_loss:4.2902 train_time:28820ms step_avg:154.12ms
step:198/3125 train_loss:4.1322 train_time:28973ms step_avg:154.11ms
step:199/3125 train_loss:4.2683 train_time:29125ms step_avg:154.10ms
step:200/3125 train_loss:4.1294 train_time:29277ms step_avg:154.09ms
step:201/3125 train_loss:4.2196 train_time:29429ms step_avg:154.08ms
step:202/3125 train_loss:4.0931 train_time:29582ms step_avg:154.07ms
step:203/3125 train_loss:4.3334 train_time:29737ms step_avg:154.08ms
step:204/3125 train_loss:4.1688 train_time:29892ms step_avg:154.08ms
step:205/3125 train_loss:4.2815 train_time:30046ms step_avg:154.08ms
step:206/3125 train_loss:4.3343 train_time:30199ms step_avg:154.07ms
step:207/3125 train_loss:4.0381 train_time:30351ms step_avg:154.06ms
step:208/3125 train_loss:4.1824 train_time:30503ms step_avg:154.06ms
step:209/3125 train_loss:4.1751 train_time:30657ms step_avg:154.05ms
step:210/3125 train_loss:4.3260 train_time:30809ms step_avg:154.05ms
step:211/3125 train_loss:4.2605 train_time:30964ms step_avg:154.05ms
step:212/3125 train_loss:4.1554 train_time:31118ms step_avg:154.05ms
step:213/3125 train_loss:4.1698 train_time:31272ms step_avg:154.05ms
step:214/3125 train_loss:4.1329 train_time:31425ms step_avg:154.04ms
step:215/3125 train_loss:4.2028 train_time:31578ms step_avg:154.04ms
step:216/3125 train_loss:4.0256 train_time:31731ms step_avg:154.03ms
step:217/3125 train_loss:4.0843 train_time:31883ms step_avg:154.02ms
step:218/3125 train_loss:4.0890 train_time:32037ms step_avg:154.03ms
step:219/3125 train_loss:4.1714 train_time:32190ms step_avg:154.02ms
step:220/3125 train_loss:4.1603 train_time:32342ms step_avg:154.01ms
step:221/3125 train_loss:4.1684 train_time:32495ms step_avg:154.01ms
step:222/3125 train_loss:4.1961 train_time:32646ms step_avg:153.99ms
step:223/3125 train_loss:4.1016 train_time:32800ms step_avg:153.99ms
step:224/3125 train_loss:4.0603 train_time:32952ms step_avg:153.98ms
step:225/3125 train_loss:4.3720 train_time:33105ms step_avg:153.98ms
step:226/3125 train_loss:3.9910 train_time:33259ms step_avg:153.98ms
step:227/3125 train_loss:4.0650 train_time:33413ms step_avg:153.98ms
step:228/3125 train_loss:4.0734 train_time:33565ms step_avg:153.97ms
step:229/3125 train_loss:4.2206 train_time:33718ms step_avg:153.96ms
step:230/3125 train_loss:4.0105 train_time:33869ms step_avg:153.95ms
step:231/3125 train_loss:4.1350 train_time:34023ms step_avg:153.95ms
step:232/3125 train_loss:3.9911 train_time:34176ms step_avg:153.95ms
step:233/3125 train_loss:4.0595 train_time:34329ms step_avg:153.94ms
step:234/3125 train_loss:4.1844 train_time:34482ms step_avg:153.94ms
step:235/3125 train_loss:4.1051 train_time:34634ms step_avg:153.93ms
step:236/3125 train_loss:3.9907 train_time:34787ms step_avg:153.92ms
step:237/3125 train_loss:4.1527 train_time:34939ms step_avg:153.92ms
step:238/3125 train_loss:4.1634 train_time:35093ms step_avg:153.92ms
step:239/3125 train_loss:4.0258 train_time:35247ms step_avg:153.92ms
step:240/3125 train_loss:4.1590 train_time:35400ms step_avg:153.91ms
step:241/3125 train_loss:4.1925 train_time:35552ms step_avg:153.91ms
step:242/3125 train_loss:4.0543 train_time:35704ms step_avg:153.90ms
step:243/3125 train_loss:4.2208 train_time:35857ms step_avg:153.89ms
step:244/3125 train_loss:4.1026 train_time:36009ms step_avg:153.89ms
step:245/3125 train_loss:4.1576 train_time:36164ms step_avg:153.89ms
step:246/3125 train_loss:4.2304 train_time:36316ms step_avg:153.88ms
step:247/3125 train_loss:4.1454 train_time:36469ms step_avg:153.88ms
step:248/3125 train_loss:4.0902 train_time:36621ms step_avg:153.87ms
step:249/3125 train_loss:4.1949 train_time:36774ms step_avg:153.87ms
step:250/3125 train_loss:4.0062 train_time:36928ms step_avg:153.87ms
step:250/3125 val_loss:4.0891 train_time:36978ms step_avg:154.07ms
step:251/3125 train_loss:4.0506 train_time:37086ms step_avg:153.88ms
step:252/3125 train_loss:4.1582 train_time:37239ms step_avg:153.88ms
step:253/3125 train_loss:4.2259 train_time:37393ms step_avg:153.88ms
step:254/3125 train_loss:4.0141 train_time:37544ms step_avg:153.87ms
step:255/3125 train_loss:3.9551 train_time:37695ms step_avg:153.86ms
step:256/3125 train_loss:4.1475 train_time:37846ms step_avg:153.84ms
step:257/3125 train_loss:4.0586 train_time:38001ms step_avg:153.85ms
step:258/3125 train_loss:4.0613 train_time:38157ms step_avg:153.86ms
step:259/3125 train_loss:4.0487 train_time:38312ms step_avg:153.86ms
step:260/3125 train_loss:4.1038 train_time:38464ms step_avg:153.86ms
step:261/3125 train_loss:4.1327 train_time:38616ms step_avg:153.85ms
step:262/3125 train_loss:4.0982 train_time:38768ms step_avg:153.84ms
step:263/3125 train_loss:4.0664 train_time:38920ms step_avg:153.83ms
step:264/3125 train_loss:3.9791 train_time:39074ms step_avg:153.84ms
step:265/3125 train_loss:4.0681 train_time:39230ms step_avg:153.84ms
step:266/3125 train_loss:3.9388 train_time:39382ms step_avg:153.84ms
step:267/3125 train_loss:3.9913 train_time:39535ms step_avg:153.83ms
step:268/3125 train_loss:4.0002 train_time:39686ms step_avg:153.82ms
step:269/3125 train_loss:4.0295 train_time:39838ms step_avg:153.81ms
step:270/3125 train_loss:3.9338 train_time:39991ms step_avg:153.81ms
step:271/3125 train_loss:4.1785 train_time:40144ms step_avg:153.81ms
step:272/3125 train_loss:4.0551 train_time:40298ms step_avg:153.81ms
step:273/3125 train_loss:3.9859 train_time:40452ms step_avg:153.81ms
step:274/3125 train_loss:4.0325 train_time:40605ms step_avg:153.81ms
step:275/3125 train_loss:4.1041 train_time:40755ms step_avg:153.79ms
step:276/3125 train_loss:4.1350 train_time:40908ms step_avg:153.79ms
step:277/3125 train_loss:4.3022 train_time:41061ms step_avg:153.79ms
step:278/3125 train_loss:4.1009 train_time:41214ms step_avg:153.78ms
step:279/3125 train_loss:4.1594 train_time:41366ms step_avg:153.78ms
step:280/3125 train_loss:4.0655 train_time:41519ms step_avg:153.77ms
step:281/3125 train_loss:4.2077 train_time:41671ms step_avg:153.77ms
step:282/3125 train_loss:4.0303 train_time:41824ms step_avg:153.76ms
step:283/3125 train_loss:4.0169 train_time:41975ms step_avg:153.76ms
step:284/3125 train_loss:3.9859 train_time:42129ms step_avg:153.76ms
step:285/3125 train_loss:4.1345 train_time:42282ms step_avg:153.75ms
step:286/3125 train_loss:4.1376 train_time:42435ms step_avg:153.75ms
step:287/3125 train_loss:4.1666 train_time:42588ms step_avg:153.75ms
step:288/3125 train_loss:3.9887 train_time:42755ms step_avg:153.79ms
step:289/3125 train_loss:4.0869 train_time:42894ms step_avg:153.74ms
step:290/3125 train_loss:3.9415 train_time:43047ms step_avg:153.74ms
step:291/3125 train_loss:3.9426 train_time:43198ms step_avg:153.73ms
step:292/3125 train_loss:4.0083 train_time:43352ms step_avg:153.73ms
step:293/3125 train_loss:3.9413 train_time:43505ms step_avg:153.73ms
step:294/3125 train_loss:3.9859 train_time:43656ms step_avg:153.72ms
step:295/3125 train_loss:4.0276 train_time:43810ms step_avg:153.72ms
step:296/3125 train_loss:3.9105 train_time:43962ms step_avg:153.71ms
step:297/3125 train_loss:3.9335 train_time:44115ms step_avg:153.71ms
step:298/3125 train_loss:3.9310 train_time:44269ms step_avg:153.71ms
step:299/3125 train_loss:4.0412 train_time:44421ms step_avg:153.71ms
step:300/3125 train_loss:3.9015 train_time:44575ms step_avg:153.71ms
step:301/3125 train_loss:4.0362 train_time:44728ms step_avg:153.71ms
step:302/3125 train_loss:4.0515 train_time:44880ms step_avg:153.70ms
step:303/3125 train_loss:4.0011 train_time:45034ms step_avg:153.70ms
step:304/3125 train_loss:4.0491 train_time:45188ms step_avg:153.70ms
step:305/3125 train_loss:4.0329 train_time:45339ms step_avg:153.69ms
step:306/3125 train_loss:4.5284 train_time:45493ms step_avg:153.69ms
step:307/3125 train_loss:4.0074 train_time:45646ms step_avg:153.69ms
step:308/3125 train_loss:3.9175 train_time:45798ms step_avg:153.68ms
step:309/3125 train_loss:4.0571 train_time:45951ms step_avg:153.68ms
step:310/3125 train_loss:3.9389 train_time:46104ms step_avg:153.68ms
step:311/3125 train_loss:4.1604 train_time:46256ms step_avg:153.68ms
step:312/3125 train_loss:3.9973 train_time:46410ms step_avg:153.68ms
step:313/3125 train_loss:3.9412 train_time:46564ms step_avg:153.68ms
step:314/3125 train_loss:4.0273 train_time:46716ms step_avg:153.67ms
step:315/3125 train_loss:4.1527 train_time:46869ms step_avg:153.67ms
step:316/3125 train_loss:4.0261 train_time:47022ms step_avg:153.67ms
step:317/3125 train_loss:3.8674 train_time:47174ms step_avg:153.66ms
step:318/3125 train_loss:3.9491 train_time:47327ms step_avg:153.66ms
step:319/3125 train_loss:3.9919 train_time:47479ms step_avg:153.65ms
step:320/3125 train_loss:3.9635 train_time:47632ms step_avg:153.65ms
step:321/3125 train_loss:4.0795 train_time:47783ms step_avg:153.64ms
step:322/3125 train_loss:4.0218 train_time:47936ms step_avg:153.64ms
step:323/3125 train_loss:3.9958 train_time:48090ms step_avg:153.64ms
step:324/3125 train_loss:4.0816 train_time:48241ms step_avg:153.64ms
step:325/3125 train_loss:4.0211 train_time:48395ms step_avg:153.64ms
step:326/3125 train_loss:4.0953 train_time:48549ms step_avg:153.63ms
step:327/3125 train_loss:3.9555 train_time:48700ms step_avg:153.63ms
step:328/3125 train_loss:4.4681 train_time:48853ms step_avg:153.63ms
step:329/3125 train_loss:4.1446 train_time:49006ms step_avg:153.63ms
step:330/3125 train_loss:3.8813 train_time:49159ms step_avg:153.62ms
step:331/3125 train_loss:3.8321 train_time:49313ms step_avg:153.62ms
step:332/3125 train_loss:4.0541 train_time:49467ms step_avg:153.62ms
step:333/3125 train_loss:3.9808 train_time:49619ms step_avg:153.62ms
step:334/3125 train_loss:3.9419 train_time:49772ms step_avg:153.62ms
step:335/3125 train_loss:3.9128 train_time:49926ms step_avg:153.62ms
step:336/3125 train_loss:4.0895 train_time:50077ms step_avg:153.61ms
step:337/3125 train_loss:4.0317 train_time:50231ms step_avg:153.61ms
step:338/3125 train_loss:4.4952 train_time:50384ms step_avg:153.61ms
step:339/3125 train_loss:4.0152 train_time:50536ms step_avg:153.60ms
step:340/3125 train_loss:3.9528 train_time:50690ms step_avg:153.61ms
step:341/3125 train_loss:4.0082 train_time:50842ms step_avg:153.60ms
step:342/3125 train_loss:3.9268 train_time:50997ms step_avg:153.60ms
step:343/3125 train_loss:3.8975 train_time:51151ms step_avg:153.61ms
step:344/3125 train_loss:3.9192 train_time:51304ms step_avg:153.60ms
step:345/3125 train_loss:4.0665 train_time:51456ms step_avg:153.60ms
step:346/3125 train_loss:3.9087 train_time:51609ms step_avg:153.60ms
step:347/3125 train_loss:3.8442 train_time:51761ms step_avg:153.59ms
step:348/3125 train_loss:3.8810 train_time:51915ms step_avg:153.59ms
step:349/3125 train_loss:3.9299 train_time:52070ms step_avg:153.60ms
step:350/3125 train_loss:3.8974 train_time:52222ms step_avg:153.59ms
step:351/3125 train_loss:3.6387 train_time:52374ms step_avg:153.59ms
step:352/3125 train_loss:3.9012 train_time:52527ms step_avg:153.59ms
step:353/3125 train_loss:4.2410 train_time:52679ms step_avg:153.58ms
step:354/3125 train_loss:3.7344 train_time:52832ms step_avg:153.58ms
step:355/3125 train_loss:4.0016 train_time:52985ms step_avg:153.58ms
step:356/3125 train_loss:3.8580 train_time:53138ms step_avg:153.58ms
step:357/3125 train_loss:3.9634 train_time:53291ms step_avg:153.58ms
step:358/3125 train_loss:3.8900 train_time:53444ms step_avg:153.57ms
step:359/3125 train_loss:3.9189 train_time:53597ms step_avg:153.57ms
step:360/3125 train_loss:3.9230 train_time:53751ms step_avg:153.57ms
step:361/3125 train_loss:3.5190 train_time:53906ms step_avg:153.58ms
step:362/3125 train_loss:4.0960 train_time:54057ms step_avg:153.57ms
step:363/3125 train_loss:4.0032 train_time:54210ms step_avg:153.57ms
step:364/3125 train_loss:3.9230 train_time:54362ms step_avg:153.56ms
step:365/3125 train_loss:3.8203 train_time:54515ms step_avg:153.56ms
step:366/3125 train_loss:3.9874 train_time:54668ms step_avg:153.56ms
step:367/3125 train_loss:3.9438 train_time:54823ms step_avg:153.57ms
step:368/3125 train_loss:3.9314 train_time:54976ms step_avg:153.56ms