-
Notifications
You must be signed in to change notification settings - Fork 148
/
cf9e4571-c5fc-4323-abf3-a98d862ec6c8.txt
2406 lines (2335 loc) · 151 KB
/
cf9e4571-c5fc-4323-abf3-a98d862ec6c8.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
# 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)
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.dim = dim
self.base = base
self.inv_freq = None
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.inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=x.device).float() / self.dim))
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)
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 CastedLinear(nn.Linear):
def forward(self, x):
return F.linear(x, self.weight.to(x.dtype))
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 = CastedLinear(self.n_embd, self.n_embd, bias=False)
self.c_k = CastedLinear(self.n_embd, self.n_embd, bias=False)
self.c_v = CastedLinear(self.n_embd, self.n_embd, bias=False)
# output projection
self.c_proj = CastedLinear(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.lamb = nn.Parameter(torch.tensor(0.5)) # @Grad62304977
def forward(self, x, v1, block_mask):
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 = (1 - self.lamb) * v + self.lamb * 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 = 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, v1
class MLP(nn.Module):
def __init__(self, config):
super().__init__()
self.c_fc = CastedLinear(config.n_embd, 4 * config.n_embd, bias=False)
self.c_proj = CastedLinear(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, block_mask):
x = self.lambdas[0] * x + self.lambdas[1] * x0
x1, v1 = self.attn(F.rms_norm(x, (x.size(-1),)), v1, block_mask)
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__()
# 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),
h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
))
self.lm_head = CastedLinear(config.n_embd, config.vocab_size, bias=False)
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 = F.rms_norm(x, (x.size(-1),)) # @Grad62304977
x0 = x
v1 = None
# 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, v1 = self.transformer.h[i](x, v1, 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, v1 = self.transformer.h[self.num_encoder_layers + i](x, v1, x0, block_mask)
x = F.rms_norm(x, (x.size(-1),))
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, 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
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.B * self.T
self.tokens = _load_data_shard(self.files[self.current_shard])
def next_batch(self):
batch_size = self.B * self.T * self.num_processes
buf = self.tokens[self.current_position:self.current_position+self.B*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
device_batch_size : int = 1 # batch size, in sequences, per device
sequence_length : int = 64*1024 # sequence length, in tokens
num_iterations : int = 1750 # number of iterations to run
warmup_iters : int = 0
cooldown_iters : int = 640 # 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('='*100 + '\n')
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
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)
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
# 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=0.6, 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)
# Set the attention blocksize for the current step, in chunks of 64
attn_blocksize = torch.tensor(64*((step/args.num_iterations * (1792 - 64) + 64)//64), dtype=torch.int, device='cuda')
# 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):
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):
# forward pass
loss = model(x, y, attn_blocksize=attn_blocksize)
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/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.dev20241124+cu124 compiled for CUDA 12.4
nvidia-smi:
Sun Nov 24 23:36:17 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 555.42.06 Driver Version: 555.42.06 CUDA Version: 12.5 |
|-----------------------------------------+------------------------+----------------------+
| 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 28C P0 68W / 700W | 4MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 1 NVIDIA H100 80GB HBM3 Off | 00000000:2A:00.0 Off | 0 |
| N/A 31C P0 114W / 700W | 34MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 2 NVIDIA H100 80GB HBM3 Off | 00000000:3A:00.0 Off | 0 |
| N/A 32C P0 111W / 700W | 530MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 3 NVIDIA H100 80GB HBM3 Off | 00000000:5D:00.0 Off | 0 |
| N/A 29C P0 112W / 700W | 530MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 4 NVIDIA H100 80GB HBM3 Off | 00000000:84:00.0 Off | 0 |
| N/A 29C P0 111W / 700W | 530MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 5 NVIDIA H100 80GB HBM3 Off | 00000000:8B:00.0 Off | 0 |
| N/A 32C P0 113W / 700W | 530MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 6 NVIDIA H100 80GB HBM3 Off | 00000000:91:00.0 Off | 0 |
| N/A 30C P0 110W / 700W | 530MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 7 NVIDIA H100 80GB HBM3 Off | 00000000:E4:00.0 Off | 0 |
| N/A 29C P0 114W / 700W | 530MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 1 N/A N/A 1796 C /usr/bin/python3 0MiB |
| 2 N/A N/A 1797 C /usr/bin/python3 0MiB |
| 3 N/A N/A 1798 C /usr/bin/python3 0MiB |
| 4 N/A N/A 1799 C /usr/bin/python3 0MiB |
| 5 N/A N/A 1800 C /usr/bin/python3 0MiB |
| 6 N/A N/A 1801 C /usr/bin/python3 0MiB |
| 7 N/A N/A 1802 C /usr/bin/python3 0MiB |
+-----------------------------------------------------------------------------------------+
====================================================================================================
Training DataLoader: total number of tokens: 1800000000 across 18 files
Validation DataLoader: total number of tokens: 100000000 across 1 files
====================================================================================================
step:0/1750 val_loss:10.8258 train_time:0ms step_avg:nanms
step:1/1750 train_loss:10.8258 train_time:163438ms step_avg:nanms
step:2/1750 train_loss:10.0893 train_time:163548ms step_avg:nanms
step:3/1750 train_loss:8.3951 train_time:163693ms step_avg:nanms
step:4/1750 train_loss:7.5780 train_time:163842ms step_avg:nanms
step:5/1750 train_loss:7.4681 train_time:163991ms step_avg:nanms
step:6/1750 train_loss:6.9705 train_time:164138ms step_avg:nanms
step:7/1750 train_loss:7.2105 train_time:164285ms step_avg:nanms
step:8/1750 train_loss:6.7367 train_time:164434ms step_avg:nanms
step:9/1750 train_loss:6.6213 train_time:164581ms step_avg:nanms
step:10/1750 train_loss:6.5087 train_time:164728ms step_avg:nanms
step:11/1750 train_loss:6.4831 train_time:110ms step_avg:nanms
step:12/1750 train_loss:6.3491 train_time:257ms step_avg:nanms
step:13/1750 train_loss:6.2939 train_time:407ms step_avg:135.58ms
step:14/1750 train_loss:6.2331 train_time:553ms step_avg:138.33ms
step:15/1750 train_loss:6.2026 train_time:701ms step_avg:140.24ms
step:16/1750 train_loss:6.1480 train_time:849ms step_avg:141.52ms
step:17/1750 train_loss:6.2293 train_time:996ms step_avg:142.24ms
step:18/1750 train_loss:6.0134 train_time:1145ms step_avg:143.09ms
step:19/1750 train_loss:6.0690 train_time:1292ms step_avg:143.54ms
step:20/1750 train_loss:5.7139 train_time:1440ms step_avg:143.97ms
step:21/1750 train_loss:6.0282 train_time:1588ms step_avg:144.36ms
step:22/1750 train_loss:6.2938 train_time:1736ms step_avg:144.65ms
step:23/1750 train_loss:5.9283 train_time:1884ms step_avg:144.90ms
step:24/1750 train_loss:6.0908 train_time:2032ms step_avg:145.13ms
step:25/1750 train_loss:5.7889 train_time:2180ms step_avg:145.32ms
step:26/1750 train_loss:5.7179 train_time:2330ms step_avg:145.60ms
step:27/1750 train_loss:5.9059 train_time:2476ms step_avg:145.67ms
step:28/1750 train_loss:5.5065 train_time:2626ms step_avg:145.89ms
step:29/1750 train_loss:5.7777 train_time:2774ms step_avg:145.98ms
step:30/1750 train_loss:5.5997 train_time:2922ms step_avg:146.11ms
step:31/1750 train_loss:5.5644 train_time:3070ms step_avg:146.17ms
step:32/1750 train_loss:5.4084 train_time:3216ms step_avg:146.18ms
step:33/1750 train_loss:5.6910 train_time:3365ms step_avg:146.30ms
step:34/1750 train_loss:5.5944 train_time:3513ms step_avg:146.36ms
step:35/1750 train_loss:5.7210 train_time:3660ms step_avg:146.39ms
step:36/1750 train_loss:5.6432 train_time:3809ms step_avg:146.49ms
step:37/1750 train_loss:5.5441 train_time:3956ms step_avg:146.51ms
step:38/1750 train_loss:5.4215 train_time:4105ms step_avg:146.59ms
step:39/1750 train_loss:5.4423 train_time:4252ms step_avg:146.62ms
step:40/1750 train_loss:5.3416 train_time:4399ms step_avg:146.62ms
step:41/1750 train_loss:5.3366 train_time:4547ms step_avg:146.68ms
step:42/1750 train_loss:5.2762 train_time:4693ms step_avg:146.67ms
step:43/1750 train_loss:5.3978 train_time:4841ms step_avg:146.71ms
step:44/1750 train_loss:5.3553 train_time:4989ms step_avg:146.74ms
step:45/1750 train_loss:5.4858 train_time:5137ms step_avg:146.77ms
step:46/1750 train_loss:5.2656 train_time:5284ms step_avg:146.79ms
step:47/1750 train_loss:5.1761 train_time:5432ms step_avg:146.82ms
step:48/1750 train_loss:5.3076 train_time:5580ms step_avg:146.83ms
step:49/1750 train_loss:5.2348 train_time:5729ms step_avg:146.88ms
step:50/1750 train_loss:5.3447 train_time:5876ms step_avg:146.90ms
step:51/1750 train_loss:5.2615 train_time:6025ms step_avg:146.94ms
step:52/1750 train_loss:5.1171 train_time:6172ms step_avg:146.95ms
step:53/1750 train_loss:5.2673 train_time:6319ms step_avg:146.95ms
step:54/1750 train_loss:5.1178 train_time:6468ms step_avg:146.99ms
step:55/1750 train_loss:5.4847 train_time:6616ms step_avg:147.02ms
step:56/1750 train_loss:5.1145 train_time:6764ms step_avg:147.04ms
step:57/1750 train_loss:4.9826 train_time:6911ms step_avg:147.05ms
step:58/1750 train_loss:5.1019 train_time:7057ms step_avg:147.03ms
step:59/1750 train_loss:5.1164 train_time:7207ms step_avg:147.08ms
step:60/1750 train_loss:5.2303 train_time:7354ms step_avg:147.08ms
step:61/1750 train_loss:4.9726 train_time:7501ms step_avg:147.08ms
step:62/1750 train_loss:5.0816 train_time:7649ms step_avg:147.10ms
step:63/1750 train_loss:5.0547 train_time:7795ms step_avg:147.08ms
step:64/1750 train_loss:4.9510 train_time:7944ms step_avg:147.11ms
step:65/1750 train_loss:4.9034 train_time:8091ms step_avg:147.11ms
step:66/1750 train_loss:5.0774 train_time:8239ms step_avg:147.13ms
step:67/1750 train_loss:4.9320 train_time:8387ms step_avg:147.14ms
step:68/1750 train_loss:5.1849 train_time:8534ms step_avg:147.14ms
step:69/1750 train_loss:4.8183 train_time:8681ms step_avg:147.13ms
step:70/1750 train_loss:4.9104 train_time:8829ms step_avg:147.16ms
step:71/1750 train_loss:5.0624 train_time:8976ms step_avg:147.15ms
step:72/1750 train_loss:4.9887 train_time:9124ms step_avg:147.17ms
step:73/1750 train_loss:4.8633 train_time:9271ms step_avg:147.16ms
step:74/1750 train_loss:4.9941 train_time:9418ms step_avg:147.16ms
step:75/1750 train_loss:4.9696 train_time:9566ms step_avg:147.17ms
step:76/1750 train_loss:4.9045 train_time:9714ms step_avg:147.17ms
step:77/1750 train_loss:5.0199 train_time:9860ms step_avg:147.17ms
step:78/1750 train_loss:5.2016 train_time:10009ms step_avg:147.19ms
step:79/1750 train_loss:4.9216 train_time:10156ms step_avg:147.19ms
step:80/1750 train_loss:4.9497 train_time:10304ms step_avg:147.20ms
step:81/1750 train_loss:4.7405 train_time:10452ms step_avg:147.22ms
step:82/1750 train_loss:4.8915 train_time:10599ms step_avg:147.22ms
step:83/1750 train_loss:4.8516 train_time:10748ms step_avg:147.23ms
step:84/1750 train_loss:4.8451 train_time:10895ms step_avg:147.23ms
step:85/1750 train_loss:4.6940 train_time:11044ms step_avg:147.25ms
step:86/1750 train_loss:4.9021 train_time:11191ms step_avg:147.25ms
step:87/1750 train_loss:4.8275 train_time:11338ms step_avg:147.25ms
step:88/1750 train_loss:4.8348 train_time:11486ms step_avg:147.25ms
step:89/1750 train_loss:4.7845 train_time:11633ms step_avg:147.26ms
step:90/1750 train_loss:4.7102 train_time:11781ms step_avg:147.26ms
step:91/1750 train_loss:4.7036 train_time:11928ms step_avg:147.27ms
step:92/1750 train_loss:4.8486 train_time:12076ms step_avg:147.27ms
step:93/1750 train_loss:4.6607 train_time:12224ms step_avg:147.28ms
step:94/1750 train_loss:4.7021 train_time:12371ms step_avg:147.27ms
step:95/1750 train_loss:4.7476 train_time:12517ms step_avg:147.26ms
step:96/1750 train_loss:4.6388 train_time:12666ms step_avg:147.28ms
step:97/1750 train_loss:4.6843 train_time:12813ms step_avg:147.28ms
step:98/1750 train_loss:4.6344 train_time:12960ms step_avg:147.27ms
step:99/1750 train_loss:4.7371 train_time:13109ms step_avg:147.29ms
step:100/1750 train_loss:4.7256 train_time:13256ms step_avg:147.29ms
step:101/1750 train_loss:4.5779 train_time:13404ms step_avg:147.30ms
step:102/1750 train_loss:4.7521 train_time:13551ms step_avg:147.30ms
step:103/1750 train_loss:4.6428 train_time:13698ms step_avg:147.29ms
step:104/1750 train_loss:4.5703 train_time:13847ms step_avg:147.31ms
step:105/1750 train_loss:4.6017 train_time:13993ms step_avg:147.30ms
step:106/1750 train_loss:4.6746 train_time:14142ms step_avg:147.31ms
step:107/1750 train_loss:4.5616 train_time:14289ms step_avg:147.31ms
step:108/1750 train_loss:4.3978 train_time:14437ms step_avg:147.32ms
step:109/1750 train_loss:4.5401 train_time:14585ms step_avg:147.32ms
step:110/1750 train_loss:4.5238 train_time:14732ms step_avg:147.32ms
step:111/1750 train_loss:4.4611 train_time:14879ms step_avg:147.32ms
step:112/1750 train_loss:4.6176 train_time:15027ms step_avg:147.32ms
step:113/1750 train_loss:4.5175 train_time:15175ms step_avg:147.33ms
step:114/1750 train_loss:4.3843 train_time:15323ms step_avg:147.34ms
step:115/1750 train_loss:4.5378 train_time:15470ms step_avg:147.34ms
step:116/1750 train_loss:4.5121 train_time:15617ms step_avg:147.33ms
step:117/1750 train_loss:4.4269 train_time:15765ms step_avg:147.34ms
step:118/1750 train_loss:4.6516 train_time:15915ms step_avg:147.36ms
step:119/1750 train_loss:4.5109 train_time:16061ms step_avg:147.35ms
step:120/1750 train_loss:4.4046 train_time:16210ms step_avg:147.37ms
step:121/1750 train_loss:4.3465 train_time:16357ms step_avg:147.36ms
step:122/1750 train_loss:4.4887 train_time:16505ms step_avg:147.37ms
step:123/1750 train_loss:4.3337 train_time:16652ms step_avg:147.37ms
step:124/1750 train_loss:4.6312 train_time:16800ms step_avg:147.36ms
step:125/1750 train_loss:4.5231 train_time:16948ms step_avg:147.37ms
step:125/1750 val_loss:4.4538 train_time:16985ms step_avg:147.70ms
step:126/1750 train_loss:4.4651 train_time:17096ms step_avg:147.38ms
step:127/1750 train_loss:4.4789 train_time:17246ms step_avg:147.41ms
step:128/1750 train_loss:4.4198 train_time:17393ms step_avg:147.40ms
step:129/1750 train_loss:4.7347 train_time:17541ms step_avg:147.40ms
step:130/1750 train_loss:4.4163 train_time:17689ms step_avg:147.41ms
step:131/1750 train_loss:4.4430 train_time:17838ms step_avg:147.42ms
step:132/1750 train_loss:4.3889 train_time:17990ms step_avg:147.46ms
step:133/1750 train_loss:4.4932 train_time:18139ms step_avg:147.47ms
step:134/1750 train_loss:4.2932 train_time:18290ms step_avg:147.50ms
step:135/1750 train_loss:4.4814 train_time:18441ms step_avg:147.53ms
step:136/1750 train_loss:4.2469 train_time:18591ms step_avg:147.55ms
step:137/1750 train_loss:4.4101 train_time:18741ms step_avg:147.57ms
step:138/1750 train_loss:4.3273 train_time:18892ms step_avg:147.59ms
step:139/1750 train_loss:4.4063 train_time:19043ms step_avg:147.62ms
step:140/1750 train_loss:4.5022 train_time:19194ms step_avg:147.64ms
step:141/1750 train_loss:4.3406 train_time:19344ms step_avg:147.67ms
step:142/1750 train_loss:4.3411 train_time:19494ms step_avg:147.69ms
step:143/1750 train_loss:4.2793 train_time:19646ms step_avg:147.71ms
step:144/1750 train_loss:4.3757 train_time:19795ms step_avg:147.72ms
step:145/1750 train_loss:4.3371 train_time:19946ms step_avg:147.75ms
step:146/1750 train_loss:4.2026 train_time:20097ms step_avg:147.77ms
step:147/1750 train_loss:4.3480 train_time:20250ms step_avg:147.81ms
step:148/1750 train_loss:4.3887 train_time:20400ms step_avg:147.83ms
step:149/1750 train_loss:4.3296 train_time:20552ms step_avg:147.86ms
step:150/1750 train_loss:4.4696 train_time:20703ms step_avg:147.88ms
step:151/1750 train_loss:4.2961 train_time:20854ms step_avg:147.90ms
step:152/1750 train_loss:4.3049 train_time:21004ms step_avg:147.92ms
step:153/1750 train_loss:4.3990 train_time:21154ms step_avg:147.93ms
step:154/1750 train_loss:4.3798 train_time:21305ms step_avg:147.95ms
step:155/1750 train_loss:4.3071 train_time:21455ms step_avg:147.97ms
step:156/1750 train_loss:4.3730 train_time:21606ms step_avg:147.99ms
step:157/1750 train_loss:4.4318 train_time:21757ms step_avg:148.00ms
step:158/1750 train_loss:4.2650 train_time:21907ms step_avg:148.02ms
step:159/1750 train_loss:4.3304 train_time:22058ms step_avg:148.04ms
step:160/1750 train_loss:4.1536 train_time:22208ms step_avg:148.05ms
step:161/1750 train_loss:4.3783 train_time:22358ms step_avg:148.07ms
step:162/1750 train_loss:4.3926 train_time:22509ms step_avg:148.08ms
step:163/1750 train_loss:4.3581 train_time:22660ms step_avg:148.10ms
step:164/1750 train_loss:4.2046 train_time:22811ms step_avg:148.12ms
step:165/1750 train_loss:4.3091 train_time:22961ms step_avg:148.13ms
step:166/1750 train_loss:4.3748 train_time:23111ms step_avg:148.15ms
step:167/1750 train_loss:4.2263 train_time:23260ms step_avg:148.16ms
step:168/1750 train_loss:4.3050 train_time:23411ms step_avg:148.17ms
step:169/1750 train_loss:4.1774 train_time:23561ms step_avg:148.18ms
step:170/1750 train_loss:4.0538 train_time:23712ms step_avg:148.20ms
step:171/1750 train_loss:4.2378 train_time:23862ms step_avg:148.21ms
step:172/1750 train_loss:4.2340 train_time:24012ms step_avg:148.22ms
step:173/1750 train_loss:4.2895 train_time:24164ms step_avg:148.24ms
step:174/1750 train_loss:4.4565 train_time:24313ms step_avg:148.25ms
step:175/1750 train_loss:4.2879 train_time:24464ms step_avg:148.27ms
step:176/1750 train_loss:4.1282 train_time:24614ms step_avg:148.28ms
step:177/1750 train_loss:4.0948 train_time:24765ms step_avg:148.29ms
step:178/1750 train_loss:4.2127 train_time:24914ms step_avg:148.30ms
step:179/1750 train_loss:4.1663 train_time:25066ms step_avg:148.32ms
step:180/1750 train_loss:4.1423 train_time:25215ms step_avg:148.32ms
step:181/1750 train_loss:4.3258 train_time:25367ms step_avg:148.34ms
step:182/1750 train_loss:4.1825 train_time:25516ms step_avg:148.35ms
step:183/1750 train_loss:4.1683 train_time:25667ms step_avg:148.37ms
step:184/1750 train_loss:4.1513 train_time:25817ms step_avg:148.37ms
step:185/1750 train_loss:4.2424 train_time:25968ms step_avg:148.39ms
step:186/1750 train_loss:4.2009 train_time:26118ms step_avg:148.40ms
step:187/1750 train_loss:4.2737 train_time:26269ms step_avg:148.41ms
step:188/1750 train_loss:4.1966 train_time:26536ms step_avg:149.08ms
step:189/1750 train_loss:4.1472 train_time:26837ms step_avg:149.93ms
step:190/1750 train_loss:4.2418 train_time:26990ms step_avg:149.94ms
step:191/1750 train_loss:4.1192 train_time:27141ms step_avg:149.95ms
step:192/1750 train_loss:4.0633 train_time:27291ms step_avg:149.95ms
step:193/1750 train_loss:4.2836 train_time:27442ms step_avg:149.96ms
step:194/1750 train_loss:4.2035 train_time:27592ms step_avg:149.96ms
step:195/1750 train_loss:4.3966 train_time:27743ms step_avg:149.96ms
step:196/1750 train_loss:4.2172 train_time:27893ms step_avg:149.96ms
step:197/1750 train_loss:4.0753 train_time:28044ms step_avg:149.97ms
step:198/1750 train_loss:4.2062 train_time:28194ms step_avg:149.97ms
step:199/1750 train_loss:4.0584 train_time:28344ms step_avg:149.97ms
step:200/1750 train_loss:4.1504 train_time:28493ms step_avg:149.96ms
step:201/1750 train_loss:4.0286 train_time:28642ms step_avg:149.96ms
step:202/1750 train_loss:4.2739 train_time:28792ms step_avg:149.96ms
step:203/1750 train_loss:4.0900 train_time:28941ms step_avg:149.95ms
step:204/1750 train_loss:4.2088 train_time:29090ms step_avg:149.95ms
step:205/1750 train_loss:4.2703 train_time:29240ms step_avg:149.95ms
step:206/1750 train_loss:3.9673 train_time:29389ms step_avg:149.95ms
step:207/1750 train_loss:4.1045 train_time:29538ms step_avg:149.94ms
step:208/1750 train_loss:4.1207 train_time:29689ms step_avg:149.94ms
step:209/1750 train_loss:4.2634 train_time:29838ms step_avg:149.94ms
step:210/1750 train_loss:4.2119 train_time:29988ms step_avg:149.94ms
step:211/1750 train_loss:4.0755 train_time:30164ms step_avg:150.07ms
step:212/1750 train_loss:4.1534 train_time:30326ms step_avg:150.13ms
step:213/1750 train_loss:4.0654 train_time:30474ms step_avg:150.12ms
step:214/1750 train_loss:4.1369 train_time:30624ms step_avg:150.12ms
step:215/1750 train_loss:3.9738 train_time:30773ms step_avg:150.11ms
step:216/1750 train_loss:4.0305 train_time:30923ms step_avg:150.11ms
step:217/1750 train_loss:4.0277 train_time:31072ms step_avg:150.11ms
step:218/1750 train_loss:4.1009 train_time:31222ms step_avg:150.11ms
step:219/1750 train_loss:4.0920 train_time:31371ms step_avg:150.10ms
step:220/1750 train_loss:4.1055 train_time:31521ms step_avg:150.10ms
step:221/1750 train_loss:4.1255 train_time:31671ms step_avg:150.10ms
step:222/1750 train_loss:4.0226 train_time:31820ms step_avg:150.09ms
step:223/1750 train_loss:4.0041 train_time:31970ms step_avg:150.09ms
step:224/1750 train_loss:4.3246 train_time:32118ms step_avg:150.09ms
step:225/1750 train_loss:3.9332 train_time:32268ms step_avg:150.09ms
step:226/1750 train_loss:4.0081 train_time:32417ms step_avg:150.08ms
step:227/1750 train_loss:4.0051 train_time:32567ms step_avg:150.08ms
step:228/1750 train_loss:4.1663 train_time:32715ms step_avg:150.07ms
step:229/1750 train_loss:3.9598 train_time:32865ms step_avg:150.07ms
step:230/1750 train_loss:4.0801 train_time:33013ms step_avg:150.06ms
step:231/1750 train_loss:3.9158 train_time:33163ms step_avg:150.06ms
step:232/1750 train_loss:3.9955 train_time:33312ms step_avg:150.05ms
step:233/1750 train_loss:4.1198 train_time:33461ms step_avg:150.05ms
step:234/1750 train_loss:4.0620 train_time:33611ms step_avg:150.05ms
step:235/1750 train_loss:3.9269 train_time:33761ms step_avg:150.05ms
step:236/1750 train_loss:4.1247 train_time:33910ms step_avg:150.04ms
step:237/1750 train_loss:4.1033 train_time:34059ms step_avg:150.04ms
step:238/1750 train_loss:3.9671 train_time:34208ms step_avg:150.04ms
step:239/1750 train_loss:4.1156 train_time:34358ms step_avg:150.03ms
step:240/1750 train_loss:4.1388 train_time:34508ms step_avg:150.03ms
step:241/1750 train_loss:3.9918 train_time:34658ms step_avg:150.03ms
step:242/1750 train_loss:4.1711 train_time:34807ms step_avg:150.03ms
step:243/1750 train_loss:4.0377 train_time:34956ms step_avg:150.03ms
step:244/1750 train_loss:4.0966 train_time:35106ms step_avg:150.02ms
step:245/1750 train_loss:4.1625 train_time:35255ms step_avg:150.02ms
step:246/1750 train_loss:4.0826 train_time:35404ms step_avg:150.02ms
step:247/1750 train_loss:4.0306 train_time:35553ms step_avg:150.01ms
step:248/1750 train_loss:4.1463 train_time:35703ms step_avg:150.01ms
step:249/1750 train_loss:3.9394 train_time:35853ms step_avg:150.01ms
step:250/1750 train_loss:3.9975 train_time:36002ms step_avg:150.01ms
step:250/1750 val_loss:4.0346 train_time:36040ms step_avg:150.17ms
step:251/1750 train_loss:4.1051 train_time:36152ms step_avg:150.01ms
step:252/1750 train_loss:4.1939 train_time:36303ms step_avg:150.01ms
step:253/1750 train_loss:3.9638 train_time:36455ms step_avg:150.02ms
step:254/1750 train_loss:3.9133 train_time:36603ms step_avg:150.01ms
step:255/1750 train_loss:4.0981 train_time:36753ms step_avg:150.01ms
step:256/1750 train_loss:4.0202 train_time:36901ms step_avg:150.01ms
step:257/1750 train_loss:4.0193 train_time:37053ms step_avg:150.01ms
step:258/1750 train_loss:4.0179 train_time:37202ms step_avg:150.01ms
step:259/1750 train_loss:4.0585 train_time:37352ms step_avg:150.01ms
step:260/1750 train_loss:4.0880 train_time:37502ms step_avg:150.01ms
step:261/1750 train_loss:4.0460 train_time:37655ms step_avg:150.02ms
step:262/1750 train_loss:4.0195 train_time:37809ms step_avg:150.03ms
step:263/1750 train_loss:3.9180 train_time:37962ms step_avg:150.05ms
step:264/1750 train_loss:4.0139 train_time:38115ms step_avg:150.06ms
step:265/1750 train_loss:3.8921 train_time:38268ms step_avg:150.07ms
step:266/1750 train_loss:3.9539 train_time:38419ms step_avg:150.07ms
step:267/1750 train_loss:3.9522 train_time:38573ms step_avg:150.09ms
step:268/1750 train_loss:3.9814 train_time:38725ms step_avg:150.10ms
step:269/1750 train_loss:3.8802 train_time:38878ms step_avg:150.11ms
step:270/1750 train_loss:4.1190 train_time:39031ms step_avg:150.12ms
step:271/1750 train_loss:4.0007 train_time:39183ms step_avg:150.13ms
step:272/1750 train_loss:3.9590 train_time:39336ms step_avg:150.14ms
step:273/1750 train_loss:3.9899 train_time:39488ms step_avg:150.14ms
step:274/1750 train_loss:4.0589 train_time:39641ms step_avg:150.16ms
step:275/1750 train_loss:4.0832 train_time:39794ms step_avg:150.17ms
step:276/1750 train_loss:4.2500 train_time:39947ms step_avg:150.18ms
step:277/1750 train_loss:4.0591 train_time:40100ms step_avg:150.19ms
step:278/1750 train_loss:4.1131 train_time:40253ms step_avg:150.20ms
step:279/1750 train_loss:4.0144 train_time:40406ms step_avg:150.21ms
step:280/1750 train_loss:4.2016 train_time:40560ms step_avg:150.22ms
step:281/1750 train_loss:3.9922 train_time:40712ms step_avg:150.23ms
step:282/1750 train_loss:3.9644 train_time:40865ms step_avg:150.24ms
step:283/1750 train_loss:3.9320 train_time:41017ms step_avg:150.25ms
step:284/1750 train_loss:4.0718 train_time:41170ms step_avg:150.26ms
step:285/1750 train_loss:4.0895 train_time:41321ms step_avg:150.26ms
step:286/1750 train_loss:4.1085 train_time:41475ms step_avg:150.27ms
step:287/1750 train_loss:3.9361 train_time:41628ms step_avg:150.28ms
step:288/1750 train_loss:4.0355 train_time:41781ms step_avg:150.29ms
step:289/1750 train_loss:3.9032 train_time:41935ms step_avg:150.30ms
step:290/1750 train_loss:3.8801 train_time:42087ms step_avg:150.31ms
step:291/1750 train_loss:3.9385 train_time:42241ms step_avg:150.32ms
step:292/1750 train_loss:3.8894 train_time:42393ms step_avg:150.33ms
step:293/1750 train_loss:3.9270 train_time:42545ms step_avg:150.33ms
step:294/1750 train_loss:3.9656 train_time:42698ms step_avg:150.34ms
step:295/1750 train_loss:3.8600 train_time:42850ms step_avg:150.35ms
step:296/1750 train_loss:3.8868 train_time:43004ms step_avg:150.36ms
step:297/1750 train_loss:3.8973 train_time:43158ms step_avg:150.38ms
step:298/1750 train_loss:3.9980 train_time:43311ms step_avg:150.39ms
step:299/1750 train_loss:3.8486 train_time:43464ms step_avg:150.39ms
step:300/1750 train_loss:3.9956 train_time:43617ms step_avg:150.40ms
step:301/1750 train_loss:3.9932 train_time:43769ms step_avg:150.41ms
step:302/1750 train_loss:3.9606 train_time:43921ms step_avg:150.41ms
step:303/1750 train_loss:4.0099 train_time:44074ms step_avg:150.42ms
step:304/1750 train_loss:3.9926 train_time:44225ms step_avg:150.43ms
step:305/1750 train_loss:4.4777 train_time:44378ms step_avg:150.43ms
step:306/1750 train_loss:3.9605 train_time:44530ms step_avg:150.44ms
step:307/1750 train_loss:3.8574 train_time:44683ms step_avg:150.45ms
step:308/1750 train_loss:4.0120 train_time:44836ms step_avg:150.46ms
step:309/1750 train_loss:3.8951 train_time:44989ms step_avg:150.46ms
step:310/1750 train_loss:4.1108 train_time:45141ms step_avg:150.47ms
step:311/1750 train_loss:3.9564 train_time:45293ms step_avg:150.48ms
step:312/1750 train_loss:3.8918 train_time:45445ms step_avg:150.48ms
step:313/1750 train_loss:3.9643 train_time:45599ms step_avg:150.49ms
step:314/1750 train_loss:4.0936 train_time:45749ms step_avg:150.49ms
step:315/1750 train_loss:3.9681 train_time:45903ms step_avg:150.50ms
step:316/1750 train_loss:3.8199 train_time:46057ms step_avg:150.51ms
step:317/1750 train_loss:3.9009 train_time:46210ms step_avg:150.52ms
step:318/1750 train_loss:3.9515 train_time:46363ms step_avg:150.53ms
step:319/1750 train_loss:3.9144 train_time:46516ms step_avg:150.54ms
step:320/1750 train_loss:4.0385 train_time:46668ms step_avg:150.54ms
step:321/1750 train_loss:3.9792 train_time:46820ms step_avg:150.55ms
step:322/1750 train_loss:3.9566 train_time:46974ms step_avg:150.56ms
step:323/1750 train_loss:4.0314 train_time:47125ms step_avg:150.56ms
step:324/1750 train_loss:3.9674 train_time:47279ms step_avg:150.57ms
step:325/1750 train_loss:4.0413 train_time:47432ms step_avg:150.58ms
step:326/1750 train_loss:3.9121 train_time:47583ms step_avg:150.58ms
step:327/1750 train_loss:4.4148 train_time:47736ms step_avg:150.59ms
step:328/1750 train_loss:4.0935 train_time:47887ms step_avg:150.59ms
step:329/1750 train_loss:3.8215 train_time:48040ms step_avg:150.59ms
step:330/1750 train_loss:3.7692 train_time:48192ms step_avg:150.60ms
step:331/1750 train_loss:4.0042 train_time:48344ms step_avg:150.60ms
step:332/1750 train_loss:3.9314 train_time:48496ms step_avg:150.61ms
step:333/1750 train_loss:3.9088 train_time:48647ms step_avg:150.61ms
step:334/1750 train_loss:3.8635 train_time:48799ms step_avg:150.62ms
step:335/1750 train_loss:4.0311 train_time:48952ms step_avg:150.62ms
step:336/1750 train_loss:3.9771 train_time:49104ms step_avg:150.63ms
step:337/1750 train_loss:4.4377 train_time:49257ms step_avg:150.63ms
step:338/1750 train_loss:3.9588 train_time:49409ms step_avg:150.64ms
step:339/1750 train_loss:3.8811 train_time:49561ms step_avg:150.64ms
step:340/1750 train_loss:3.9563 train_time:49713ms step_avg:150.65ms
step:341/1750 train_loss:3.8855 train_time:49865ms step_avg:150.65ms
step:342/1750 train_loss:3.8353 train_time:50016ms step_avg:150.65ms
step:343/1750 train_loss:3.8606 train_time:50170ms step_avg:150.66ms
step:344/1750 train_loss:4.0127 train_time:50321ms step_avg:150.66ms
step:345/1750 train_loss:3.8382 train_time:50475ms step_avg:150.67ms
step:346/1750 train_loss:3.7932 train_time:50626ms step_avg:150.67ms
step:347/1750 train_loss:3.8214 train_time:50778ms step_avg:150.68ms
step:348/1750 train_loss:3.8751 train_time:50930ms step_avg:150.68ms
step:349/1750 train_loss:3.8520 train_time:51082ms step_avg:150.69ms
step:350/1750 train_loss:3.5925 train_time:51235ms step_avg:150.69ms
step:351/1750 train_loss:3.8523 train_time:51386ms step_avg:150.69ms
step:352/1750 train_loss:4.2137 train_time:51538ms step_avg:150.70ms
step:353/1750 train_loss:3.6772 train_time:51689ms step_avg:150.70ms
step:354/1750 train_loss:3.9490 train_time:51841ms step_avg:150.70ms
step:355/1750 train_loss:3.8082 train_time:51993ms step_avg:150.70ms
step:356/1750 train_loss:3.9056 train_time:52144ms step_avg:150.71ms