-
Notifications
You must be signed in to change notification settings - Fork 257
/
Copy pathFSDP_adavnced_tutorial.html
1374 lines (1147 loc) ยท 118 KB
/
FSDP_adavnced_tutorial.html
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
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="ko" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="ko" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta property="og:title" content="Advanced Model Training with Fully Sharded Data Parallel (FSDP)" />
<meta property="og:type" content="article" />
<meta property="og:url" content="https://tutorials.pytorch.kr/intermediate/FSDP_adavnced_tutorial.html" />
<meta property="og:site_name" content="PyTorch Tutorials KR" />
<meta property="og:description" content="Author: Hamid Shojanazeri, Less Wright, Rohan Varma, Yanli Zhao This tutorial introduces more advanced features of Fully Sharded Data Parallel (FSDP) as part of the PyTorch 1.12 release. To get familiar with FSDP, please refer to the FSDP getting started tutorial. In this tutorial, we fine-tune a..." />
<meta property="og:image" content="https://tutorials.pytorch.kr/_static/logos/logo-kr-sm-dark.png" />
<meta property="og:image:alt" content="PyTorch Tutorials KR" />
<meta name="description" content="Author: Hamid Shojanazeri, Less Wright, Rohan Varma, Yanli Zhao This tutorial introduces more advanced features of Fully Sharded Data Parallel (FSDP) as part of the PyTorch 1.12 release. To get familiar with FSDP, please refer to the FSDP getting started tutorial. In this tutorial, we fine-tune a..." />
<meta property="og:ignore_canonical" content="true" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Advanced Model Training with Fully Sharded Data Parallel (FSDP) — ํ์ดํ ์น ํ๊ตญ์ด ํํ ๋ฆฌ์ผ (PyTorch tutorials in Korean)</title>
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="../_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.10/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-binder.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-dataframe.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-rendered-html.css" type="text/css" />
<link rel="stylesheet" href="../_static/sphinx-design.5ea377869091fd0449014c60fc090103.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.10.0-beta/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/custom.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/custom2.css" type="text/css" />
<link rel="index" title="์์ธ" href="../genindex.html" />
<link rel="search" title="๊ฒ์" href="../search.html" />
<link rel="next" title="Large Scale Transformer model training with Tensor Parallel (TP)" href="TP_tutorial.html" />
<link rel="prev" title="Getting Started with Fully Sharded Data Parallel(FSDP)" href="FSDP_tutorial.html" />
<script src="../_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.kr/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.kr/get-started">์์ํ๊ธฐ</a>
</li>
<li class="active">
<a href="https://tutorials.pytorch.kr/">ํํ ๋ฆฌ์ผ</a>
</li>
<li>
<a href="https://pytorch.kr/hub">ํ๋ธ</a>
</li>
<li>
<a href="https://discuss.pytorch.kr/">์ปค๋ฎค๋ํฐ</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
2.3.1+cu121
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="../search.html" method="get">
<input type="text" name="q" placeholder="Search Tutorials" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption" role="heading"><span class="caption-text">ํ์ดํ ์น(PyTorch) ๋ ์ํผ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../recipes/recipes_index.html">๋ชจ๋ ๋ ์ํผ ๋ณด๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../prototype/prototype_index.html">๋ชจ๋ ํ๋กํ ํ์
๋ ์ํผ ๋ณด๊ธฐ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ์ดํ ์น(PyTorch) ์์ํ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/intro.html">ํ์ดํ ์น(PyTorch) ๊ธฐ๋ณธ ์ตํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/quickstart_tutorial.html">๋น ๋ฅธ ์์(Quickstart)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/tensorqs_tutorial.html">ํ
์(Tensor)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/data_tutorial.html">Dataset๊ณผ DataLoader</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/transforms_tutorial.html">๋ณํ(Transform)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/buildmodel_tutorial.html">์ ๊ฒฝ๋ง ๋ชจ๋ธ ๊ตฌ์ฑํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/autogradqs_tutorial.html"><code class="docutils literal notranslate"><span class="pre">torch.autograd</span></code>๋ฅผ ์ฌ์ฉํ ์๋ ๋ฏธ๋ถ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/optimization_tutorial.html">๋ชจ๋ธ ๋งค๊ฐ๋ณ์ ์ต์ ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/saveloadrun_tutorial.html">๋ชจ๋ธ ์ ์ฅํ๊ณ ๋ถ๋ฌ์ค๊ธฐ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Introduction to PyTorch on YouTube</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt.html">PyTorch ์๊ฐ - YouTube ์๋ฆฌ์ฆ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/introyt1_tutorial.html">PyTorch ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensors_deeper_tutorial.html">Pytorch Tensor ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/autogradyt_tutorial.html">The Fundamentals of Autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/modelsyt_tutorial.html">Building Models with PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensorboardyt_tutorial.html">PyTorch TensorBoard Support</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/trainingyt.html">Training with PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/captumyt.html">Model Understanding with Captum</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ์ดํ ์น(PyTorch) ๋ฐฐ์ฐ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/deep_learning_60min_blitz.html">PyTorch๋ก ๋ฅ๋ฌ๋ํ๊ธฐ: 60๋ถ๋ง์ ๋์ฅ๋ด๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/pytorch_with_examples.html">์์ ๋ก ๋ฐฐ์ฐ๋ ํ์ดํ ์น(PyTorch)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/nn_tutorial.html"><cite>torch.nn</cite> ์ด <em>์ค์ ๋ก</em> ๋ฌด์์ธ๊ฐ์?</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard_tutorial.html">TensorBoard๋ก ๋ชจ๋ธ, ๋ฐ์ดํฐ, ํ์ต ์๊ฐํํ๊ธฐ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">์ด๋ฏธ์ง/๋น๋์ค</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchvision_tutorial.html">TorchVision Object Detection Finetuning Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/transfer_learning_tutorial.html">์ปดํจํฐ ๋น์ (Vision)์ ์ํ ์ ์ดํ์ต(Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/fgsm_tutorial.html">์ ๋์ ์์ ์์ฑ(Adversarial Example Generation)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/dcgan_faces_tutorial.html">DCGAN ํํ ๋ฆฌ์ผ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">๋ฐฐํฌ๋ฅผ ์ํด ๋น์ ํธ๋์คํฌ๋จธ(Vision Transformer) ๋ชจ๋ธ ์ต์ ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="tiatoolbox_tutorial.html">Whole Slide Image Classification Using PyTorch and TIAToolbox</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">์ค๋์ค</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_io_tutorial.html">Audio I/O</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_resampling_tutorial.html">Audio Resampling</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_data_augmentation_tutorial.html">Audio Data Augmentation</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_feature_extractions_tutorial.html">Audio Feature Extractions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_feature_augmentation_tutorial.html">Audio Feature Augmentation</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_datasets_tutorial.html">Audio Datasets</a></li>
<li class="toctree-l1"><a class="reference internal" href="speech_recognition_pipeline_tutorial.html">Speech Recognition with Wav2Vec2</a></li>
<li class="toctree-l1"><a class="reference internal" href="text_to_speech_with_torchaudio.html">Text-to-speech with Tacotron2</a></li>
<li class="toctree-l1"><a class="reference internal" href="forced_alignment_with_torchaudio_tutorial.html">wav2vec2์ ์ด์ฉํ ๊ฐ์ ์ ๋ ฌ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ
์คํธ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/bettertransformer_tutorial.html">Fast Transformer Inference with Better Transformer</a></li>
<li class="toctree-l1"><a class="reference internal" href="char_rnn_classification_tutorial.html">๊ธฐ์ด๋ถํฐ ์์ํ๋ NLP: ๋ฌธ์-๋จ์ RNN์ผ๋ก ์ด๋ฆ ๋ถ๋ฅํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="char_rnn_generation_tutorial.html">๊ธฐ์ด๋ถํฐ ์์ํ๋ NLP: ๋ฌธ์-๋จ์ RNN์ผ๋ก ์ด๋ฆ ์์ฑํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="seq2seq_translation_tutorial.html">๊ธฐ์ด๋ถํฐ ์์ํ๋ NLP: Sequence to Sequence ๋คํธ์ํฌ์ Attention์ ์ด์ฉํ ๋ฒ์ญ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/text_sentiment_ngrams_tutorial.html">torchtext ๋ผ์ด๋ธ๋ฌ๋ฆฌ๋ก ํ
์คํธ ๋ถ๋ฅํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/translation_transformer.html"><code class="docutils literal notranslate"><span class="pre">nn.Transformer</span></code> ์ torchtext๋ก ์ธ์ด ๋ฒ์ญํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/torchtext_custom_dataset_tutorial.html">Preprocess custom text dataset using Torchtext</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">๋ฐฑ์๋</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/onnx/intro_onnx.html">Introduction to ONNX</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">๊ฐํํ์ต</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="reinforcement_q_learning.html">๊ฐํ ํ์ต (DQN) ํํ ๋ฆฌ์ผ</a></li>
<li class="toctree-l1"><a class="reference internal" href="reinforcement_ppo.html">Reinforcement Learning (PPO) with TorchRL Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="mario_rl_tutorial.html">๋ง๋ฆฌ์ค ๊ฒ์ RL ์์ด์ ํธ๋ก ํ์ตํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/pendulum.html">Pendulum: Writing your environment and transforms with TorchRL</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch ๋ชจ๋ธ์ ํ๋ก๋์
ํ๊ฒฝ์ ๋ฐฐํฌํ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/onnx/intro_onnx.html">Introduction to ONNX</a></li>
<li class="toctree-l1"><a class="reference internal" href="flask_rest_api_tutorial.html">Flask๋ฅผ ์ฌ์ฉํ์ฌ Python์์ PyTorch๋ฅผ REST API๋ก ๋ฐฐํฌํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/Intro_to_TorchScript_tutorial.html">TorchScript ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_export.html">C++์์ TorchScript ๋ชจ๋ธ ๋ก๋ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/super_resolution_with_onnxruntime.html">(์ ํ) PyTorch ๋ชจ๋ธ์ ONNX์ผ๋ก ๋ณํํ๊ณ ONNX ๋ฐํ์์์ ์คํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="realtime_rpi.html">Raspberry Pi 4 ์์ ์ค์๊ฐ ์ถ๋ก (Inference) (30fps!)</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch ํ๋กํ์ผ๋ง</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch ๋ชจ๋ ํ๋กํ์ผ๋งํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hta_intro_tutorial.html">Introduction to Holistic Trace Analysis</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hta_trace_diff_tutorial.html">Trace Diff using Holistic Trace Analysis</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Code Transforms with FX</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="fx_conv_bn_fuser.html">(๋ฒ ํ) FX์์ ํฉ์ฑ๊ณฑ/๋ฐฐ์น ์ ๊ทํ(Convolution/Batch Norm) ๊ฒฐํฉ๊ธฐ(Fuser) ๋ง๋ค๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx_profiling_tutorial.html">(beta) Building a Simple CPU Performance Profiler with FX</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ๋ก ํธ์๋ API</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="memory_format_tutorial.html">(๋ฒ ํ) PyTorch๋ฅผ ์ฌ์ฉํ Channels Last ๋ฉ๋ชจ๋ฆฌ ํ์</a></li>
<li class="toctree-l1"><a class="reference internal" href="forward_ad_usage.html">Forward-mode Automatic Differentiation (Beta)</a></li>
<li class="toctree-l1"><a class="reference internal" href="jacobians_hessians.html">Jacobians, Hessians, hvp, vhp, and more: composing function transforms</a></li>
<li class="toctree-l1"><a class="reference internal" href="ensembling.html">๋ชจ๋ธ ์์๋ธ</a></li>
<li class="toctree-l1"><a class="reference internal" href="per_sample_grads.html">Per-sample-gradients</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_frontend.html">PyTorch C++ ํ๋ก ํธ์๋ ์ฌ์ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch-script-parallelism.html">TorchScript์ ๋์ ๋ณ๋ ฌ ์ฒ๋ฆฌ(Dynamic Parallelism)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_autograd.html">C++ ํ๋ก ํธ์๋์ ์๋ ๋ฏธ๋ถ (autograd)</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch ํ์ฅํ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="custom_function_double_backward_tutorial.html">Double Backward with Custom Functions</a></li>
<li class="toctree-l1"><a class="reference internal" href="custom_function_conv_bn_tutorial.html">Fusing Convolution and Batch Norm using Custom Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_extension.html">Custom C++ and CUDA Extensions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch_script_custom_ops.html">Extending TorchScript with Custom C++ Operators</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch_script_custom_classes.html">์ปค์คํ
C++ ํด๋์ค๋ก TorchScript ํ์ฅํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/dispatcher.html">Registering a Dispatched Operator in C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/extend_dispatcher.html">Extending dispatcher for a new backend in C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/privateuseone.html">Facilitating New Backend Integration by PrivateUse1</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">๋ชจ๋ธ ์ต์ ํ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch ๋ชจ๋ ํ๋กํ์ผ๋งํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard_profiler_tutorial.html">ํ
์๋ณด๋๋ฅผ ์ด์ฉํ ํ์ดํ ์น ํ๋กํ์ผ๋ฌ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hyperparameter_tuning_tutorial.html">Ray Tune์ ์ฌ์ฉํ ํ์ดํผํ๋ผ๋ฏธํฐ ํ๋</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">๋ฐฐํฌ๋ฅผ ์ํด ๋น์ ํธ๋์คํฌ๋จธ(Vision Transformer) ๋ชจ๋ธ ์ต์ ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="parametrizations.html">Parametrizations Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="pruning_tutorial.html">๊ฐ์ง์น๊ธฐ ๊ธฐ๋ฒ(Pruning) ํํ ๋ฆฌ์ผ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/dynamic_quantization_tutorial.html">(๋ฒ ํ) LSTM ๊ธฐ๋ฐ ๋จ์ด ๋จ์ ์ธ์ด ๋ชจ๋ธ์ ๋์ ์์ํ</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamic_quantization_bert_tutorial.html">(๋ฒ ํ) BERT ๋ชจ๋ธ ๋์ ์์ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantized_transfer_learning_tutorial.html">(๋ฒ ํ) ์ปดํจํฐ ๋น์ ํํ ๋ฆฌ์ผ์ ์ํ ์์ํ๋ ์ ์ดํ์ต(Quantized Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/static_quantization_tutorial.html">(๋ฒ ํ) PyTorch์์ Eager Mode๋ฅผ ์ด์ฉํ ์ ์ ์์ํ</a></li>
<li class="toctree-l1"><a class="reference internal" href="torchserve_with_ipex.html">Grokking PyTorch Intel CPU performance from first principles</a></li>
<li class="toctree-l1"><a class="reference internal" href="torchserve_with_ipex_2.html">Grokking PyTorch Intel CPU performance from first principles (Part 2)</a></li>
<li class="toctree-l1"><a class="reference internal" href="nvfuser_intro_tutorial.html">Getting Started - Accelerate Your Scripts with nvFuser</a></li>
<li class="toctree-l1"><a class="reference internal" href="ax_multiobjective_nas_tutorial.html">Multi-Objective NAS with Ax</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_compile_tutorial.html">Introduction to <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code></a></li>
<li class="toctree-l1"><a class="reference internal" href="inductor_debug_cpu.html">Inductor CPU backend debugging and profiling</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html">(Beta) Scaled Dot Product Attention (SDPA)๋ก ๊ณ ์ฑ๋ฅ ํธ๋์คํฌ๋จธ(Transformers) ๊ตฌํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#torch-compile-sdpa"><code class="docutils literal notranslate"><span class="pre">torch.compile</span></code> ๊ณผ ํจ๊ป SDPA ์ฌ์ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#sdpa-atteition-bias">SDPA๋ฅผ <code class="docutils literal notranslate"><span class="pre">atteition.bias</span></code> ํ์ ํด๋์ค์ ์ฌ์ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#id8">๊ฒฐ๋ก </a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/knowledge_distillation_tutorial.html">Knowledge Distillation Tutorial</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">๋ณ๋ ฌ ๋ฐ ๋ถ์ฐ ํ์ต</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../distributed/home.html">Distributed and Parallel Training Tutorials</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/dist_overview.html">PyTorch Distributed Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/ddp_series_intro.html">Distributed Data Parallel in PyTorch - Video Tutorials</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_parallel_tutorial.html">๋จ์ผ ๋จธ์ ์ ์ฌ์ฉํ ๋ชจ๋ธ ๋ณ๋ ฌํ ๋ชจ๋ฒ ์ฌ๋ก</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_tutorial.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ์์ํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="dist_tuto.html">PyTorch๋ก ๋ถ์ฐ ์ดํ๋ฆฌ์ผ์ด์
๊ฐ๋ฐํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="FSDP_tutorial.html">Getting Started with Fully Sharded Data Parallel(FSDP)</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Advanced Model Training with Fully Sharded Data Parallel (FSDP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="TP_tutorial.html">Large Scale Transformer model training with Tensor Parallel (TP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="process_group_cpp_extension_tutorial.html">Cpp ํ์ฅ์ ์ฌ์ฉํ ํ๋ก์ธ์ค ๊ทธ๋ฃน ๋ฐฑ์๋ ์ฌ์ฉ์ ์ ์</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_tutorial.html">Getting Started with Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_param_server_tutorial.html">Implementing a Parameter Server Using Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="dist_pipeline_parallel_tutorial.html">Distributed Pipeline Parallelism Using RPC</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_async_execution.html">Implementing Batch RPC Processing Using Asynchronous Executions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/rpc_ddp_tutorial.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ(DDP)๊ณผ ๋ถ์ฐ RPC ํ๋ ์์ํฌ ๊ฒฐํฉ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/ddp_pipeline.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ์ ๋ณ๋ ฌ ์ฒ๋ฆฌ ํ์ดํ๋ผ์ธ์ ์ฌ์ฉํ ํธ๋์คํฌ๋จธ ๋ชจ๋ธ ํ์ต</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/generic_join.html">Distributed Training with Uneven Inputs Using the Join Context Manager</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Edge with ExecuTorch</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/tutorials/export-to-executorch-tutorial.html">Exporting to ExecuTorch Tutorial</a></li>
<li class="toctree-l1"><a class="reference external" href=" https://pytorch.org/executorch/stable/running-a-model-cpp-tutorial.html">Running an ExecuTorch Model in C++ Tutorial</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/tutorials/sdk-integration-tutorial.html">Using the ExecuTorch SDK to Profile a Model</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/demo-apps-ios.html">Building an ExecuTorch iOS Demo App</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/demo-apps-android.html">Building an ExecuTorch Android Demo App</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/examples-end-to-end-to-lower-model-to-delegate.html">Lowering a Model as a Delegate</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">์ถ์ฒ ์์คํ
</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchrec_tutorial.html">TorchRec ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/sharding.html">Exploring TorchRec sharding</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Multimodality</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/flava_finetuning_tutorial.html">TorchMultimodal ํํ ๋ฆฌ์ผ: FLAVA ๋ฏธ์ธ์กฐ์ </a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="../index.html">
Tutorials
</a> >
</li>
<li>Advanced Model Training with Fully Sharded Data Parallel (FSDP)</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/intermediate/FSDP_adavnced_tutorial.rst.txt" rel="nofollow"><img src="../_static/images/view-page-source-icon.svg"></a>
</li>
</ul>
</div>
</div>
<div class="pytorch-shortcuts-wrapper" id="pytorch-shortcuts-wrapper">
Shortcuts
</div>
</div>
<section data-toggle="wy-nav-shift" id="pytorch-content-wrap" class="pytorch-content-wrap">
<div class="pytorch-content-left">
<div class="pytorch-call-to-action-links">
<div id="tutorial-type">intermediate/FSDP_adavnced_tutorial</div>
<div id="google-colab-link">
<img class="call-to-action-img" src="../_static/images/pytorch-colab.svg"/>
<div class="call-to-action-desktop-view">Run in Google Colab</div>
<div class="call-to-action-mobile-view">Colab</div>
</div>
<div id="download-notebook-link">
<img class="call-to-action-notebook-img" src="../_static/images/pytorch-download.svg"/>
<div class="call-to-action-desktop-view">Download Notebook</div>
<div class="call-to-action-mobile-view">Notebook</div>
</div>
<div id="github-view-link">
<img class="call-to-action-img" src="../_static/images/pytorch-github.svg"/>
<div class="call-to-action-desktop-view">View on GitHub</div>
<div class="call-to-action-mobile-view">GitHub</div>
</div>
</div>
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<div class="section" id="advanced-model-training-with-fully-sharded-data-parallel-fsdp">
<h1>Advanced Model Training with Fully Sharded Data Parallel (FSDP)<a class="headerlink" href="#advanced-model-training-with-fully-sharded-data-parallel-fsdp" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h1>
<p><strong>Author</strong>: <a class="reference external" href="https://github.com/HamidShojanazeri">Hamid Shojanazeri</a>, <a class="reference external" href="https://github.com/lessw2020">Less
Wright</a>, <a class="reference external" href="https://github.com/rohan-varma/">Rohan Varma</a>, <a class="reference external" href="https://github.com/zhaojuanmao">Yanli Zhao</a></p>
<p>This tutorial introduces more advanced features of Fully Sharded Data Parallel
(FSDP) as part of the PyTorch 1.12 release. To get familiar with FSDP, please
refer to the <a class="reference external" href="https://tutorials.pytorch.kr/intermediate/FSDP_tutorial.html">FSDP getting started tutorial</a>.</p>
<p>In this tutorial, we fine-tune a HuggingFace (HF) T5 model with FSDP for text
summarization as a working example.</p>
<p>The example uses Wikihow and for simplicity, we will showcase the training on a
single node, P4dn instance with 8 A100 GPUs. We will soon have a blog post on
large scale FSDP training on a multi-node cluster, please stay tuned for that on
the PyTorch medium channel.</p>
<p>FSDP is a production ready package with focus on ease of use, performance, and
long-term support. One of the main benefits of FSDP is reducing the memory
footprint on each GPU. This enables training of larger models with lower total
memory vs DDP, and leverages the overlap of computation and communication to
train models efficiently.
This reduced memory pressure can be leveraged to either train larger models or
increase batch size, potentially helping overall training throughput. You can
read more about PyTorch FSDP <a class="reference external" href="https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api/">here</a>.</p>
<div class="section" id="fsdp-features-in-this-tutorial">
<h2>FSDP Features in This Tutorial<a class="headerlink" href="#fsdp-features-in-this-tutorial" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<ul class="simple">
<li><p>Transformer Auto Wrap Policy</p></li>
<li><p>Mixed Precision</p></li>
<li><p>Initializing FSDP Model on Device</p></li>
<li><p>Sharding Strategy</p></li>
<li><p>Backward Prefetch</p></li>
<li><p>Model Checkpoint Saving via Streaming to CPU</p></li>
</ul>
</div>
<div class="section" id="recap-on-how-fsdp-works">
<h2>Recap on How FSDP Works<a class="headerlink" href="#recap-on-how-fsdp-works" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>At a high level FDSP works as follow:</p>
<p><em>In constructor</em></p>
<ul class="simple">
<li><p>Shard model parameters and each rank only keeps its own shard</p></li>
</ul>
<p><em>In forward pass</em></p>
<ul class="simple">
<li><p>Run <cite>all_gather</cite> to collect all shards from all ranks to recover the full
parameter for this FSDP unit Run forward computation</p></li>
<li><p>Discard non-owned parameter shards it has just collected to free memory</p></li>
</ul>
<p><em>In backward pass</em></p>
<ul class="simple">
<li><p>Run <cite>all_gather</cite> to collect all shards from all ranks to recover the full
parameter in this FSDP unit Run backward computation</p></li>
<li><p>Discard non-owned parameters to free memory.</p></li>
<li><p>Run reduce_scatter to sync gradients</p></li>
</ul>
</div>
<div class="section" id="fine-tuning-hf-t5">
<h2>Fine-tuning HF T5<a class="headerlink" href="#fine-tuning-hf-t5" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>HF T5 pre-trained models are available in four different sizes, ranging from
small with 60 Million parameters to XXL with 11 Billion parameters. In this
tutorial, we demonstrate the fine-tuning of a T5 3B with FSDP for text
summarization using WikiHow dataset. The main focus of this tutorial is to
highlight different available features in FSDP that are helpful for training
large scale model above 3B parameters. Also, we cover specific features for
Transformer based models. The code for this tutorial is available in <a class="reference external" href="https://github.com/pytorch/examples/tree/main/distributed/FSDP/">Pytorch
examples</a>.</p>
<p><em>Setup</em></p>
<p>1.1 Install PyTorch Nightlies</p>
<p>We will install PyTorch nightlies, as some of the features such as activation
checkpointing is available in nightlies and will be added in next PyTorch
release after 1.12.</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span>pip3<span class="w"> </span>install<span class="w"> </span>--pre<span class="w"> </span>torch<span class="w"> </span>torchvision<span class="w"> </span>torchaudio<span class="w"> </span>-f<span class="w"> </span>https://download.pytorch.org/whl/nightly/cu113/torch_nightly.html
</pre></div>
</div>
<p>1.2 Dataset Setup</p>
<p>Please create a <cite>data</cite> folder, download the WikiHow dataset from <a class="reference external" href="https://ucsb.app.box.com/s/ap23l8gafpezf4tq3wapr6u8241zz358">wikihowAll.csv</a> and
<a class="reference external" href="https://ucsb.app.box.com/s/7yq601ijl1lzvlfu4rjdbbxforzd2oag">wikihowSep.cs</a>,
and place them in the <cite>data</cite> folder. We will use the wikihow dataset from
<a class="reference external" href="https://github.com/pytorch/examples/blob/main/distributed/FSDP/summarization_dataset.py">summarization_dataset</a>.</p>
<p>Next, we add the following code snippets to a Python script โT5_training.pyโ.</p>
<div class="admonition note">
<p class="admonition-title">์ฐธ๊ณ </p>
<p>The full source code for this tutorial is available in <a class="reference external" href="https://github.com/pytorch/examples/tree/main/distributed/FSDP/">PyTorch examples</a>.</p>
</div>
<p>1.3 Import necessary packages:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">os</span>
<span class="kn">import</span> <span class="nn">argparse</span>
<span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.nn</span> <span class="k">as</span> <span class="nn">nn</span>
<span class="kn">import</span> <span class="nn">torch.nn.functional</span> <span class="k">as</span> <span class="nn">F</span>
<span class="kn">import</span> <span class="nn">torch.optim</span> <span class="k">as</span> <span class="nn">optim</span>
<span class="kn">from</span> <span class="nn">transformers</span> <span class="kn">import</span> <span class="n">AutoTokenizer</span><span class="p">,</span> <span class="n">GPT2TokenizerFast</span>
<span class="kn">from</span> <span class="nn">transformers</span> <span class="kn">import</span> <span class="n">T5Tokenizer</span><span class="p">,</span> <span class="n">T5ForConditionalGeneration</span>
<span class="kn">import</span> <span class="nn">functools</span>
<span class="kn">from</span> <span class="nn">torch.optim.lr_scheduler</span> <span class="kn">import</span> <span class="n">StepLR</span>
<span class="kn">import</span> <span class="nn">torch.nn.functional</span> <span class="k">as</span> <span class="nn">F</span>
<span class="kn">import</span> <span class="nn">torch.distributed</span> <span class="k">as</span> <span class="nn">dist</span>
<span class="kn">import</span> <span class="nn">torch.multiprocessing</span> <span class="k">as</span> <span class="nn">mp</span>
<span class="kn">from</span> <span class="nn">torch.nn.parallel</span> <span class="kn">import</span> <span class="n">DistributedDataParallel</span> <span class="k">as</span> <span class="n">DDP</span>
<span class="kn">from</span> <span class="nn">torch.utils.data.distributed</span> <span class="kn">import</span> <span class="n">DistributedSampler</span>
<span class="kn">from</span> <span class="nn">transformers.models.t5.modeling_t5</span> <span class="kn">import</span> <span class="n">T5Block</span>
<span class="kn">from</span> <span class="nn">torch.distributed.algorithms._checkpoint.checkpoint_wrapper</span> <span class="kn">import</span> <span class="p">(</span>
<span class="n">checkpoint_wrapper</span><span class="p">,</span>
<span class="n">CheckpointImpl</span><span class="p">,</span>
<span class="n">apply_activation_checkpointing_wrapper</span><span class="p">)</span>
<span class="kn">from</span> <span class="nn">torch.distributed.fsdp</span> <span class="kn">import</span> <span class="p">(</span>
<span class="n">FullyShardedDataParallel</span> <span class="k">as</span> <span class="n">FSDP</span><span class="p">,</span>
<span class="n">MixedPrecision</span><span class="p">,</span>
<span class="n">BackwardPrefetch</span><span class="p">,</span>
<span class="n">ShardingStrategy</span><span class="p">,</span>
<span class="n">FullStateDictConfig</span><span class="p">,</span>
<span class="n">StateDictType</span><span class="p">,</span>
<span class="p">)</span>
<span class="kn">from</span> <span class="nn">torch.distributed.fsdp.wrap</span> <span class="kn">import</span> <span class="p">(</span>
<span class="n">transformer_auto_wrap_policy</span><span class="p">,</span>
<span class="n">enable_wrap</span><span class="p">,</span>
<span class="n">wrap</span><span class="p">,</span>
<span class="p">)</span>
<span class="kn">from</span> <span class="nn">functools</span> <span class="kn">import</span> <span class="n">partial</span>
<span class="kn">from</span> <span class="nn">torch.utils.data</span> <span class="kn">import</span> <span class="n">DataLoader</span>
<span class="kn">from</span> <span class="nn">pathlib</span> <span class="kn">import</span> <span class="n">Path</span>
<span class="kn">from</span> <span class="nn">summarization_dataset</span> <span class="kn">import</span> <span class="o">*</span>
<span class="kn">from</span> <span class="nn">transformers.models.t5.modeling_t5</span> <span class="kn">import</span> <span class="n">T5Block</span>
<span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">Type</span>
<span class="kn">import</span> <span class="nn">time</span>
<span class="kn">import</span> <span class="nn">tqdm</span>
<span class="kn">from</span> <span class="nn">datetime</span> <span class="kn">import</span> <span class="n">datetime</span>
</pre></div>
</div>
<p>1.4 Distributed training setup.
Here we use two helper functions to initialize the processes for distributed
training, and then to clean up after training completion. In this tutorial, we
are going to use torch elastic, using <a class="reference external" href="https://pytorch.org/docs/stable/elastic/run.html">torchrun</a> , which will set the
worker <cite>RANK</cite> and <cite>WORLD_SIZE</cite> automatically.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">setup</span><span class="p">():</span>
<span class="c1"># initialize the process group</span>
<span class="n">dist</span><span class="o">.</span><span class="n">init_process_group</span><span class="p">(</span><span class="s2">"nccl"</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">cleanup</span><span class="p">():</span>
<span class="n">dist</span><span class="o">.</span><span class="n">destroy_process_group</span><span class="p">()</span>
</pre></div>
</div>
<p>2.1 Set up the HuggingFace T5 model:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">setup_model</span><span class="p">(</span><span class="n">model_name</span><span class="p">):</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">T5ForConditionalGeneration</span><span class="o">.</span><span class="n">from_pretrained</span><span class="p">(</span><span class="n">model_name</span><span class="p">)</span>
<span class="n">tokenizer</span> <span class="o">=</span> <span class="n">T5Tokenizer</span><span class="o">.</span><span class="n">from_pretrained</span><span class="p">(</span><span class="n">model_name</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span><span class="p">,</span> <span class="n">tokenizer</span>
</pre></div>
</div>
<p>We also, add couple of helper functions here for date and formatting memory
metrics.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">get_date_of_run</span><span class="p">():</span>
<span class="w"> </span><span class="sd">"""create date and time for file save uniqueness</span>
<span class="sd"> example: 2022-05-07-08:31:12_PM'</span>
<span class="sd"> """</span>
<span class="n">date_of_run</span> <span class="o">=</span> <span class="n">datetime</span><span class="o">.</span><span class="n">now</span><span class="p">()</span><span class="o">.</span><span class="n">strftime</span><span class="p">(</span><span class="s2">"%Y-%m-</span><span class="si">%d</span><span class="s2">-%I:%M:%S_%p"</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"--> current date and time of run = </span><span class="si">{</span><span class="n">date_of_run</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="k">return</span> <span class="n">date_of_run</span>
<span class="k">def</span> <span class="nf">format_metrics_to_gb</span><span class="p">(</span><span class="n">item</span><span class="p">):</span>
<span class="w"> </span><span class="sd">"""quick function to format numbers to gigabyte and round to 4 digit precision"""</span>
<span class="n">metric_num</span> <span class="o">=</span> <span class="n">item</span> <span class="o">/</span> <span class="n">g_gigabyte</span>
<span class="n">metric_num</span> <span class="o">=</span> <span class="nb">round</span><span class="p">(</span><span class="n">metric_num</span><span class="p">,</span> <span class="n">ndigits</span><span class="o">=</span><span class="mi">4</span><span class="p">)</span>
<span class="k">return</span> <span class="n">metric_num</span>
</pre></div>
</div>
<p>2.2 Define a train function:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">train</span><span class="p">(</span><span class="n">args</span><span class="p">,</span> <span class="n">model</span><span class="p">,</span> <span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="p">,</span> <span class="n">train_loader</span><span class="p">,</span> <span class="n">optimizer</span><span class="p">,</span> <span class="n">epoch</span><span class="p">,</span> <span class="n">sampler</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
<span class="n">model</span><span class="o">.</span><span class="n">train</span><span class="p">()</span>
<span class="n">local_rank</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'LOCAL_RANK'</span><span class="p">])</span>
<span class="n">fsdp_loss</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="mi">2</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">local_rank</span><span class="p">)</span>
<span class="k">if</span> <span class="n">sampler</span><span class="p">:</span>
<span class="n">sampler</span><span class="o">.</span><span class="n">set_epoch</span><span class="p">(</span><span class="n">epoch</span><span class="p">)</span>
<span class="k">if</span> <span class="n">rank</span><span class="o">==</span><span class="mi">0</span><span class="p">:</span>
<span class="n">inner_pbar</span> <span class="o">=</span> <span class="n">tqdm</span><span class="o">.</span><span class="n">tqdm</span><span class="p">(</span>
<span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="n">train_loader</span><span class="p">)),</span> <span class="n">colour</span><span class="o">=</span><span class="s2">"blue"</span><span class="p">,</span> <span class="n">desc</span><span class="o">=</span><span class="s2">"r0 Training Epoch"</span>
<span class="p">)</span>
<span class="k">for</span> <span class="n">batch</span> <span class="ow">in</span> <span class="n">train_loader</span><span class="p">:</span>
<span class="k">for</span> <span class="n">key</span> <span class="ow">in</span> <span class="n">batch</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
<span class="n">batch</span><span class="p">[</span><span class="n">key</span><span class="p">]</span> <span class="o">=</span> <span class="n">batch</span><span class="p">[</span><span class="n">key</span><span class="p">]</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">local_rank</span><span class="p">)</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">model</span><span class="p">(</span><span class="n">input_ids</span><span class="o">=</span><span class="n">batch</span><span class="p">[</span><span class="s2">"source_ids"</span><span class="p">],</span><span class="n">attention_mask</span><span class="o">=</span><span class="n">batch</span><span class="p">[</span><span class="s2">"source_mask"</span><span class="p">],</span><span class="n">labels</span><span class="o">=</span><span class="n">batch</span><span class="p">[</span><span class="s2">"target_ids"</span><span class="p">]</span> <span class="p">)</span>
<span class="n">loss</span> <span class="o">=</span> <span class="n">output</span><span class="p">[</span><span class="s2">"loss"</span><span class="p">]</span>
<span class="n">loss</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="n">fsdp_loss</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">+=</span> <span class="n">loss</span><span class="o">.</span><span class="n">item</span><span class="p">()</span>
<span class="n">fsdp_loss</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="o">+=</span> <span class="nb">len</span><span class="p">(</span><span class="n">batch</span><span class="p">)</span>
<span class="k">if</span> <span class="n">rank</span><span class="o">==</span><span class="mi">0</span><span class="p">:</span>
<span class="n">inner_pbar</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
<span class="n">dist</span><span class="o">.</span><span class="n">all_reduce</span><span class="p">(</span><span class="n">fsdp_loss</span><span class="p">,</span> <span class="n">op</span><span class="o">=</span><span class="n">dist</span><span class="o">.</span><span class="n">ReduceOp</span><span class="o">.</span><span class="n">SUM</span><span class="p">)</span>
<span class="n">train_accuracy</span> <span class="o">=</span> <span class="n">fsdp_loss</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">/</span> <span class="n">fsdp_loss</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span>
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="n">inner_pbar</span><span class="o">.</span><span class="n">close</span><span class="p">()</span>
<span class="nb">print</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">"Train Epoch: </span><span class="se">\t</span><span class="si">{</span><span class="n">epoch</span><span class="si">}</span><span class="s2">, Loss: </span><span class="se">\t</span><span class="si">{</span><span class="n">train_accuracy</span><span class="si">:</span><span class="s2">.4f</span><span class="si">}</span><span class="s2">"</span>
<span class="p">)</span>
<span class="k">return</span> <span class="n">train_accuracy</span>
</pre></div>
</div>
<p>2.3 Define a validation function:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">validation</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="p">,</span> <span class="n">val_loader</span><span class="p">):</span>
<span class="n">model</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="n">correct</span> <span class="o">=</span> <span class="mi">0</span>
<span class="n">local_rank</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'LOCAL_RANK'</span><span class="p">])</span>
<span class="n">fsdp_loss</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="mi">3</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">local_rank</span><span class="p">)</span>
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="n">inner_pbar</span> <span class="o">=</span> <span class="n">tqdm</span><span class="o">.</span><span class="n">tqdm</span><span class="p">(</span>
<span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="n">val_loader</span><span class="p">)),</span> <span class="n">colour</span><span class="o">=</span><span class="s2">"green"</span><span class="p">,</span> <span class="n">desc</span><span class="o">=</span><span class="s2">"Validation Epoch"</span>
<span class="p">)</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">():</span>
<span class="k">for</span> <span class="n">batch</span> <span class="ow">in</span> <span class="n">val_loader</span><span class="p">:</span>
<span class="k">for</span> <span class="n">key</span> <span class="ow">in</span> <span class="n">batch</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
<span class="n">batch</span><span class="p">[</span><span class="n">key</span><span class="p">]</span> <span class="o">=</span> <span class="n">batch</span><span class="p">[</span><span class="n">key</span><span class="p">]</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">local_rank</span><span class="p">)</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">model</span><span class="p">(</span><span class="n">input_ids</span><span class="o">=</span><span class="n">batch</span><span class="p">[</span><span class="s2">"source_ids"</span><span class="p">],</span><span class="n">attention_mask</span><span class="o">=</span><span class="n">batch</span><span class="p">[</span><span class="s2">"source_mask"</span><span class="p">],</span><span class="n">labels</span><span class="o">=</span><span class="n">batch</span><span class="p">[</span><span class="s2">"target_ids"</span><span class="p">])</span>
<span class="n">fsdp_loss</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">+=</span> <span class="n">output</span><span class="p">[</span><span class="s2">"loss"</span><span class="p">]</span><span class="o">.</span><span class="n">item</span><span class="p">()</span> <span class="c1"># sum up batch loss</span>
<span class="n">fsdp_loss</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="o">+=</span> <span class="nb">len</span><span class="p">(</span><span class="n">batch</span><span class="p">)</span>
<span class="k">if</span> <span class="n">rank</span><span class="o">==</span><span class="mi">0</span><span class="p">:</span>
<span class="n">inner_pbar</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
<span class="n">dist</span><span class="o">.</span><span class="n">all_reduce</span><span class="p">(</span><span class="n">fsdp_loss</span><span class="p">,</span> <span class="n">op</span><span class="o">=</span><span class="n">dist</span><span class="o">.</span><span class="n">ReduceOp</span><span class="o">.</span><span class="n">SUM</span><span class="p">)</span>
<span class="n">val_loss</span> <span class="o">=</span> <span class="n">fsdp_loss</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">/</span> <span class="n">fsdp_loss</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span>
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="n">inner_pbar</span><span class="o">.</span><span class="n">close</span><span class="p">()</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Validation Loss: </span><span class="si">{</span><span class="n">val_loss</span><span class="si">:</span><span class="s2">.4f</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="k">return</span> <span class="n">val_loss</span>
</pre></div>
</div>
<p>2.4 Define a distributed train function that wraps the model in FSDP:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">fsdp_main</span><span class="p">(</span><span class="n">args</span><span class="p">):</span>
<span class="n">model</span><span class="p">,</span> <span class="n">tokenizer</span> <span class="o">=</span> <span class="n">setup_model</span><span class="p">(</span><span class="s2">"t5-base"</span><span class="p">)</span>
<span class="n">local_rank</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'LOCAL_RANK'</span><span class="p">])</span>
<span class="n">rank</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'RANK'</span><span class="p">])</span>
<span class="n">world_size</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'WORLD_SIZE'</span><span class="p">])</span>
<span class="n">dataset</span> <span class="o">=</span> <span class="n">load_dataset</span><span class="p">(</span><span class="s1">'wikihow'</span><span class="p">,</span> <span class="s1">'all'</span><span class="p">,</span> <span class="n">data_dir</span><span class="o">=</span><span class="s1">'data/'</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">dataset</span><span class="o">.</span><span class="n">keys</span><span class="p">())</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Size of train dataset: "</span><span class="p">,</span> <span class="n">dataset</span><span class="p">[</span><span class="s1">'train'</span><span class="p">]</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Size of Validation dataset: "</span><span class="p">,</span> <span class="n">dataset</span><span class="p">[</span><span class="s1">'validation'</span><span class="p">]</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span>
<span class="c1">#wikihow(tokenizer, type_path, num_samples, input_length, output_length, print_text=False)</span>
<span class="n">train_dataset</span> <span class="o">=</span> <span class="n">wikihow</span><span class="p">(</span><span class="n">tokenizer</span><span class="p">,</span> <span class="s1">'train'</span><span class="p">,</span> <span class="mi">1500</span><span class="p">,</span> <span class="mi">512</span><span class="p">,</span> <span class="mi">150</span><span class="p">,</span> <span class="kc">False</span><span class="p">)</span>
<span class="n">val_dataset</span> <span class="o">=</span> <span class="n">wikihow</span><span class="p">(</span><span class="n">tokenizer</span><span class="p">,</span> <span class="s1">'validation'</span><span class="p">,</span> <span class="mi">300</span><span class="p">,</span> <span class="mi">512</span><span class="p">,</span> <span class="mi">150</span><span class="p">,</span> <span class="kc">False</span><span class="p">)</span>
<span class="n">sampler1</span> <span class="o">=</span> <span class="n">DistributedSampler</span><span class="p">(</span><span class="n">train_dataset</span><span class="p">,</span> <span class="n">rank</span><span class="o">=</span><span class="n">rank</span><span class="p">,</span> <span class="n">num_replicas</span><span class="o">=</span><span class="n">world_size</span><span class="p">,</span> <span class="n">shuffle</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="n">sampler2</span> <span class="o">=</span> <span class="n">DistributedSampler</span><span class="p">(</span><span class="n">val_dataset</span><span class="p">,</span> <span class="n">rank</span><span class="o">=</span><span class="n">rank</span><span class="p">,</span> <span class="n">num_replicas</span><span class="o">=</span><span class="n">world_size</span><span class="p">)</span>
<span class="n">setup</span><span class="p">()</span>
<span class="n">train_kwargs</span> <span class="o">=</span> <span class="p">{</span><span class="s1">'batch_size'</span><span class="p">:</span> <span class="n">args</span><span class="o">.</span><span class="n">batch_size</span><span class="p">,</span> <span class="s1">'sampler'</span><span class="p">:</span> <span class="n">sampler1</span><span class="p">}</span>
<span class="n">test_kwargs</span> <span class="o">=</span> <span class="p">{</span><span class="s1">'batch_size'</span><span class="p">:</span> <span class="n">args</span><span class="o">.</span><span class="n">test_batch_size</span><span class="p">,</span> <span class="s1">'sampler'</span><span class="p">:</span> <span class="n">sampler2</span><span class="p">}</span>
<span class="n">cuda_kwargs</span> <span class="o">=</span> <span class="p">{</span><span class="s1">'num_workers'</span><span class="p">:</span> <span class="mi">2</span><span class="p">,</span>
<span class="s1">'pin_memory'</span><span class="p">:</span> <span class="kc">True</span><span class="p">,</span>
<span class="s1">'shuffle'</span><span class="p">:</span> <span class="kc">False</span><span class="p">}</span>
<span class="n">train_kwargs</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">cuda_kwargs</span><span class="p">)</span>
<span class="n">test_kwargs</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">cuda_kwargs</span><span class="p">)</span>
<span class="n">train_loader</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">train_dataset</span><span class="p">,</span><span class="o">**</span><span class="n">train_kwargs</span><span class="p">)</span>
<span class="n">val_loader</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">val_dataset</span><span class="p">,</span> <span class="o">**</span><span class="n">test_kwargs</span><span class="p">)</span>
<span class="n">t5_auto_wrap_policy</span> <span class="o">=</span> <span class="n">functools</span><span class="o">.</span><span class="n">partial</span><span class="p">(</span>
<span class="n">transformer_auto_wrap_policy</span><span class="p">,</span>
<span class="n">transformer_layer_cls</span><span class="o">=</span><span class="p">{</span>
<span class="n">T5Block</span><span class="p">,</span>
<span class="p">},</span>
<span class="p">)</span>
<span class="n">sharding_strategy</span><span class="p">:</span> <span class="n">ShardingStrategy</span> <span class="o">=</span> <span class="n">ShardingStrategy</span><span class="o">.</span><span class="n">SHARD_GRAD_OP</span> <span class="c1">#for Zero2 and FULL_SHARD for Zero3</span>
<span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">set_device</span><span class="p">(</span><span class="n">local_rank</span><span class="p">)</span>
<span class="c1">#init_start_event = torch.cuda.Event(enable_timing=True)</span>
<span class="c1">#init_end_event = torch.cuda.Event(enable_timing=True)</span>
<span class="c1">#init_start_event.record()</span>
<span class="n">bf16_ready</span> <span class="o">=</span> <span class="p">(</span>
<span class="n">torch</span><span class="o">.</span><span class="n">version</span><span class="o">.</span><span class="n">cuda</span>
<span class="ow">and</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">is_bf16_supported</span><span class="p">()</span>
<span class="ow">and</span> <span class="n">LooseVersion</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">version</span><span class="o">.</span><span class="n">cuda</span><span class="p">)</span> <span class="o">>=</span> <span class="s2">"11.0"</span>
<span class="ow">and</span> <span class="n">dist</span><span class="o">.</span><span class="n">is_nccl_available</span><span class="p">()</span>
<span class="ow">and</span> <span class="n">nccl</span><span class="o">.</span><span class="n">version</span><span class="p">()</span> <span class="o">>=</span> <span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">10</span><span class="p">)</span>
<span class="p">)</span>
<span class="k">if</span> <span class="n">bf16_ready</span><span class="p">:</span>
<span class="n">mp_policy</span> <span class="o">=</span> <span class="n">bfSixteen</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">mp_policy</span> <span class="o">=</span> <span class="kc">None</span> <span class="c1"># defaults to fp32</span>
<span class="c1"># model is on CPU before input to FSDP</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">FSDP</span><span class="p">(</span><span class="n">model</span><span class="p">,</span>
<span class="n">auto_wrap_policy</span><span class="o">=</span><span class="n">t5_auto_wrap_policy</span><span class="p">,</span>
<span class="n">mixed_precision</span><span class="o">=</span><span class="n">mp_policy</span><span class="p">,</span>
<span class="c1">#sharding_strategy=sharding_strategy,</span>
<span class="n">device_id</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">current_device</span><span class="p">())</span>
<span class="n">optimizer</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">AdamW</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="n">args</span><span class="o">.</span><span class="n">lr</span><span class="p">)</span>
<span class="n">scheduler</span> <span class="o">=</span> <span class="n">StepLR</span><span class="p">(</span><span class="n">optimizer</span><span class="p">,</span> <span class="n">step_size</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">gamma</span><span class="o">=</span><span class="n">args</span><span class="o">.</span><span class="n">gamma</span><span class="p">)</span>
<span class="n">best_val_loss</span> <span class="o">=</span> <span class="nb">float</span><span class="p">(</span><span class="s2">"inf"</span><span class="p">)</span>
<span class="n">curr_val_loss</span> <span class="o">=</span> <span class="nb">float</span><span class="p">(</span><span class="s2">"inf"</span><span class="p">)</span>
<span class="n">file_save_name</span> <span class="o">=</span> <span class="s2">"T5-model-"</span>
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="n">time_of_run</span> <span class="o">=</span> <span class="n">get_date_of_run</span><span class="p">()</span>
<span class="n">dur</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">train_acc_tracking</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">val_acc_tracking</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">training_start_time</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span>
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span> <span class="ow">and</span> <span class="n">args</span><span class="o">.</span><span class="n">track_memory</span><span class="p">:</span>
<span class="n">mem_alloc_tracker</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">mem_reserved_tracker</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">args</span><span class="o">.</span><span class="n">epochs</span> <span class="o">+</span> <span class="mi">1</span><span class="p">):</span>
<span class="n">t0</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span>
<span class="n">train_accuracy</span> <span class="o">=</span> <span class="n">train</span><span class="p">(</span><span class="n">args</span><span class="p">,</span> <span class="n">model</span><span class="p">,</span> <span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="p">,</span> <span class="n">train_loader</span><span class="p">,</span> <span class="n">optimizer</span><span class="p">,</span> <span class="n">epoch</span><span class="p">,</span> <span class="n">sampler</span><span class="o">=</span><span class="n">sampler1</span><span class="p">)</span>
<span class="k">if</span> <span class="n">args</span><span class="o">.</span><span class="n">run_validation</span><span class="p">:</span>
<span class="n">curr_val_loss</span> <span class="o">=</span> <span class="n">validation</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="p">,</span> <span class="n">val_loader</span><span class="p">)</span>
<span class="n">scheduler</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"--> epoch </span><span class="si">{</span><span class="n">epoch</span><span class="si">}</span><span class="s2"> completed...entering save and stats zone"</span><span class="p">)</span>
<span class="n">dur</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span> <span class="o">-</span> <span class="n">t0</span><span class="p">)</span>
<span class="n">train_acc_tracking</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">train_accuracy</span><span class="o">.</span><span class="n">item</span><span class="p">())</span>
<span class="k">if</span> <span class="n">args</span><span class="o">.</span><span class="n">run_validation</span><span class="p">:</span>
<span class="n">val_acc_tracking</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">curr_val_loss</span><span class="o">.</span><span class="n">item</span><span class="p">())</span>
<span class="k">if</span> <span class="n">args</span><span class="o">.</span><span class="n">track_memory</span><span class="p">:</span>
<span class="n">mem_alloc_tracker</span><span class="o">.</span><span class="n">append</span><span class="p">(</span>
<span class="n">format_metrics_to_gb</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">memory_allocated</span><span class="p">())</span>
<span class="p">)</span>
<span class="n">mem_reserved_tracker</span><span class="o">.</span><span class="n">append</span><span class="p">(</span>
<span class="n">format_metrics_to_gb</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">memory_reserved</span><span class="p">())</span>
<span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"completed save and stats zone..."</span><span class="p">)</span>
<span class="k">if</span> <span class="n">args</span><span class="o">.</span><span class="n">save_model</span> <span class="ow">and</span> <span class="n">curr_val_loss</span> <span class="o"><</span> <span class="n">best_val_loss</span><span class="p">:</span>
<span class="c1"># save</span>
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"--> entering save model state"</span><span class="p">)</span>
<span class="n">save_policy</span> <span class="o">=</span> <span class="n">FullStateDictConfig</span><span class="p">(</span><span class="n">offload_to_cpu</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">rank0_only</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="k">with</span> <span class="n">FSDP</span><span class="o">.</span><span class="n">state_dict_type</span><span class="p">(</span>
<span class="n">model</span><span class="p">,</span> <span class="n">StateDictType</span><span class="o">.</span><span class="n">FULL_STATE_DICT</span><span class="p">,</span> <span class="n">save_policy</span>
<span class="p">):</span>
<span class="n">cpu_state</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">state_dict</span><span class="p">()</span>
<span class="c1">#print(f"saving process: rank {rank} done w state_dict")</span>
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"--> saving model ..."</span><span class="p">)</span>
<span class="n">currEpoch</span> <span class="o">=</span> <span class="p">(</span>
<span class="s2">"-"</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="n">epoch</span><span class="p">)</span> <span class="o">+</span> <span class="s2">"-"</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="nb">round</span><span class="p">(</span><span class="n">curr_val_loss</span><span class="o">.</span><span class="n">item</span><span class="p">(),</span> <span class="mi">4</span><span class="p">))</span> <span class="o">+</span> <span class="s2">".pt"</span>
<span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"--> attempting to save model prefix </span><span class="si">{</span><span class="n">currEpoch</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="n">save_name</span> <span class="o">=</span> <span class="n">file_save_name</span> <span class="o">+</span> <span class="s2">"-"</span> <span class="o">+</span> <span class="n">time_of_run</span> <span class="o">+</span> <span class="s2">"-"</span> <span class="o">+</span> <span class="n">currEpoch</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"--> saving as model name </span><span class="si">{</span><span class="n">save_name</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">cpu_state</span><span class="p">,</span> <span class="n">save_name</span><span class="p">)</span>
<span class="k">if</span> <span class="n">curr_val_loss</span> <span class="o"><</span> <span class="n">best_val_loss</span><span class="p">:</span>
<span class="n">best_val_loss</span> <span class="o">=</span> <span class="n">curr_val_loss</span>
<span class="k">if</span> <span class="n">rank</span><span class="o">==</span><span class="mi">0</span><span class="p">:</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"-->>>> New Val Loss Record: </span><span class="si">{</span><span class="n">best_val_loss</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="n">dist</span><span class="o">.</span><span class="n">barrier</span><span class="p">()</span>
<span class="n">cleanup</span><span class="p">()</span>
</pre></div>
</div>
<p>2.5 Parse the arguments and set the main function:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">if</span> <span class="vm">__name__</span> <span class="o">==</span> <span class="s1">'__main__'</span><span class="p">:</span>
<span class="c1"># Training settings</span>
<span class="n">parser</span> <span class="o">=</span> <span class="n">argparse</span><span class="o">.</span><span class="n">ArgumentParser</span><span class="p">(</span><span class="n">description</span><span class="o">=</span><span class="s1">'PyTorch T5 FSDP Example'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--batch-size'</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">int</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mi">4</span><span class="p">,</span> <span class="n">metavar</span><span class="o">=</span><span class="s1">'N'</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'input batch size for training (default: 64)'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--test-batch-size'</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">int</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mi">4</span><span class="p">,</span> <span class="n">metavar</span><span class="o">=</span><span class="s1">'N'</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'input batch size for testing (default: 1000)'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--epochs'</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">int</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mi">2</span><span class="p">,</span> <span class="n">metavar</span><span class="o">=</span><span class="s1">'N'</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'number of epochs to train (default: 3)'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--lr'</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">float</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mf">.002</span><span class="p">,</span> <span class="n">metavar</span><span class="o">=</span><span class="s1">'LR'</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'learning rate (default: .002)'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--gamma'</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">float</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mf">0.7</span><span class="p">,</span> <span class="n">metavar</span><span class="o">=</span><span class="s1">'M'</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'Learning rate step gamma (default: 0.7)'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--no-cuda'</span><span class="p">,</span> <span class="n">action</span><span class="o">=</span><span class="s1">'store_true'</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'disables CUDA training'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--seed'</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">int</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">metavar</span><span class="o">=</span><span class="s1">'S'</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'random seed (default: 1)'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--track_memory'</span><span class="p">,</span> <span class="n">action</span><span class="o">=</span><span class="s1">'store_false'</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'track the gpu memory'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--run_validation'</span><span class="p">,</span> <span class="n">action</span><span class="o">=</span><span class="s1">'store_false'</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'running the validation'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--save-model'</span><span class="p">,</span> <span class="n">action</span><span class="o">=</span><span class="s1">'store_false'</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'For Saving the current Model'</span><span class="p">)</span>
<span class="n">args</span> <span class="o">=</span> <span class="n">parser</span><span class="o">.</span><span class="n">parse_args</span><span class="p">()</span>
<span class="n">torch</span><span class="o">.</span><span class="n">manual_seed</span><span class="p">(</span><span class="n">args</span><span class="o">.</span><span class="n">seed</span><span class="p">)</span>
<span class="n">fsdp_main</span><span class="p">(</span><span class="n">args</span><span class="p">)</span>
</pre></div>
</div>
<p>To run the the training using torchrun:</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span>torchrun<span class="w"> </span>--nnodes<span class="w"> </span><span class="m">1</span><span class="w"> </span>--nproc_per_node<span class="w"> </span><span class="m">4</span><span class="w"> </span>T5_training.py
</pre></div>
</div>
</div>
<div class="section" id="transformer-wrapping-policy">
<span id="id1"></span><h2>Transformer Wrapping Policy<a class="headerlink" href="#transformer-wrapping-policy" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>As discussed in the <a class="reference external" href="https://tutorials.pytorch.kr/intermediate/FSDP_tutorial.html">previous tutorial</a>,
auto_wrap_policy is one of the FSDP features that make it easy to automatically
shard a given model and put the model, optimizer and gradient shards into
distinct FSDP units.</p>
<p>For some architectures such as Transformer encoder-decoders, some parts of the
model such as embedding table is being shared with both encoder and decoder. In
this case, we need to place the embedding table in the outer FSDP unit so that
it could be accessed from both encoder and decoder. In addition, by registering
the layer class for a transformer, the sharding plan can be made much more
communication efficient. In PyTorch 1.12, FSDP added this support and now we
have a wrapping policy for transfomers.</p>
<p>It can be created as follows, where the T5Block represents the T5 transformer
layer class (holding MHSA and FFN).</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">t5_auto_wrap_policy</span> <span class="o">=</span> <span class="n">functools</span><span class="o">.</span><span class="n">partial</span><span class="p">(</span>
<span class="n">transformer_auto_wrap_policy</span><span class="p">,</span>
<span class="n">transformer_layer_cls</span><span class="o">=</span><span class="p">{</span>
<span class="n">T5Block</span><span class="p">,</span>
<span class="p">},</span>
<span class="p">)</span>
<span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">set_device</span><span class="p">(</span><span class="n">local_rank</span><span class="p">)</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">FSDP</span><span class="p">(</span><span class="n">model</span><span class="p">,</span>
<span class="n">fsdp_auto_wrap_policy</span><span class="o">=</span><span class="n">t5_auto_wrap_policy</span><span class="p">)</span>
</pre></div>
</div>
<p>To see the wrapped model, you can easily print the model and visually inspect
the sharding and FSDP units as well.</p>
</div>
<div class="section" id="mixed-precision">
<h2>Mixed Precision<a class="headerlink" href="#mixed-precision" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>FSDP supports flexible mixed precision training allowing for arbitrary reduced
precision types (such as fp16 or bfloat16). Currently BFloat16 is only available
on Ampere GPUs, so you need to confirm native support before you use it. On
V100s for example, BFloat16 can still be run but due to it running non-natively,
it can result in significant slowdowns.</p>
<p>To check if BFloat16 is natively supported, you can use the following :</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">bf16_ready</span> <span class="o">=</span> <span class="p">(</span>
<span class="n">torch</span><span class="o">.</span><span class="n">version</span><span class="o">.</span><span class="n">cuda</span>
<span class="ow">and</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">is_bf16_supported</span><span class="p">()</span>
<span class="ow">and</span> <span class="n">LooseVersion</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">version</span><span class="o">.</span><span class="n">cuda</span><span class="p">)</span> <span class="o">>=</span> <span class="s2">"11.0"</span>
<span class="ow">and</span> <span class="n">dist</span><span class="o">.</span><span class="n">is_nccl_available</span><span class="p">()</span>
<span class="ow">and</span> <span class="n">nccl</span><span class="o">.</span><span class="n">version</span><span class="p">()</span> <span class="o">>=</span> <span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">10</span><span class="p">)</span>
<span class="p">)</span>
</pre></div>
</div>
<p>One of the advantages of mixed percision in FSDP is providing granular control
over different precision levels for parameters, gradients, and buffers as
follows:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">fpSixteen</span> <span class="o">=</span> <span class="n">MixedPrecision</span><span class="p">(</span>
<span class="n">param_dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float16</span><span class="p">,</span>
<span class="c1"># Gradient communication precision.</span>
<span class="n">reduce_dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float16</span><span class="p">,</span>
<span class="c1"># Buffer precision.</span>
<span class="n">buffer_dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float16</span><span class="p">,</span>
<span class="p">)</span>
<span class="n">bfSixteen</span> <span class="o">=</span> <span class="n">MixedPrecision</span><span class="p">(</span>
<span class="n">param_dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">bfloat16</span><span class="p">,</span>
<span class="c1"># Gradient communication precision.</span>
<span class="n">reduce_dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">bfloat16</span><span class="p">,</span>
<span class="c1"># Buffer precision.</span>
<span class="n">buffer_dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">bfloat16</span><span class="p">,</span>
<span class="p">)</span>
<span class="n">fp32_policy</span> <span class="o">=</span> <span class="n">MixedPrecision</span><span class="p">(</span>
<span class="n">param_dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">,</span>
<span class="c1"># Gradient communication precision.</span>
<span class="n">reduce_dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">,</span>
<span class="c1"># Buffer precision.</span>
<span class="n">buffer_dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">,</span>
<span class="p">)</span>
</pre></div>
</div>
<p>Note that if a certain type (parameter, reduce, buffer) is not specified, they
will not be casted at all.</p>
<p>This flexibility allows users fine grained control, such as only setting
gradient communication to happen in reduced precision, and all parameters /
buffer computation to be done in full precision. This is potentially useful in
cases where intra-node communication is the main bottleneck and parameters /
buffers must be in full precision to avoid accuracy issues. This can be done
with the following policy:</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span><span class="nv">grad_bf16</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>MixedPrecision<span class="o">(</span><span class="nv">reduce_dtype</span><span class="o">=</span>torch.bfloat16<span class="o">)</span>
</pre></div>
</div>
<p>In 2.4 we just add the relevant mixed precision policy to the FSDP wrapper:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">model</span> <span class="o">=</span> <span class="n">FSDP</span><span class="p">(</span><span class="n">model</span><span class="p">,</span>
<span class="n">auto_wrap_policy</span><span class="o">=</span><span class="n">t5_auto_wrap_policy</span><span class="p">,</span>
<span class="n">mixed_precision</span><span class="o">=</span><span class="n">bfSixteen</span><span class="p">)</span>
</pre></div>
</div>
<p>In our experiments, we have observed up to 4x speed up by using BFloat16 for
training and memory reduction of approximately 30% in some experiments that can
be used for batch size increases.</p>
</div>
<div class="section" id="intializing-fsdp-model-on-device">
<h2>Intializing FSDP Model on Device<a class="headerlink" href="#intializing-fsdp-model-on-device" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>In 1.12, FSDP supports a <cite>device_id</cite> argument meant to initialize input CPU
module on the device given by <cite>device_id</cite>. This is useful when the entire model
does not fit on a single GPU, but fits in a hostโs CPU memory. When <cite>device_id</cite>
is specified, FSDP will move the model to the specified device on a per-FSDP
unit basis, avoiding GPU OOM issues while initializing several times faster than
CPU-based initialization:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">set_device</span><span class="p">(</span><span class="n">local_rank</span><span class="p">)</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">FSDP</span><span class="p">(</span><span class="n">model</span><span class="p">,</span>
<span class="n">auto_wrap_policy</span><span class="o">=</span><span class="n">t5_auto_wrap_policy</span><span class="p">,</span>
<span class="n">mixed_precision</span><span class="o">=</span><span class="n">bfSixteen</span><span class="p">,</span>
<span class="n">device_id</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">current_device</span><span class="p">())</span>
</pre></div>
</div>
</div>
<div class="section" id="sharding-strategy">
<h2>Sharding Strategy<a class="headerlink" href="#sharding-strategy" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>FSDP sharding strategy by default is set to fully shard the model parameters,
gradients and optimizer states get sharded across all ranks. (also termed Zero3
sharding). In case you are interested to have the Zero2 sharding strategy, where
only optimizer states and gradients are sharded, FSDP support this feature by
passing the Sharding strategy by using ใShardingStrategy.SHARD_GRAD_OPใ,
instead of ใShardingStrategy.FULL_SHARDใ to the FSDP initialization as follows:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">set_device</span><span class="p">(</span><span class="n">local_rank</span><span class="p">)</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">FSDP</span><span class="p">(</span><span class="n">model</span><span class="p">,</span>
<span class="n">auto_wrap_policy</span><span class="o">=</span><span class="n">t5_auto_wrap_policy</span><span class="p">,</span>
<span class="n">mixed_precision</span><span class="o">=</span><span class="n">bfSixteen</span><span class="p">,</span>
<span class="n">device_id</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">current_device</span><span class="p">(),</span>
<span class="n">sharding_strategy</span><span class="o">=</span><span class="n">ShardingStrategy</span><span class="o">.</span><span class="n">SHARD_GRAD_OP</span> <span class="c1"># ZERO2)</span>
</pre></div>
</div>
<p>This will reduce the communication overhead in FSDP, in this case, it holds full
parameters after forward and through the backwards pass.</p>
<p>This saves an all_gather during backwards so there is less communication at the
cost of a higher memory footprint. Note that full model params are freed at the
end of backwards and all_gather will happen on the next forward pass.</p>
</div>
<div class="section" id="backward-prefetch">
<h2>Backward Prefetch<a class="headerlink" href="#backward-prefetch" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>The backward prefetch setting controls the timing of when the next FSDP unitโs
parameters should be requested. By setting it to <cite>BACKWARD_PRE</cite>, the next
FSDPโs unit params can begin to be requested and arrive sooner before the
computation of the current unit starts. This overlaps the <cite>all_gather</cite>
communication and gradient computation which can increase the training speed in
exchange for slightly higher memory consumption. It can be utilized in the FSDP