File size: 55,930 Bytes
131da64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
import os
import sys
from contextlib import ExitStack
from pathlib import Path

from constants import CONFIG_PATH, LIB_DIR
sys.path.append(str(LIB_DIR / "hydra_submitit_launcher"))

import builtins
import random
import re
import signal
import traceback
from copy import deepcopy
from datetime import datetime

import hydra
import numpy as np
import omegaconf
from hydra.core.hydra_config import HydraConfig
from omegaconf import DictConfig, OmegaConf, open_dict, read_write
from safetensors.torch import load_file, save_file

import dataloader
from model import Diffusion
import utils
import wandb
from decoupled_utils import (check_gpu_memory_usage, get_hostname,
                             get_local_rank, get_rank, get_slurm_filename_info,
                             get_slurm_log_prefix, get_tpu_devices,
                             get_world_size, gprint, is_local_main_process,
                             is_main_process, is_torch_cuda_available,
                             is_torch_xla_available, print_params,
                             process_file_prefix, profile_memory, rank_zero_fn,
                             rprint, set_global_breakpoint, set_global_exists,
                             set_timing_builtins, try_except)
from utils import (ErrorHandler, _print_config, convert_state_dict_keys, set_omega_conf_resolvers, set_torch_defaults)

# Only needed when debugging hydra
# os.environ["HYDRA_FULL_ERROR"] = "1"

set_global_breakpoint()  # Overrides breakpoint() to use ipdb.set_trace() instead and handle distributed training
set_global_exists()
set_omega_conf_resolvers()

if is_torch_xla_available():
    from jax_smi import initialise_tracking

def _load_from_checkpoint(config, tokenizer):
    OmegaConf.resolve(config)
    if "hf" in config.backbone:
        return Diffusion(config=config, tokenizer=tokenizer).to("cuda")

    return Diffusion.load_from_checkpoint(config.eval.checkpoint_path, tokenizer=tokenizer, config=config)

@rank_zero_fn
def _print_batch(train_ds, valid_ds, tokenizer, k=64):
    for dl_type, dl in [("train", train_ds), ("valid", valid_ds)]:
        rprint(f"Printing {dl_type} dataloader batch.")
        batch = next(iter(dl))
        rprint("Batch input_ids.shape", batch["input_ids"].shape)
        first = batch["input_ids"][0, :k]
        last = batch["input_ids"][0, -k:]
        rprint(f"First {k} tokens:", tokenizer.decode(first))
        rprint("ids:", first)
        rprint(f"Last {k} tokens:", tokenizer.decode(last))
        rprint("ids:", last)


def generate_samples(config, tokenizer):
    rprint("Generating samples.")
    model = _load_from_checkpoint(config=config, tokenizer=tokenizer)
    model.gen_ppl_metric.reset()
    if config.eval.disable_ema:
        rprint("Disabling EMA.")
        model.ema = None
    stride_length = config.sampling.stride_length
    num_strides = config.sampling.num_strides
    for _ in range(config.sampling.num_sample_batches):
        if config.sampling.semi_ar:
            _, intermediate_samples, _ = model.restore_model_and_semi_ar_sample(
                stride_length=stride_length, num_strides=num_strides, dt=1 / config.sampling.steps
            )
            text_samples = intermediate_samples[-1]
            # Note: Samples generated using semi-ar method
            # need to to be processed before computing generative perplexity
            # since these samples contain numerous <|endoftext|> tokens
            # and diffusion.compute_generative_perplexity() discards
            # any text after the first EOS token.
        else:
            samples = model.restore_model_and_sample(num_steps=config.sampling.steps)
            text_samples = model.tokenizer.batch_decode(samples)
            model.compute_generative_perplexity(text_samples)
    
    rprint("Text samples:", text_samples)
    if not config.sampling.semi_ar:
        rprint("Generative perplexity:", model.gen_ppl_metric.compute())
    return text_samples


def instantiate_wandb(config, accelerator):
    if is_torch_xla_available():
        gprint("Initializing wandb for XLA")
    if config.mode == 'eval':
        config.wandb.project = f"{config.wandb.project}-eval"
    elif config.mode == 'zero-shot-eval':
        config.wandb.project = f"{config.wandb.project}-zero-shot-eval"

    if config.wandb.group is not None:
        config.wandb.group = str(config.wandb.group)

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    wandb_kwargs = dict(config.wandb)

    if getattr(config, "sweep_id", None) is not None:
        rprint(f"Setting Wandb group to {config.sweep_id}")
        wandb_kwargs["group"] = config.sweep_id
    del wandb_kwargs["project"]
    accelerator.init_trackers(
        config.wandb.project, config=OmegaConf.to_container(config, resolve=True, throw_on_missing=True), init_kwargs=dict(wandb=wandb_kwargs)
    )

    if getattr(config.trainer, "log_code", True) and is_main_process():
        if "matrix" in get_hostname():
            rprint(f"Not logging code to wandb on {get_hostname()}")
        else:
            rprint(f"Logging code to wandb from {Path(__file__).parent}")
            try:
                wandb.run.log_code(
                    root=str(Path(__file__).parent),
                    include_fn=lambda path: any(path.endswith(f) for f in (".py", ".yaml", ".yml", ".txt", ".md")),
                    exclude_fn=lambda path, root: any(x in os.path.relpath(path, root) for x in ("output", "multirun", "logs", "wandb")),
                )
            except Exception as e:
                rprint(f"Failed to log code to wandb: {e}")

    with open_dict(config):
        try:
            config.wandb_url = wandb.run.get_url()
            wandb.define_metric("global_samples")
            wandb.define_metric("effective_global_tokens")
            wandb.define_metric("effective_global_step")
            wandb.define_metric("train_metrics/samples")
            wandb.define_metric("trainer/loss", step_metric="global_samples")
        except Exception as e:
            rprint(f"Failed to get wandb url: {e}")

def instantiate_model(config, tokenizer):
    model = _load_from_checkpoint(config=config, tokenizer=tokenizer)
    if config.eval.disable_ema:
        rprint("Disabling EMA.")
        model.ema = None

    return model

def gconf(config, attr):
    return getattr(config, attr, None)


def has_ckpt(config, attr):
    return gconf(config, attr) is not None and utils.fsspec_exists(gconf(config, attr))


def set_env_vars(config):
    import torch
    hostname = __import__("socket").gethostname()
    rprint(f"Starting Training on {hostname}")
    import torch
    # os.environ["TORCHINDUCTOR_CACHE_DIR"] = str((Path.home() / ".cache" / "torchinductor").resolve())

    if not is_torch_xla_available():
        try:
            # Applies the equivalent of ulimit -l unlimited to this process [and children]
            # This caused a significant amount of pain to figure out
            import resource
            soft, hard = resource.getrlimit(resource.RLIMIT_MEMLOCK)
            resource.setrlimit(resource.RLIMIT_MEMLOCK, (hard, hard))
            if is_local_main_process():
                gprint(f"Successfully set RLIMIT_MEMLOCK to {hard}")
        except ValueError as e:
            rprint(f"Failed to set RLIMIT_MEMLOCK: {e}")
        except resource.error as e:
            rprint(f"Error setting RLIMIT_MEMLOCK: {e}")
    else:
        rprint(f"Not setting RLIMIT_MEMLOCK on XLA")

    if "matrix-3-28" in hostname or "matrix-3-26" in hostname:
        rprint(f"Disabling NCCL P2P")
        os.environ["NCCL_P2P_DISABLE"] = "1"

    if os.environ.get("TORCH_DISTRIBUTED_DEBUG", "") != "":
        assert False, f"TORCH_DISTRIBUTED_DEBUG is set to: {os.environ.get('TORCH_DISTRIBUTED_DEBUG')}. Please unset it as it starts a gloo backend."

    if config.model.use_spda_attn:
        os.environ["TORCH_CUDNN_SDPA_ENABLED"] = "1"
        os.environ["TORCH_CUDNN_MHA_ENABLED"] = "1"
        rprint("Setting SPDA Flags")

    if config.trainer.detect_anomaly:
        torch.autograd.set_detect_anomaly(True)

def update_config_before_resolution(config):
    import torch
    if hasattr(config, "training"):
        rprint(f"'training' has been refactored to 'trainer'. Please update the config.")
        
    with open_dict(config):
        config.output_dir = os.getcwd()
        config.logging_dir = os.getcwd()
        if config.model.use_kv_cache is False and config.mode == "eval" and config.loader.eval_batch_size > 1:
            config.loader.eval_batch_size = max(config.loader.eval_batch_size, 16)
        
        # todo revert?
        if getattr(config.eval, 'txt_img_ratio', None) is not None:
            # 2,1,0.5,0.25
            tot = config.model.length
            # if its 2:1, then distribute the tokens as 2/3, 1/3
            # if its 1:1, then distribute the tokens as 1/2, 1/2
            # if its 0.5:1, then distribute the tokens as 2/3, 1/3
            # if its 0.25:1, then distribute the tokens as 1/4, 3/4
            if config.eval.txt_img_ratio == 2:
                # do first 2/3 tokens as text, last 1/3 as image
                config.model.txt_length = int(tot * 2/3)
            elif config.eval.txt_img_ratio == 1:
                config.model.txt_length = int(tot / 2)
            elif config.eval.txt_img_ratio == 0.5:
                config.model.txt_length = int(tot * 2/3)
            elif config.eval.txt_img_ratio == 0.25:
                config.model.txt_length = int(tot / 4)
            config.model.img_length = tot - config.model.txt_length
            config.model.length = config.model.txt_length + config.model.img_length
            # config.eval.attention_caching_txt_to_img_ratio = config.model.txt_length // 20
            
        if getattr(config.eval, "varying_seq_len_ratio", False):
            assert getattr(config.eval, "sampling_step_ratio", None) is not None, "Must set both varying_seq_len_ratio and sampling_step_ratio"
            config.sampling.steps = int(config.model.length * config.eval.sampling_step_ratio)

        if getattr(config.eval, "ablation_config", False):
            if config.parameterization == "ar":
                rprint(f"WARNING!!!!! FORCING AR PARAMS")
                config.trainer.ar_shift = True
                config.model.full_attention = False

            config.data.keep_tensordict_on_disk = True
            if is_torch_cuda_available():
                if any(x.lower() in torch.cuda.get_device_name().lower() for x in ["v100", "1080", "2080", "quadro", "titan"]) or torch.cuda.get_device_capability()[0] <= 7:
                    rprint(f"Using 2080Ti/V100, setting precision to fp32")
                    config.trainer.precision = "no"
                    config.model.force_optimized_native_attn = False
                    config.trainer.compile = False
                    if any(x.lower() in torch.cuda.get_device_name().lower() for x in ["2080", "quadro"]):
                        config.loader.eval_batch_size = config.loader.eval_batch_size // 7
                        config.loader.batch_size = config.loader.batch_size // 7
                    elif any(x.lower() in torch.cuda.get_device_name().lower() for x in ["1080", "titan"]):
                        config.loader.eval_batch_size = config.loader.eval_batch_size // 6
                        config.loader.batch_size = config.loader.batch_size // 6
                    else:
                        config.loader.eval_batch_size = config.loader.eval_batch_size // 2
                        config.loader.batch_size = config.loader.batch_size // 2
                elif "a5000" in torch.cuda.get_device_name().lower() or "a4500" in torch.cuda.get_device_name().lower():
                    config.loader.eval_batch_size = config.loader.eval_batch_size // 2
                    config.loader.batch_size = config.loader.batch_size // 2
                else:
                    rprint(f"Found {torch.cuda.get_device_name()}")
                    config.loader.eval_batch_size = config.loader.eval_batch_size // 2
                    config.loader.batch_size = config.loader.batch_size // 2
            
            if getattr(config, "parametierzation", None) == "ar" and config.eval.cfg is not None:
                config.loader.eval_batch_size = config.loader.eval_batch_size // 2
                config.loader.batch_size = config.loader.batch_size // 2

            config.loader.eval_batch_size = max(config.loader.eval_batch_size, 1)
            config.loader.batch_size = max(config.loader.batch_size, 1)

            if getattr(config, "parametierzation", None) == "ar":
                config.trainer.compile = False
            
        if getattr(config.sampling, "sampling_step_frac", None) is not None:
            config.sampling.steps = int(config.model.length * config.sampling.sampling_step_frac)
            rprint(f"Setting sampling steps to {config.sampling.steps}")
        
        if os.environ.get("SUBMITIT_FOLDER") is not None or os.environ.get("CUSTOM_SBATCH_LAUNCHER", "0") == "1":
            rprint(f'Using submitit folder: {os.environ.get("SUBMITIT_FOLDER", "")}, setting slurm=True')
            config.slurm = True

        if (config.debug is False or os.environ.get("HYDRA_RUN_DIR_NAME", None) is not None) and torch.distributed.is_torchelastic_launched():
            config.trainer.restart_on_failure = True
            rprint(f"Setting restart_on_failure to True")

        if config.trainer.restart_on_failure and config.mode == 'train':
            if os.environ.get("HYDRA_RUN_DIR", None) is None and os.environ.get("HYDRA_RUN_DIR_NAME", None) is None:
                os.environ["HYDRA_RUN_DIR"] = config.output_dir
                rprint(f"Setting HYDRA_RUN_DIR to {os.environ['HYDRA_RUN_DIR']}")
            else:
                rprint(f"Not setting HYDRA_RUN_DIR, already set to {os.environ.get('HYDRA_RUN_DIR', 'N/A')}, and HYDRA_RUN_DIR_NAME is set to {os.environ.get('HYDRA_RUN_DIR_NAME', 'N/A')}")

            os.environ["RESTART_FAULT_TOLERANT"] = "1"
            rprint(f"Setting RESTART_FAULT_TOLERANT to 1")
        elif config.trainer.restart_on_failure:
            rprint(f"Restart_on_failure is True, but mode is not 'train', so not setting restart fault tolerant")

        relevant_vars = {}
        for key, value in os.environ.items():
            if "SLURM" in key or "NCCL" in key or "TORCH" in key:
                relevant_vars[key] = value

        config.env_vars = relevant_vars

        if config.trainer.profile_memory:
            config.trainer.max_steps = 2

        if config.debug and config.trainer.force_enable_checkpointing is False and (config.trainer.ckpt_steps is None or config.trainer.ckpt_steps > 0):
            config.trainer.ckpt_steps = 10000
            rprint(f"Only checkpointing every {config.trainer.ckpt_steps} steps in debug mode")

        if config.loader.global_batch_size is None:
            config.loader.global_batch_size = config.loader.batch_size * config.trainer.accumulate_grad_batches * (1 if is_torch_xla_available() else get_world_size())
            config.loader.eval_global_batch_size = config.loader.global_batch_size
            if config.trainer.scale_lr_by_batch_size:
                config.optim.lr = config.optim.lr * (config.loader.global_batch_size / 512)
            rprint(f"Setting global batch size to {config.loader.global_batch_size}, lr to {config.optim.lr}")

        if config.mode != 'train':
            config.checkpointing.resume_wandb = False
            config.wandb.resume = None

        if config.trainer.use_spmd_distributed_checkpointing is None:
            config.trainer.use_spmd_distributed_checkpointing = is_torch_xla_available() and config.trainer.xla_spmd

        if config.trainer.disable_all_eval_generation:
            config.eval.num_masking_viz_batches=0
            config.eval.num_uncond_sample_batches=0
            config.eval.num_sample_batches=0
            config.eval.num_random_masking=0
            config.eval.generate_samples=False
            config.trainer.log_flops=False
            config.eval.log_every_n_evals=-1
            config.eval.log_every_n_fid = -1
            config.model.image_model_fid_eval = False
            rprint("Disabling all eval generation!!!")

        if os.environ.get("XLA_IR_DEBUG", "0") == "1":
            config.trainer.tpu_profile = True

        if config.checkpointing_root_dir is not None:
            assert "checkpoints" in config.checkpointing.save_dir
            relative_path = Path(*Path(config.checkpointing.save_dir).relative_to(config.root_output_dir).parts[1:])
            full_checkpointing_dir = Path(config.checkpointing_root_dir) / relative_path
            if config.checkpointing_root_dir is not None:
                old_save_dir = Path(config.output_dir) / "checkpoints"
                full_checkpointing_dir.mkdir(parents=True, exist_ok=True)
                try:
                    if old_save_dir.exists():
                        rprint(f"WARNING: Cannot create symlink from {old_save_dir} to {full_checkpointing_dir} because {old_save_dir} exists.")
                        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
                        old_save_dir = Path(*old_save_dir.parts[:-1]) / f"checkpoints_{timestamp}"
                    
                    old_save_dir.symlink_to(full_checkpointing_dir, target_is_directory=True)
                    rprint(f"Created softlink from {old_save_dir} to {full_checkpointing_dir}")

                    # Create a symlink from the parent of full_checkpointing_dir named "original" back to config.output_dir
                    original_link = full_checkpointing_dir.parent / "original_output_dir"
                    if not original_link.exists():
                        original_link.symlink_to(Path(config.output_dir).resolve(), target_is_directory=True)
                        rprint(f"Created softlink from {original_link} to {config.output_dir}")
                    else:
                        rprint(f"WARNING: Symlink {original_link} already exists. Skipping creation.")

                except OSError as e:
                    rprint(f"Error creating softlinks: {e}")

        assert getattr(config.data, "allow_label", False) == getattr(config.trainer, "add_label", False) == (getattr(config.model, "add_labels", None) is not None) == getattr(config.eval, "class_conditional_fid", False), f"Mismatching values: data.allow_label={config.data.allow_label}, trainer.add_label={config.trainer.add_label}, model.add_labels={config.model.add_labels}, eval.class_conditional_fid={config.eval.class_conditional_fid}"

        if getattr(config.loader, "num_eval_workers", None) is not None and config.loader.num_workers == 0:
            rprint(f"Setting num_eval_workers to 0 because num_workers is 0")
            config.loader.num_eval_workers = 0

    if config.trainer.disable_all_checkpointing:
        gprint("-"*50)
        gprint(f"WARNING: DISABLING ALL CHECKPOINTING!!!!")
        gprint("-"*50)
        gprint(f"WARNING: DISABLING ALL CHECKPOINTING!!!!")
        gprint("-"*50)
        config.trainer.ckpt_steps = 100000000

    if config.sampling.steps != config.sampling.max_sampling_steps:
        rprint(f"WARNING!!!! steps {config.sampling.steps} != max_sampling_steps {config.sampling.max_sampling_steps}")
        config.sampling.max_sampling_steps = config.sampling.steps

def get_latest_ckpt(config, input_dir):
    if input_dir is None or not Path(input_dir).exists():
        rprint(f"Project dir {input_dir} does not exist")
        return None
    
    if config.trainer.xla_spmd and is_torch_xla_available():
        rprint(f"XLA SPMD detected, using XLA checkpointing")
        if any(Path(input_dir).iterdir()):
            rprint(f"Found existing files/folders in {input_dir}")
            return input_dir
        else:
            rprint(f"No folders found in {input_dir}")
            return None

    folders = [str(folder) for folder in Path(input_dir).iterdir() if folder.is_dir() and ((folder / "model.safetensors").exists() or (folder / "config.yaml").exists())]

    if len(folders) == 0:
        rprint(f"No folders found in {input_dir}")
        return None

    def _inner(folder):
        return list(map(int, re.findall(r"[\/]?([0-9]+)(?=[^\/]*$)", folder)))[0]

    folders.sort(key=_inner)
    rprint(f"Found folders: {folders}")
    input_dir = folders[-1]
    return input_dir

def is_sweep():
    try:
        subdir = HydraConfig.get().sweep.subdir
        rprint(f"Found sweep subdir: {subdir}")
        return True
    except omegaconf.errors.InterpolationToMissingValueError:
        return False
    
def get_sweep_run_name(config):
    try:
        subdir = HydraConfig.get().sweep.subdir
        sweep_str = f"{subdir}_"
        is_sweep = True
    except omegaconf.errors.InterpolationToMissingValueError:
        is_sweep = False
        sweep_str = f"{os.environ.get('SLURM_JOB_ID', '')}_"

    if getattr(config, "training", None) is not None and getattr(getattr(config, "training", None), "force_keys", None) is not None:
        rprint("Using legacy keys")
        forced_keys = set(config.training.force_keys)
    else:
        forced_keys = set(getattr(config.trainer, "forced_keys", []))

    if is_sweep:
        print(
            f"Getting sweep keys: {HydraConfig.get().job.sweep_keys}, Tasks: {HydraConfig.get().overrides.task}, {getattr(config.trainer, 'forced_keys', [])}"
        )
        valid_keys = set(HydraConfig.get().job.sweep_keys)
        for task in HydraConfig.get().overrides.task:
            if task.removeprefix("+").split("=")[0] in valid_keys or task.removeprefix("+").split("=")[0] in forced_keys:
                sweep_str += f"{task.removeprefix('+').split('=')[0].split('.')[-1]}={task.removeprefix('+').split('=')[1]}__"
                if task.removeprefix("+").split("=")[0] in forced_keys:
                    forced_keys.remove(task.removeprefix("+").split("=")[0])
                    print(f"Forced key: {task.removeprefix('+').split('=')[0]}={task.removeprefix('+').split('=')[1]}")

    for key in sorted(list(forced_keys)):
        sweep_str += f"{key.split('.')[-1]}={OmegaConf.select(config, key)}__"

    rprint(f"Sweep: {is_sweep=}, {sweep_str=}")
    return "" if sweep_str == "" else sweep_str[:-2]

def save_config_to_ckpt(config, output_dir, model):
    with try_except(write_error_to_file=True, clear_cuda_cache=True):
        with read_write(config):
            with open_dict(config):
                config.state.ckpt_step = model.global_step
                config.state.num_evals = model.num_evals

        OmegaConf.save(config=config, f=Path(output_dir) / "config.yaml")
        rprint(f"Saved global step {model.global_step}")

def determine_ckpt(config):
    has_recent_ckpt = False
    rprint(f"Looking at checkpoint path: {getattr(config.checkpointing, 'resume_ckpt_path', None)}")
    if (
        config.checkpointing.resume_from_ckpt
        and (latest_ckpt := get_latest_ckpt(config, getattr(config.checkpointing, "resume_ckpt_path", None))) is not None
        and (Path(latest_ckpt) / "config.yaml").exists()
    ):
        ckpt_path = latest_ckpt
        has_recent_ckpt = True
        if config.slurm:
            config.wandb.resume = "must"
        rprint(f"Resuming from checkpoint {ckpt_path}")
    elif config.checkpointing.resume_from_ckpt and getattr(config.checkpointing, "initial_resume_ckpt_path", None) is not None:
        ckpt_path = config.checkpointing.initial_resume_ckpt_path
        rprint(f"Resuming from initial checkpoint {ckpt_path}")
    else:
        ckpt_path = None

    if ckpt_path is not None and (config.checkpointing.resume_wandb or has_recent_ckpt):
        loaded = OmegaConf.load(Path(ckpt_path) / "config.yaml")
        if loaded.wandb.id is not None:
            config.wandb.id = str(loaded.wandb.id)
            rprint(f"Found wandb id: {config.wandb.id}")
        else:
            rprint(f"No wandb id found in checkpoint {ckpt_path}")

    if config.checkpointing.resume_wandb and config.wandb.id is not None:
        config.wandb.resume = "must"
        rprint(f"Resuming wandb, setting must, run id: {config.wandb.id}")
    elif config.slurm and config.wandb.id is None:
        if os.environ.get("SLURM_ARRAY_TASK_COUNT", "") != "" and int(os.environ.get("SLURM_ARRAY_TASK_COUNT", "")) > 1:
            config.wandb.id = str(os.environ.get("SLURM_ARRAY_JOB_ID")) + f"_{os.environ.get('SLURM_ARRAY_TASK_ID')}"
        else:
            config.wandb.id = str(os.environ.get("SLURM_JOB_ID"))
        rprint(f"Setting wandb id to {config.wandb.id}")

    if config.checkpointing.initial_resume_ckpt_path is not None and config.checkpointing.resume_wandb:
        assert config.wandb.id is not None

    if config.ckpt is not None:
        ckpt_path = config.ckpt
        rprint(f"Running eval with checkpoint {ckpt_path}")

    if config.wandb.id is not None:
        config.wandb.id = str(config.wandb.id)

    if config.wandb.id is None or getattr(config.trainer, "force_new_wandb_id", False):
        config.wandb.id = wandb.util.generate_id()
        config.wandb.resume = "allow"
        rprint(f"Set wandb id: {config.wandb.id}")

    rprint(f"Using wandb id: {config.wandb.id}")
    subdir = get_sweep_run_name(config)
    rprint(f"Wandb name: {config.wandb.name}, Wandb subdir: {subdir}")

    if config.wandb.name == 'default':
        config.wandb.name = None
    else:
        config.wandb.name = (
            (f"{config.wandb.name}_" if config.wandb.name else "")
            + (f"{subdir}_" if (subdir is not None and subdir != "") else "")
            + f"{datetime.now().strftime('%Y-%m-%d-%H-%M-%S')}"
        )

    if getattr(config.wandb, "group", None) is None and subdir is not None and config.debug and os.environ.get("SLURM_ARRAY_JOB_ID", "") != "":
        config.wandb.group = os.environ.get("SLURM_ARRAY_JOB_ID")
        rprint(f"Wandb group: {config.wandb.group}")

    return ckpt_path

def run(config, tokenizer):
    import torch
    from accelerate import (Accelerator, DataLoaderConfiguration,
                            DDPCommunicationHookType,
                            DistributedDataParallelKwargs,
                            FullyShardedDataParallelPlugin)
    from accelerate.state import AcceleratorState
    from accelerate.utils import GradientAccumulationPlugin, ProjectConfiguration

    set_torch_defaults(config.trainer.benchmark)

    set_env_vars(config)
    update_config_before_resolution(config)
    ckpt_path = determine_ckpt(config)
    OmegaConf.resolve(config)
    if is_torch_cuda_available():
        check_gpu_memory_usage()

    if is_torch_cuda_available():
        rprint(f"pt={torch.__version__}, cuda={torch.version.cuda}, nccl={torch.cuda.nccl.version()}")
        rprint(f"GPU={torch.cuda.get_device_name()}, device compute capabilities={torch.cuda.get_device_capability()}, pytorch compute capabilities={torch.cuda.get_arch_list()}")
    elif is_torch_xla_available():
        rprint(f"XLA Devices={get_tpu_devices()}")

    rprint(
        f"Initial GROUP_RANK: {os.environ.get('GROUP_RANK', 'N/A')}, RANK: {os.environ.get('RANK', 'N/A')}, LOCAL_RANK: {os.environ.get('LOCAL_RANK', 'N/A')}, WORLD_SIZE: {os.environ.get('WORLD_SIZE', 'N/A')}, MASTER_ADDR: {os.environ.get('MASTER_ADDR', 'N/A')}, MASTER_PORT: {os.environ.get('MASTER_PORT', 'N/A')}, TORCHELASTIC_RUN_ID: {os.environ.get('TORCHELASTIC_RUN_ID', 'N/A')}, TORCHELASTIC_RESTART_COUNT: {os.environ.get('TORCHELASTIC_RESTART_COUNT', 'N/A')}, TORCHELASTIC_MAX_RESTARTS: {os.environ.get('TORCHELASTIC_MAX_RESTARTS', 'N/A')}, LOCAL_WORLD_SIZE: {os.environ.get('LOCAL_WORLD_SIZE', 'N/A')}, Elastic: {torch.distributed.is_torchelastic_launched()}"
    )
    rprint(f"Computed Rank: {get_rank()}, Local Rank: {get_local_rank()}, World Size: {get_world_size()}")

    # This lets us have start_timing and end_timing functions and a global enable/disable
    # We always use torch.cuda.synchronize before/after as otherwise the timing is not very meaningful
    sync_timing = (config.trainer.nvtx_profile and getattr(config.trainer, "sync_nvtx_timing", True)) or getattr(config.trainer, "sync_timing", False)
    set_timing_builtins(enable=config.trainer.nvtx_profile, sync=sync_timing)

    num_nodes = config.trainer.num_nodes
    with open_dict(config):
        config.trainer = OmegaConf.merge(config.trainer, dict(mixed_precision=config.trainer.precision, log_with="wandb", log_gradients=None))
    if getattr(config.trainer, "process_dataloader_only", False):
        gprint("Processing dataloader only")
        train_ds, valid_ds = dataloader.get_dataloaders(config, tokenizer, device="cpu", skip_train=(config.mode == 'eval' and not config.eval.val_with_train_data))
        gprint(f"Exiting after processing dataloader")
        return

    accelerator_project_config = ProjectConfiguration(
        project_dir=config.output_dir,
        logging_dir=config.logging_dir,
        automatic_checkpoint_naming=config.checkpointing.use_automatic_naming,
        save_on_each_node=False,
    )

    accelerate_kwargs = dict()
    gradient_kwargs = dict()
    if config.trainer.fsdp and not (config.trainer.xla_spmd and is_torch_xla_available()):
        rprint("Using FSDP...")
        if config.backbone == "llama" or config.backbone == "gemma":
            os.environ["ACCELERATE_USE_FSDP"] = "true"
            os.environ["FSDP_AUTO_WRAP_POLICY"] = "TRANSFORMER_BASED_WRAP"
            os.environ["FSDP_BACKWARD_PREFETCH"] = "NO_PREFETCH" # Saved memory
            os.environ["FSDP_CPU_RAM_EFFICIENT_LOADING"] = "true"
            os.environ["FSDP_FORWARD_PREFETCH"] = "false"
            os.environ["FSDP_OFFLOAD_PARAMS"] = "false"
            os.environ["FSDP_SHARDING_STRATEGY"] = "FULL_SHARD"
            os.environ["FSDP_STATE_DICT_TYPE"] = "SHARDED_STATE_DICT"
            os.environ["FSDP_SYNC_MODULE_STATES"] = "true"
            os.environ["FSDP_USE_ORIG_PARAMS"] = "true"
            fsdp_plugin = FullyShardedDataParallelPlugin()
        else:
            os.environ["ACCELERATE_USE_FSDP"] = "true"
            os.environ["FSDP_AUTO_WRAP_POLICY"] = "TRANSFORMER_BASED_WRAP"  # or "SIZE_BASED_WRAP"
            if config.backbone == "elm":
                os.environ["FSDP_TRANSFORMER_CLS_TO_WRAP"] = "OpenELMDecoderLayer"
                os.environ["FSDP_BACKWARD_PREFETCH"] = "BACKWARD_PRE" 
                os.environ["FSDP_SHARDING_STRATEGY"] = "HYBRID_SHARD_ZERO2"
            else:
                # Fastest but requires more memory: https://pytorch.org/docs/stable/fsdp.html#torch.distributed.fsdp.BackwardPrefetch
                os.environ["FSDP_BACKWARD_PREFETCH"] = "BACKWARD_PRE" 
                # See: https://pytorch.org/docs/stable/fsdp.html#torch.distributed.fsdp.ShardingStrategy
                os.environ["FSDP_SHARDING_STRATEGY"] = "HYBRID_SHARD_ZERO2"
                os.environ["FSDP_TRANSFORMER_CLS_TO_WRAP"] = "DDiTBlock" 

            # SHARDED_STATE_DICT is a bit faster, but more complicated as later on we need to merge the shards.
            from torch.distributed.fsdp.fully_sharded_data_parallel import (
                FullOptimStateDictConfig, FullStateDictConfig)
            fsdp_plugin = FullyShardedDataParallelPlugin(
                state_dict_config=FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
                optim_state_dict_config=FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True), # SHARDED_STATE_DICT
            )

        if config.trainer.compile or config.trainer.use_orig_params is True:
            # https://github.com/huggingface/transformers/pull/24591/files
            fsdp_plugin.use_orig_params = True
            rprint("Using orig params for FSDP. This is required for torch.compile to work.")

        accelerate_kwargs["fsdp_plugin"] = fsdp_plugin
        gradient_kwargs["sync_each_batch"] = False

        if getattr(config.trainer, "fsdp_sync_each_batch", False): # Reduce memory usage: https://huggingface.co/docs/accelerate/en/concept_guides/gradient_synchronization#nosync-requires-additional-gpu-memory-when-using-fsdp
            rprint("Using sync each batch for Chameleon")
            gradient_kwargs["sync_each_batch"] = True

    elif config.trainer.xla_spmd is False: # For XLA FSDP, we init where we normally prepare()
        rprint("Using DDP...")
        ddp_kwargs = DistributedDataParallelKwargs(
            find_unused_parameters=config.trainer.find_unused_parameters,
            comm_hook=DDPCommunicationHookType.BF16,
            static_graph=config.trainer.accumulate_grad_batches == 1,
            gradient_as_bucket_view=True,
        )
        # bucket_cap_mb=32,

        # Not needed right now
        from datetime import timedelta

        from accelerate.utils import InitProcessGroupKwargs
        init_process_group_kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=1800))
        accelerate_kwargs["kwargs_handlers"] = [ddp_kwargs, init_process_group_kwargs]
    else:
        rprint(f"Did not choose DDP or FSDP.")

    if config.trainer.accumulate_grad_batches <= 0:
        gprint("WARNING!!!!!! Accumulate grad batches is <= 0, setting to 1")
        config.trainer.accumulate_grad_batches = 1

    gradient_accumulation_plugin = GradientAccumulationPlugin(
        num_steps=config.trainer.accumulate_grad_batches,
        adjust_scheduler=False, # We manually adjust our LR for accumulate_grad_batches
        sync_with_dataloader=False,
        **gradient_kwargs
    )

    if config.trainer.mixed_precision == "bf16" and (is_torch_cuda_available() and not torch.cuda.is_bf16_supported()):
        rprint(f"No BF16 GPU found, falling back to FP16")
        config.trainer.mixed_precision = "fp16"

    if config.trainer.mixed_precision == "fp32":
        config.trainer.mixed_precision = "no"
    else:
        if is_torch_xla_available():
            os.environ["ACCELERATE_DOWNCAST_BF16"] = "true"

    rprint(f"Mixed precision: {config.trainer.mixed_precision}")

    if config.seed is None or getattr(config.eval, 'set_random_gen_seed', False):
         # do not ask why, has to do something with seeds being reset by val_epoch_end so if you don't execute this code, your generations in val_epoch_end will be same across gpus
        accelerate_kwargs["rng_types"] = []
        rprint("No seed provided, disabling accelerate RNG synchronization")

    accelerator = Accelerator(
        mixed_precision=config.trainer.mixed_precision,
        log_with=config.trainer.log_with,
        project_config=accelerator_project_config,
        gradient_accumulation_plugin=gradient_accumulation_plugin,
        dataloader_config=DataLoaderConfiguration(split_batches=False, dispatch_batches=False, non_blocking=False),
        **accelerate_kwargs,
    )

    gprint(f"Distributed Type: {accelerator.distributed_type}, Accelerator state: {AcceleratorState()}")
    num_processes = AcceleratorState().num_processes
    if getattr(config.trainer, "global_num_warmup_steps", None) is not None:
        rprint(f"Global num_warmup_steps was: {config.lr_scheduler.num_warmup_steps}. Applying to num_warmup_steps")
        config.lr_scheduler.num_warmup_steps = config.trainer.global_num_warmup_steps

    if getattr(config.trainer, "global_num_training_steps", None) is not None:
        rprint(f"Global num_training_steps was: {config.lr_scheduler.num_training_steps}. Applying to num_training_steps")
        config.lr_scheduler.num_training_steps = config.trainer.global_num_training_steps

    if not config.trainer.disable_adjust_num_warmup_steps:
        rprint(f"Original num_warmup_steps was: {config.lr_scheduler.num_warmup_steps}")
        config.lr_scheduler.num_warmup_steps = config.lr_scheduler.num_warmup_steps * num_processes
        rprint(f"Setting num_warmup_steps to: {config.lr_scheduler.num_warmup_steps}")

        if hasattr(config.lr_scheduler, "num_training_steps"):
            rprint(f"Original num_training_steps was: {config.lr_scheduler.num_training_steps}")
            config.lr_scheduler.num_training_steps = config.lr_scheduler.num_training_steps * num_processes
            rprint(f"Setting num_training_steps to: {config.lr_scheduler.num_training_steps}")

    assert config.trainer.allow_dynamic_nodes or (os.environ.get("XLA_USE_SPMD", "0") == "1") or accelerator.num_processes == (
        config.trainer.devices * num_nodes
    ), f"Expected {config.trainer.devices * num_nodes} GPUs but got {accelerator.num_processes} processes."

    compute_dtyle = torch.float32
    if accelerator.mixed_precision == "fp16":
        compute_dtyle = torch.float16
    elif accelerator.mixed_precision == "bf16":
        compute_dtyle = torch.bfloat16

    if compute_dtyle != torch.bfloat16:
        rprint(f"WARNING!!!! Compute dtype is: {compute_dtyle}")
    else:
        rprint(f"Compute dtype is: {compute_dtyle}")

    if is_main_process():
        instantiate_wandb(config, accelerator)

    run_cmd = get_run_cmd(config)
    with open_dict(config):
        config.trainer.devices = accelerator.num_processes
        config.trainer.dtype = str(compute_dtyle)
        if hasattr(config, "state"):
            config.state.cmd = run_cmd
        else:
            config.state = OmegaConf.create(dict(cmd=run_cmd))

    OmegaConf.set_readonly(config, True)

    if getattr(config.trainer, "attach_oom_observer", False):
        from torchtnt.utils.oom import attach_oom_observer
        attach_oom_observer(output_dir=str(os.getcwd()), trace_max_entries=500000)
        rprint(f"Attached OOM observer to {os.getcwd()}")
    train_ds, valid_ds = dataloader.get_dataloaders(config, tokenizer, device=accelerator.device, skip_train=(config.mode == 'eval' and not config.eval.val_with_train_data))
    model = Diffusion(config=config, tokenizer=valid_ds.tokenizer, device=accelerator.device)

    if is_main_process():
        print_params(model.backbone)

    try:
        if getattr(config.model, "image_model", False) is False:
            _print_batch(train_ds, valid_ds, tokenizer)
    except:
        pass

    get_ema_path = lambda x: Path(x) / "ema.ckpt"
    SAMPLER_NAME = "weighted_dataset_sampler"

    def save_model_hook(models, weights, output_dir):
        nonlocal model, accelerator, train_ds

        if is_main_process():
            with try_except(write_error_to_file=True):
                if getattr(model, "ema", None) is not None:
                    torch.save(accelerator.unwrap_model(model).ema.state_dict(), get_ema_path(output_dir))
                    rprint(f"Saved EMA to {get_ema_path(output_dir)}")

            save_config_to_ckpt(config, output_dir, model)

            with try_except(write_error_to_file=True):
                if config.data.use_weighted_tensordict_sampler:
                    from accelerate.utils import save
                    output_sampler_file = output_dir.joinpath(f"{SAMPLER_NAME}_train.bin")
                    save(train_ds.sampler.state_dict(), output_sampler_file, save_on_each_node=False, safe_serialization=False)
                    rprint(f"Sampler state for dataloader saved in {output_sampler_file}")

    initial_global_step = None
    def load_model_hook(models, input_dir):
        nonlocal initial_global_step, model, train_ds
        config_path = os.path.join(input_dir, "config.yaml")
        ckpt_config = OmegaConf.load(config_path)
        initial_global_step = ckpt_config.state.ckpt_step
        model.global_step = initial_global_step
        try:
            if hasattr(config.state, "num_evals"):
                model.num_evals = config.state.num_evals
        except Exception as e:
            rprint(f"Error loading model: {e}")
        rprint(f"Loaded global step {initial_global_step}")

        state_dict = None
        if getattr(config.checkpointing, "load_from_old_attention_format", False):
            state_dict = load_file(os.path.join(input_dir, "model.safetensors"))
            state_dict = convert_state_dict_keys(state_dict)

        if getattr(model, "ema", None) is not None:
            if get_ema_path(input_dir).exists():
                rprint(f"Loading EMA from {get_ema_path(input_dir)}")
                model.ema.load_state_dict(torch.load(get_ema_path(input_dir), map_location='cpu'))
            else:
                rprint(f"No EMA found, initializing EMA with state_dict")
                if state_dict is None:
                    state_dict = load_file(os.path.join(input_dir, "model.safetensors"))

                # We likely don't need the unwrap, but just to be safe
                accelerator.unwrap_model(models[0]).load_state_dict(state_dict)
                from models.ema import EMAModel
                model.ema = EMAModel(accelerator.unwrap_model(models[0]).parameters(), decay=config.trainer.ema)

        if config.data.use_weighted_tensordict_sampler and not is_torch_xla_available(): # and not config.eval.test_eval_speed:
            input_sampler_file = Path(input_dir).joinpath(f"{SAMPLER_NAME}_train.bin")
            if train_ds is not None and input_sampler_file.exists():
                train_ds.sampler.load_state_dict(torch.load(input_sampler_file))
            rprint("All dataloader sampler states loaded successfully")

    accelerator.register_save_state_pre_hook(save_model_hook)
    accelerator.register_load_state_pre_hook(load_model_hook)
    model.init_dataloader(train_ds, valid_ds)
    model.set_accelerator(accelerator, ckpt_path)
    model.set_callbacks()

    if getattr(config.checkpointing, "load_from_text_model", None) is not None:
        rprint(f"Loading from text model")
        model.custom_load_checkpoint()

    if getattr(config.checkpointing, "load_from_lightning_ckpt", None) is not None:
        ckpt = torch.load(config.checkpointing.load_from_lightning_ckpt)
        initial_global_step = ckpt["global_step"]
        state_dict_ = {k.removeprefix("backbone."): v for k, v in ckpt["state_dict"].items() if "backbone" in k}
        state_dict_ = {k.replace(".attn_", ".attention.attn_"): v for k, v in state_dict_.items()}
        accelerator.unwrap_model(model.backbone).load_state_dict(state_dict_)

        if config.trainer.ema > 0:
            model.ema.load_state_dict(ckpt["ema"])

        rprint(f"Loaded lightning ckpt: {config.checkpointing.load_from_lightning_ckpt}")

    if initial_global_step is not None:
        # The load_hooks are before accelerate does it's loading and it overwrites model.global_step if we set it there
        model.global_step = initial_global_step
        rprint(f"Set global step to {initial_global_step}")

    contexts = []
    if config.trainer.nvtx_profile:
        contexts.append(torch.autograd.profiler.emit_nvtx(record_shapes=True))

    if config.trainer.profile_memory:
        contexts.append(profile_memory())

    using_torch_elastic = torch.distributed.is_torchelastic_launched()
    if using_torch_elastic:
        rprint(f"Torchelastic launched: {torch.distributed.is_torchelastic_launched()}")
        contexts.append(ErrorHandler())

    with ExitStack() as stack:
        for ctx in contexts:
            stack.enter_context(ctx)

        rprint(f"output_dir: {config.output_dir}")
        model.to(accelerator.device)
        if config.mode == 'train':
            model.train()
        elif config.mode == 'eval':
            if config.eval.standalone_fid:
                model.validate(None)
            else:
                model.validate(None)
        elif config.mode == 'zero-shot-eval':
            model.zero_shot_eval()
        else:
            raise ValueError(f"Invalid mode: {config.mode}")

    accelerator.end_training()


def get_run_cmd(config):
    orig_argv = deepcopy(sys.argv)

    prepend_argv = []
    if "HYDRA_RUN_DIR" in os.environ:
        prepend_argv.append(f"HYDRA_RUN_DIR='{os.environ['HYDRA_RUN_DIR']}'")
    else:
        prepend_argv.append(f"HYDRA_RUN_DIR='{str(Path(config.output_dir).resolve())}'")

    if orig_argv[1].startswith("experiments=["):
        orig_argv[1] = orig_argv[1].removeprefix("experiments=[").removesuffix("]")
        orig_argv[1] = f"experiments=\'[{orig_argv[1]}]\'"

    if os.environ.get("CUSTOM_SBATCH_LAUNCHER", "0") == "1":
        sbatch_script_path = 'scripts/slurm.sh'
        orig_argv.pop(0)
        orig_argv = ["sbatch", f"--nodes={os.environ.get('SLURM_NNODES', '1')}", f"--gpus-per-node={os.environ.get('SLURM_GPUS_PER_NODE', '1')}", f"--partition={os.environ.get('SLURM_JOB_PARTITION', 'all')}", sbatch_script_path] + orig_argv
    else:
        prepend_argv.append("accelerate launch")

    full_run_cmd = " ".join(prepend_argv + orig_argv)
    rprint(f"Full run cmd: {full_run_cmd}")
    return full_run_cmd

@hydra.main(version_base=None, config_path=CONFIG_PATH, config_name="config")
@try_except()
def main(config):
    if is_sweep():
        print(f"Checking if we need to requeue for job id {os.environ['SLURM_JOB_ID']}")
        from unidisc.utils.slurm_requeue import check_requeue
        check_requeue()
        print(f"Done checking if we need to requeue for job id {os.environ['SLURM_JOB_ID']}.")

    """Main entry point for trainer."""
    import torch  # Causes issue pickling if imported by default.
    if is_torch_xla_available():
        builtins.HAS_XLA_SPAWNED = True
        os.environ['PJRT_DEVICE'] = 'TPU'

        if config.trainer.precision == "bf16":
            os.environ['XLA_USE_BF16'] = '1'

        if config.devices == 1 and config.trainer.xla_spmd is False and config.trainer.fsdp is False:
            os.environ['TPU_PROCESS_BOUNDS'] = '1,1,1'
            os.environ['TPU_VISIBLE_CHIPS'] = '0'
            gprint(f"Setting TPU_PROCESS_BOUNDS: {os.environ['TPU_PROCESS_BOUNDS']}")
            gprint(f"Setting TPU_VISIBLE_CHIPS: {os.environ['TPU_VISIBLE_CHIPS']}")

        if config.trainer.tpu_eager:
            os.environ['XLA_USE_EAGER_DEBUG_MODE'] = '1'

        if config.trainer.tpu_compile_debug:
            os.environ['PT_XLA_DEBUG'] = '1'
            os.environ['PT_XLA_DEBUG_LEVEL'] = '2'
            os.environ['XLA_IR_DEBUG'] = '1'
            os.environ['XLA_HLO_DEBUG'] = '1'
            os.environ['TF_CPP_MIN_LOG_LEVEL'] = '0'
            os.environ['TF_CPP_VMODULE'] = 'xla_graph_executor=5,pjrt_computation_client=3'

        # We intentionally set these after to avoid import side effects
        spmd_mesh, axis_names, num_nodes = None, None, None
        if config.trainer.xla_spmd:
            import torch_xla.core.xla_model as xm
            import torch_xla.distributed.spmd as xs
            import torch_xla.runtime as xr
            from accelerate import PartialState
            from torch_xla._internal import tpu
            auto_spmd = getattr(config.trainer, "auto_spmd", False)

            xr.use_spmd(auto=auto_spmd) # Auto causes a crash
            force_global_devices = getattr(config.trainer, "force_global_devices", None)
            force_local_devices = getattr(config.trainer, "force_local_devices", None)
            assert (force_global_devices is None) == (force_local_devices is None), "Must set both or neither"

            if force_global_devices is not None:
                num_global_devices = force_global_devices
                num_local_devices = force_local_devices
                gprint(f"Using force global devices: num_global_devices={num_global_devices}, num_local_devices={num_local_devices}")
            else:
                num_global_devices = xr.global_runtime_device_count()
                num_local_devices = tpu.num_available_devices()
                assert num_global_devices == tpu.num_expected_global_devices()
                assert tpu.num_available_devices() == tpu.num_available_chips() == tpu.num_local_processes()

            num_nodes = num_global_devices // num_local_devices
            spmd_mesh_shape = getattr(config.trainer, "spmd_mesh", None)
            if spmd_mesh_shape is None:
                spmd_mesh_shape = (num_nodes, num_local_devices, 1)

            if getattr(config.trainer, "force_disable_replicas", False):
                spmd_mesh_shape = (1, num_global_devices, 1)
                rprint(f"Forcing disable replicas: {spmd_mesh_shape}")

            if auto_spmd is False:
                if getattr(config.trainer, "spmd_multislice", None) is not None:
                    from torch_xla.distributed.spmd import HybridMesh
                    ici_mesh_shape = spmd_mesh_shape
                    dcn_mesh_shape = (config.trainer.spmd_multislice, 1, 1)
                    spmd_mesh = HybridMesh(ici_mesh_shape=ici_mesh_shape, dcn_mesh_shape=dcn_mesh_shape, axis_names=('data','fsdp','tensor'))
                    rprint(f"Using multislice: {config.trainer.spmd_multislice}: {ici_mesh_shape} {dcn_mesh_shape}, {spmd_mesh.shape()}")
                else:
                    spmd_mesh = xs.Mesh(np.array(range(num_global_devices)), spmd_mesh_shape, ('dcn', 'fsdp', 'model'))
                xs.set_global_mesh(spmd_mesh)

            config.devices = 1
            config.nodes = 1

            with read_write(config):
                with open_dict(config):
                    config.state = OmegaConf.create(dict(spmd_mesh=spmd_mesh_shape))
                    config.state.axis_names = axis_names
                    config.state.num_nodes = num_nodes
                    config.state.num_global_devices = num_global_devices
                    config.state.num_local_devices = num_local_devices
                    config.state.worker_ips = tpu.get_worker_ips()
                    if os.environ.get("TPU_NAME") is not None:
                        config.state.tpu_name = os.environ.get("TPU_NAME")

        if config.trainer.tpu_eager:
            import torch_xla
            torch_xla.experimental.eager_mode(True)

        if config.trainer.tpu_profile:
            if config.trainer.tpu_profile_markers:
                os.environ['XLA_IR_DEBUG'] = '1'
                os.environ['XLA_HLO_DEBUG'] = '1'
            import torch_xla.debug.profiler as xp
            server = xp.start_server(9012)

        if config.trainer.tpu_cache:
            import torch_xla.runtime as xr
            readonly = not is_main_process()
            rprint(f"Initializing TPU cache with readonly={readonly}")
            xr.initialize_cache(str((Path.home() / '.cache' / 'unidisc' / f"tpu_{get_rank()}_{get_hostname().replace('-', '_')}").resolve()), readonly=readonly)

        if config.trainer.enable_jax_smi:
            initialise_tracking()
            rprint("Initializing jax-smi")

    from unidisc.utils.logging_utils import set_logger
    set_logger(f"{__name__} {get_slurm_log_prefix()}", Path(f"{get_slurm_filename_info()}_{get_hostname().replace('-', '_')}.out"))

    if is_torch_xla_available():
        import torch_xla.runtime as xr
        gprint(
                f"Computed Rank: {get_rank()}, "
                f"Is Main Process: {is_main_process()}, "
                f"Is Local Main Process: {is_local_main_process()}, "
                f"XLA world size: {xr.world_size()}, "
                f"XLA Model Ordinal: {xm.get_ordinal()}, "
                f"XLA Global Ordinal: {xr.global_ordinal()}, "
                f"XLA Supported Devices: {xm.get_xla_supported_devices()}, "
                f"Accelerate Local Process Index: {PartialState().local_process_index}, "
                f"Task ID: {tpu.task_id()}, "
                f"Worker ID: {tpu.worker_id()} "
                f"global device count: {xr.global_runtime_device_count()}, "
                f"local process count: {xr.local_process_count()}, "
                f"local device count: {xr.local_device_count()}, "
                f"addressable device count: {xr.addressable_device_count()}, "
                f"num_expected_global_devices: {tpu.num_expected_global_devices()}, "
                f"num_available_devices: {tpu.num_available_devices()}, "
                f"num_available_chips: {tpu.num_available_chips()}, "
                f"num_local_processes: {tpu.num_local_processes()}, "
                f"process_bounds_size: {tpu.process_bounds_size()}, "
                f"get_worker_ips: {tpu.get_worker_ips()}, "
                f"Computed Num Nodes: {num_nodes}, "
                f"Specified Mesh: {spmd_mesh_shape}, "
                f"Specified Mesh Axes: {axis_names}"
            )

        gprint(f"LIBTPU_INIT_ARGS: {os.environ.get('LIBTPU_INIT_ARGS', 'None')}")
        gprint(f"XLA_FLAGS: {os.environ.get('XLA_FLAGS', 'None')}")
    
    if getattr(config.trainer, "disable_ddp_optimizer", False):
        torch._dynamo.config.optimize_ddp = False

    if config.seed is not None:
        if config.mode == 'eval':
            config.seed = config.seed + 1000 * int(get_rank())
        else:
            config.seed = config.seed + int(get_rank())
        np.random.seed(config.seed)
        random.seed(config.seed)
        torch.manual_seed(config.seed)
        if is_torch_cuda_available():
            # TODO: Is seed all desired? Does it set the same one on all GPUs even in multi-process?
            torch.cuda.manual_seed_all(config.seed)

        if is_torch_xla_available():
            import torch_xla.core.xla_model as xm
            xm.set_rng_state(config.seed)
        gprint(f"Set seed: {config.seed}")
    else:
        rprint("No seed provided")

    _print_config(config, resolve=True, save_cfg=True)

    with open(f"env_vars_{get_slurm_filename_info()}_{get_hostname().replace('-', '_')}.txt", "w") as f:
        for key, value in os.environ.items():
            f.write(f"{key}={value}\n")

    tokenizer = dataloader.get_tokenizer(config)

    if "tokens" in config.data.train and (config.loader.num_workers > 0 or getattr(config.data, "force_mp_spawn", False)):
        from torch import multiprocessing as mp
        try:
            rprint(f"Start already method set to: {mp.get_start_method()}")
        except:
            mp.set_start_method("spawn")
            rprint(f"Start method set to: {mp.get_start_method()}")

    rprint(f"Mode: {config.mode}")
    if config.mode == "sample_eval":
        generate_samples(config, tokenizer)
    else:
        try:
            run(config, tokenizer)
        except Exception as e:
            rprint(f"Traceback: {traceback.format_exc()}")
            rprint(f"Exception: {e}")
    
            timestamp = int(__import__("time").time_ns())
            error_filepath = f"exception_{timestamp}_{process_file_prefix()}.out"
            with open(error_filepath, "w") as file:
                file.write(traceback.format_exc())
            rprint(f"See error file {Path(error_filepath).resolve()} for traceback")
            
            if is_torch_xla_available():
                exit(1)

            if ("SLURM_JOB_ID" not in os.environ) and ("RESTART_FAULT_TOLERANT" not in os.environ) and not is_torch_xla_available():
                gprint(f"Entering debugger")
                breakpoint(traceback=e.__traceback__)
            else:
                rprint(f"Not breaking, SLURM_JOB_ID: {os.environ.get('SLURM_JOB_ID')}, RESTART_FAULT_TOLERANT: {os.environ.get('RESTART_FAULT_TOLERANT')}")

            if "RESTART_FAULT_TOLERANT" in os.environ:
                sigterm_handler = signal.getsignal(signal.SIGTERM)
                if callable(sigterm_handler):
                    rprint(f"Calling SIGTERM handler")
                    sigterm_handler(signal.SIGTERM, None)

                try:
                    if config.trainer.num_nodes > 1 and config.debug is False and is_main_process():
                        wandb.alert(title="Exception!", text=f"{e}, {traceback.format_exc()}")
                except:
                    pass
            raise e
        finally:
            pass

if __name__ == "__main__":
    main()