-
Notifications
You must be signed in to change notification settings - Fork 257
/
Copy pathddp_tutorial.html
1076 lines (866 loc) ยท 87.7 KB
/
ddp_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="๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ์์ํ๊ธฐ" />
<meta property="og:type" content="article" />
<meta property="og:url" content="https://tutorials.pytorch.kr/intermediate/ddp_tutorial.html" />
<meta property="og:site_name" content="PyTorch Tutorials KR" />
<meta property="og:description" content="์ ์: Shen Li ๊ฐ์: Joe Zhu ๋ฒ์ญ: ์กฐ๋ณ๊ทผ ์ ์๊ณผ๋ชฉ(Prerequisites): PyTorch ๋ถ์ฐ ์ฒ๋ฆฌ ๊ฐ์, ๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ API ๋ฌธ์, ๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ๋ฌธ์. ๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ DistributedDataParallel(DDP)๋ ์ฌ๋ฌ ๊ธฐ๊ธฐ์์ ์คํํ ์ ์๋ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ๋ฅผ ๋ชจ๋ ์์ค์์ ๊ตฌํํฉ๋๋ค. DDP๋ฅผ ์ฌ์ฉํ๋ ์ดํ๋ฆฌ์ผ์ด์
์ ์ฌ๋ฌ ์์
(process)์ ์์ฑํ๊ณ ์์
๋น ๋จ์ผ DDP ์ธ์คํด์ค๋ฅผ ์์ฑํด์ผ ํฉ๋๋ค. DDP๋ torch.distributed ํจํค์ง์ ์งํฉ ํต์ (c..." />
<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="์ ์: Shen Li ๊ฐ์: Joe Zhu ๋ฒ์ญ: ์กฐ๋ณ๊ทผ ์ ์๊ณผ๋ชฉ(Prerequisites): PyTorch ๋ถ์ฐ ์ฒ๋ฆฌ ๊ฐ์, ๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ API ๋ฌธ์, ๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ๋ฌธ์. ๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ DistributedDataParallel(DDP)๋ ์ฌ๋ฌ ๊ธฐ๊ธฐ์์ ์คํํ ์ ์๋ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ๋ฅผ ๋ชจ๋ ์์ค์์ ๊ตฌํํฉ๋๋ค. DDP๋ฅผ ์ฌ์ฉํ๋ ์ดํ๋ฆฌ์ผ์ด์
์ ์ฌ๋ฌ ์์
(process)์ ์์ฑํ๊ณ ์์
๋น ๋จ์ผ DDP ์ธ์คํด์ค๋ฅผ ์์ฑํด์ผ ํฉ๋๋ค. DDP๋ torch.distributed ํจํค์ง์ ์งํฉ ํต์ (c..." />
<meta property="og:ignore_canonical" content="true" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ์์ํ๊ธฐ — ํ์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="PyTorch๋ก ๋ถ์ฐ ์ดํ๋ฆฌ์ผ์ด์
๊ฐ๋ฐํ๊ธฐ" href="dist_tuto.html" />
<link rel="prev" title="๋จ์ผ ๋จธ์ ์ ์ฌ์ฉํ ๋ชจ๋ธ ๋ณ๋ ฌํ ๋ชจ๋ฒ ์ฌ๋ก" href="model_parallel_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 current"><a class="current reference internal" href="#">๋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"><a class="reference internal" href="FSDP_adavnced_tutorial.html">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>๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ์์ํ๊ธฐ</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/intermediate/ddp_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/ddp_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="id1">
<h1>๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ์์ํ๊ธฐ<a class="headerlink" href="#id1" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h1>
<p><strong>์ ์</strong>: <a class="reference external" href="https://mrshenli.github.io/">Shen Li</a></p>
<p><strong>๊ฐ์</strong>: <a class="reference external" href="https://github.com/gunandrose4u">Joe Zhu</a></p>
<p><strong>๋ฒ์ญ</strong>: <a class="reference external" href="https://github.com/Jo-byung-geun">์กฐ๋ณ๊ทผ</a></p>
<div class="admonition note">
<p class="admonition-title">์ฐธ๊ณ </p>
<p><a class="reference internal" href="../_images/pencil-16.png"><img alt="edit" src="../_images/pencil-16.png" style="width: 16px; height: 16px;" /></a> ์ด ํํ ๋ฆฌ์ผ์ ์์ค ์ฝ๋๋ <a class="reference external" href="https://github.com/pytorch/tutorials/blob/main/intermediate_source/ddp_tutorial.rst">GitHub</a> ์์ ํ์ธํ๊ณ ๋ณ๊ฒฝํด ๋ณผ ์ ์์ต๋๋ค.</p>
</div>
<p>์ ์๊ณผ๋ชฉ(Prerequisites):</p>
<ul class="simple">
<li><p><a class="reference external" href="../beginner/dist_overview.html">PyTorch ๋ถ์ฐ ์ฒ๋ฆฌ ๊ฐ์</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/docs/master/generated/torch.nn.parallel.DistributedDataParallel.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ API ๋ฌธ์</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/docs/master/notes/ddp.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ๋ฌธ์</a></p></li>
</ul>
<p><a class="reference external" href="https://pytorch.org/docs/stable/nn.html#module-torch.nn.parallel">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ DistributedDataParallel</a>(DDP)๋
์ฌ๋ฌ ๊ธฐ๊ธฐ์์ ์คํํ ์ ์๋ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ๋ฅผ ๋ชจ๋ ์์ค์์ ๊ตฌํํฉ๋๋ค.
DDP๋ฅผ ์ฌ์ฉํ๋ ์ดํ๋ฆฌ์ผ์ด์
์ ์ฌ๋ฌ ์์
(process)์ ์์ฑํ๊ณ ์์
๋น ๋จ์ผ DDP ์ธ์คํด์ค๋ฅผ ์์ฑํด์ผ ํฉ๋๋ค.
DDP๋ <a class="reference external" href="https://tutorials.pytorch.kr/intermediate/dist_tuto.html">torch.distributed</a>
ํจํค์ง์ ์งํฉ ํต์ (collective communication)์ ์ฌ์ฉํ์ฌ ๋ณํ๋(gradient)์ ๋ฒํผ๋ฅผ ๋๊ธฐํํฉ๋๋ค.
์ข ๋ ๊ตฌ์ฒด์ ์ผ๋ก, DDP๋ <code class="docutils literal notranslate"><span class="pre">model.parameters()</span></code>์ ์ํด ์ฃผ์ด์ง ๊ฐ ํ๋ผ๋ฏธํฐ์ ๋ํด Autograd hook์ ๋ฑ๋กํ๊ณ ,
hook์ ์ญ๋ฐฉํฅ ์ ๋ฌ์์ ํด๋น ๋ณํ๋๊ฐ ๊ณ์ฐ๋ ๋ ์๋ํฉ๋๋ค.
๋ค์์ผ๋ก DDP๋ ์ด ์ ํธ๋ฅผ ์ฌ์ฉํ์ฌ ์์
๊ฐ์ ๋ณํ๋ ๋๊ธฐํ๋ฅผ ๋ฐ์์ํต๋๋ค. ์์ธํ ๋ด์ฉ์
<a class="reference external" href="https://pytorch.org/docs/master/notes/ddp.html">DDP design note</a>๋ฅผ ์ฐธ์กฐํ์ญ์์ค.</p>
<p>DDP์ ๊ถ์ฅ ์ฌ์ฉ๋ฒ์, ์ฌ๋ฌ ์ฅ์น์ ์์ ์ ์๋ ๊ฐ ๋ชจ๋ธ ๋ณต์ ๋ณธ๋น ํ๋์ ์์
์ ์์ฑํ๋ ๊ฒ์
๋๋ค.
DDP ์์
์ ๋์ผํ ๊ธฐ๊ธฐ ๋๋ ์ฌ๋ฌ ๊ธฐ๊ธฐ์ ๋ฐฐ์นํ ์ ์์ง๋ง GPU ์ฅ์น๋ ์์
๊ฐ์ ๊ณต์ ํ ์ ์์ต๋๋ค.
์ด ํํ ๋ฆฌ์ผ์์๋ ๊ธฐ๋ณธ DDP ์ฌ์ฉ ์ฌ๋ก์์ ์์ํ์ฌ,
checkpointing ๋ชจ๋ธ ๋ฐ DDP์ ๋ชจ๋ธ ๋ณ๋ ฌ ์ฒ๋ฆฌ์ ๊ฒฐํฉ์ ํฌํจํ ์ถ๊ฐ์ ์ธ ์ฌ์ฉ ์ฌ๋ก๋ฅผ ๋ณด์ฌ์ค๋๋ค.</p>
<div class="admonition note">
<p class="admonition-title">์ฐธ๊ณ </p>
<p>์ด ํํ ๋ฆฌ์ผ์ ์ฝ๋๋ 8-GPU ์๋ฒ์์ ์คํ๋์ง๋ง ๋ค๋ฅธ ํ๊ฒฝ์์๋ ์ฝ๊ฒ ์ ์ฉํ ์ ์์ต๋๋ค.</p>
</div>
<p><code class="docutils literal notranslate"><span class="pre">DataParallel</span></code>๊ณผ <code class="docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code> ๊ฐ์ ๋น๊ต
โโโโโโโโโโโโโโโโโโโ-===</p>
<p>๋ด์ฉ์ ๋ค์ด๊ฐ๊ธฐ์ ์์ ๋ณต์ก์ฑ์ด ์ฆ๊ฐํ์์๋ ๋ถ๊ตฌํ๊ณ
<code class="docutils literal notranslate"><span class="pre">DataParallel</span></code>์ <code class="docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code> ์ฌ์ฉ์ ๊ณ ๋ คํ๋ ์ด์ ๋ฅผ ์๊ฐํด๋ด
์๋ค.</p>
<ul class="simple">
<li><p>์ฒซ์งธ, <code class="docutils literal notranslate"><span class="pre">DataParallel</span></code>์ ๋จ์ผ ์์
, ๋ฉํฐ์ฐ๋ ๋์ด๋ฉฐ ๋จ์ผ ๊ธฐ๊ธฐ์์๋ง ์๋ํ๋ ๋ฐ๋ฉด,
<code class="docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code>์ ๋ค์ค ์์
์ด๋ฉฐ ๋จ์ผ ๋ฐ ๋ค์ค ๊ธฐ๊ธฐ ํ์ต์ ์ ๋ถ ์ง์ํฉ๋๋ค.
<code class="docutils literal notranslate"><span class="pre">DataParallel</span></code>์ ์ฐ๋ ๋๊ฐ GIL ๊ฒฝํฉ, ๋ณต์ ๋ชจ๋ธ์ ๋ฐ๋ณต ๋น ์์ฑ, ์ฐ๋ ์
๋ ฅ ๋ฐ ์์ง ์ถ๋ ฅ์ผ๋ก ์ธํ
์ถ๊ฐ์ ์ธ ์ค๋ฒํค๋๋ก ์ธํด ์ผ๋ฐ์ ์ผ๋ก ๋จ์ผ ์์คํ
์์์กฐ์ฐจ <code class="docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code>๋ณด๋ค ๋๋ฆฝ๋๋ค.</p></li>
<li><p>๋ชจ๋ธ์ด ๋๋ฌด ์ปค์ ๋จ์ผ GPU์ ๋ง์ง ์์ ๊ฒฝ์ฐ <strong>model parallel</strong>์ ์ฌ์ฉํ์ฌ ์ฌ๋ฌ GPU๋ก ๋ถํ ํด์ผ ํ๋ค๋
<a class="reference external" href="https://tutorials.pytorch.kr/intermediate/model_parallel_tutorial.html">prior tutorial</a>์ ๋ ์ฌ๋ ค ๋ณด์ธ์.
<code class="docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code>์ <strong>model parallel</strong>์์ ์คํ๋์ง๋ง <code class="docutils literal notranslate"><span class="pre">DataParallel</span></code>์ ์ด๋ ์คํ๋์ง ์์ต๋๋ค.
DDP๋ฅผ ๋ชจ๋ธ ๋ณ๋ ฌ ์ฒ๋ฆฌ์ ๊ฒฐํฉํ๋ฉด ๊ฐ DDP ์์
์ ๋ชจ๋ธ ๋ณ๋ ฌ ์ฒ๋ฆฌ๋ฅผ ์ฌ์ฉํ๋ฉฐ
๋ชจ๋ ์์
์ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ๋ฅผ ์ฌ์ฉํฉ๋๋ค.</p></li>
<li><p>๋ชจ๋ธ์ด ์ฌ๋ฌ ๋์ ๊ธฐ๊ธฐ์ ์กด์ฌํด์ผ ํ๊ฑฐ๋ ์ฌ์ฉ ์ฌ๋ก๊ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌํ ํจ๋ฌ๋ค์์ ๋ง์ง ์๋ ๊ฒฝ์ฐ,
์ผ๋ฐ์ ์ธ ๋ถ์ฐ ํ์ต ์ง์์ ๋ณด๋ ค๋ฉด <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html">the RPC API</a>๋ฅผ ์ฐธ์กฐํ์ญ์์ค.</p></li>
</ul>
<div class="section" id="id3">
<h2>๊ธฐ๋ณธ์ ์ธ ์ฌ์ฉ๋ฒ<a class="headerlink" href="#id3" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>DDP ๋ชจ๋์ ์์ฑํ๊ธฐ ์ ์ ๋ฐ๋์ ์ฐ์ ์์
๊ทธ๋ฃน์ ์ฌ๋ฐ๋ฅด๊ฒ ์ค์ ํด์ผ ํฉ๋๋ค. ์์ธํ ๋ด์ฉ์
<a class="reference external" href="https://tutorials.pytorch.kr/intermediate/dist_tuto.html">PYTORCH๋ก ๋ถ์ฐ ์ดํ๋ฆฌ์ผ์ด์
๊ฐ๋ฐํ๊ธฐ</a>์์ ํ์ธํ ์ ์์ต๋๋ค.</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">sys</span>
<span class="kn">import</span> <span class="nn">tempfile</span>
<span class="kn">import</span> <span class="nn">torch</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.nn</span> <span class="k">as</span> <span class="nn">nn</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">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="c1"># ์๋์ฐ ํ๋ซํผ์์ torch.distributed ํจํค์ง๋</span>
<span class="c1"># Gloo backend, FileStore ๋ฐ TcpStore ๋ง์ ์ง์ํฉ๋๋ค.</span>
<span class="c1"># FileStore์ ๊ฒฝ์ฐ, init_process_group ์์</span>
<span class="c1"># init_method ๋งค๊ฐ๋ณ์๋ฅผ ๋ก์ปฌ ํ์ผ๋ก ์ค์ ํฉ๋๋ค.</span>
<span class="c1"># ๋ค์ ์์:</span>
<span class="c1"># init_method="file:///f:/libtmp/some_file"</span>
<span class="c1"># dist.init_process_group(</span>
<span class="c1"># "gloo",</span>
<span class="c1"># rank=rank,</span>
<span class="c1"># init_method=init_method,</span>
<span class="c1"># world_size=world_size)</span>
<span class="c1"># TcpStore์ ๊ฒฝ์ฐ ๋ฆฌ๋
์ค์ ๋์ผํ ๋ฐฉ์์
๋๋ค.</span>
<span class="k">def</span> <span class="nf">setup</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">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'MASTER_ADDR'</span><span class="p">]</span> <span class="o">=</span> <span class="s1">'localhost'</span>
<span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'MASTER_PORT'</span><span class="p">]</span> <span class="o">=</span> <span class="s1">'12355'</span>
<span class="c1"># ์์
๊ทธ๋ฃน ์ด๊ธฐํ</span>
<span class="n">dist</span><span class="o">.</span><span class="n">init_process_group</span><span class="p">(</span><span class="s2">"gloo"</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">world_size</span><span class="o">=</span><span class="n">world_size</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>์ด์ DDP๋ก ๊ฐ์ธ์ฌ์ง Toy ๋ชจ๋์ ์์ฑํ๊ณ ๋๋ฏธ ์
๋ ฅ ๋ฐ์ดํฐ๋ฅผ ์
๋ ฅํด ๋ณด๊ฒ ์ต๋๋ค.
์ฐ์ DDP๋ 0์์ ์์
์์๋ถํฐ DDP ์์ฑ์์ ๋ค๋ฅธ ๋ชจ๋ ์์
๋ค์๊ฒ ๋ชจ๋ธ์ ์ํ๋ฅผ ์ ๋ฌํ๋ฏ๋ก,
๋ค๋ฅธ ๋ชจ๋ธ์ ๋งค๊ฐ ๋ณ์ ์ด๊ธฐ๊ฐ๋ค์์ ์์ํ๋ ๋ค๋ฅธ DDP ์์
๋ค์ ๋ํ์ฌ ๊ฑฑ์ ํ ํ์๊ฐ ์์ต๋๋ค.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">ToyModel</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">ToyModel</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">net1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">10</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">relu</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">net2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">5</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">net2</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">net1</span><span class="p">(</span><span class="n">x</span><span class="p">)))</span>
<span class="k">def</span> <span class="nf">demo_basic</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="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Running basic DDP example on rank </span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s2">."</span><span class="p">)</span>
<span class="n">setup</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="c1"># ๋ชจ๋ธ์ ์์ฑํ๊ณ ์์ ์์ด๋๊ฐ ์๋ GPU๋ก ์ ๋ฌ</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">ToyModel</span><span class="p">()</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">rank</span><span class="p">)</span>
<span class="n">ddp_model</span> <span class="o">=</span> <span class="n">DDP</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">device_ids</span><span class="o">=</span><span class="p">[</span><span class="n">rank</span><span class="p">])</span>
<span class="n">loss_fn</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">MSELoss</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">SGD</span><span class="p">(</span><span class="n">ddp_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="mf">0.001</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">outputs</span> <span class="o">=</span> <span class="n">ddp_model</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">20</span><span class="p">,</span> <span class="mi">10</span><span class="p">))</span>
<span class="n">labels</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">20</span><span class="p">,</span> <span class="mi">5</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">rank</span><span class="p">)</span>
<span class="n">loss_fn</span><span class="p">(</span><span class="n">outputs</span><span class="p">,</span> <span class="n">labels</span><span class="p">)</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">cleanup</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">run_demo</span><span class="p">(</span><span class="n">demo_fn</span><span class="p">,</span> <span class="n">world_size</span><span class="p">):</span>
<span class="n">mp</span><span class="o">.</span><span class="n">spawn</span><span class="p">(</span><span class="n">demo_fn</span><span class="p">,</span>
<span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">world_size</span><span class="p">,),</span>
<span class="n">nprocs</span><span class="o">=</span><span class="n">world_size</span><span class="p">,</span>
<span class="n">join</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
</pre></div>
</div>
<p>๋ณด์ฌ์ง๋ ๋ฐ์ ๊ฐ์ด DDP๋ ํ์ ์์ค์ ๋ถ์ฐ ์ปค๋ฎค๋์ผ์ด์
์ธ๋ถ ์ฌํญ์ ํฌํจํ๊ณ
๋ก์ปฌ ๋ชจ๋ธ์ฒ๋ผ ๊น๋ํ API๋ฅผ ์ ๊ณตํฉ๋๋ค. ๋ณํ๋ ๋๊ธฐํ ํต์ (gradient synchronization communications)์
์ญ์ ํ ์ ๋ฌ(backward pass)๊ฐ ์ํ๋๋ฉฐ ์ญ์ ํ ๊ณ์ฐ(backward computation)๊ณผ ๊ฒน์น๊ฒ ๋ฉ๋๋ค.
<code class="docutils literal notranslate"><span class="pre">backword()</span></code>๊ฐ ๋ฐํ๋๋ฉด <code class="docutils literal notranslate"><span class="pre">param.grad</span></code>์๋ ๋๊ธฐํ๋ ๋ณํ๋ ํ
์(synchronized gradient tensor)๊ฐ ํฌํจ๋์ด ์์ต๋๋ค.
๊ธฐ๋ณธ์ ์ผ๋ก DDP๋ ์์
๊ทธ๋ฃน์ ์ค์ ํ๋๋ฐ ๋ช ๊ฐ์ LoCs๋ง์ด ํ์ํ์ง๋ง ๋ณด๋ค ๋ค์ํ๊ฒ ์ฌ์ฉํ๋ ๊ฒฝ์ฐ ์ฃผ์๊ฐ ํ์ํฉ๋๋ค.</p>
</div>
<div class="section" id="id4">
<h2>๋น๋์นญ ์์
์๋<a class="headerlink" href="#id4" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>DDP์์๋ ์์ฑ์, ์์ ํ(forward pass) ๋ฐ ์ญ์ ํ ์ ๋ฌ ํธ์ถ ์ง์ ์ด ๋ถ์ฐ ๋๊ธฐํ ์ง์ (distribute synchronization point)์
๋๋ค.
์๋ก ๋ค๋ฅธ ์์
์ด ๋์ผํ ์์ ๋๊ธฐํ๋ฅผ ์์ํ๊ณ ๋์ผํ ์์๋ก ์ด๋ฌํ ๋๊ธฐํ ์ง์ ์ ๋๋ฌํ์ฌ
๊ฐ ๋๊ธฐํ ์ง์ ์ ๊ฑฐ์ ๋์์ ์ง์
์ ์๊ตฌํฉ๋๋ค.
๊ทธ๋ ์ง ์์ผ๋ฉด ๋น ๋ฅธ ์์
์ด ์ผ์ฐ ๋์ฐฉํ๊ณ ๋ค๋ฅธ ์์
์ ๋๊ธฐ ์๊ฐ์ด ์ด๊ณผ๋ ์ ์์ต๋๋ค.
๋ฐ๋ผ์ ์ฌ์ฉ์๋ ์์
๊ฐ์ ์์
๋์ ๊ท ํ ์๊ฒ ๋ถ๋ฐฐํ ํ์๊ฐ ์์ต๋๋ค.
๋๋๋ก ๋น๋์นญ ์์
(skewed processing) ์๋๋ ๋ค์๊ณผ ๊ฐ์ ์ด์ ๋ก ์ธํ์ฌ ๋ถ๊ฐํผํ๊ฒ ๋ฐ์ํฉ๋๋ค.
์๋ฅผ ๋ค์ด, ๋คํธ์ํฌ ์ง์ฐ, ๋ฆฌ์์ค ๊ฒฝ์(resource contentions), ์์ธกํ์ง ๋ชปํ ์์
๋ ๊ธ์ฆ ๋ฑ์
๋๋ค.
์ด๋ฌํ ์ํฉ์์ ์๊ฐ ์ด๊ณผ๋ฅผ ๋ฐฉ์งํ๋ ค๋ฉด, <a class="reference external" href="https://pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group">init_process_group</a>๋ฅผ
ํธ์ถํ ๋ ์ถฉ๋ถํ <code class="docutils literal notranslate"><span class="pre">timeout</span></code>๊ฐ์ ์ ๋ฌํด์ผ ํฉ๋๋ค.</p>
</div>
<div class="section" id="id5">
<h2>์ฒดํฌํฌ์ธํธ๋ฅผ ์ ์ฅํ๊ณ ์ฝ์ด์ค๊ธฐ<a class="headerlink" href="#id5" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>ํ์ต ์ค์ <code class="docutils literal notranslate"><span class="pre">torch.save</span></code>์ <code class="docutils literal notranslate"><span class="pre">torch.load</span></code> ๋ก ๋ชจ๋์ ์ฒดํฌํฌ์ธํธ๋ฅผ ๋ง๋ค๊ณ ๊ทธ ์ฒดํฌํฌ์ธํธ๋ก๋ถํฐ ๋ณต๊ตฌํ๋ ๊ฒ์ด ์ผ๋ฐ์ ์
๋๋ค.
๋ ์์ธํ ๋ด์ฉ์ <a class="reference external" href="https://tutorials.pytorch.kr/beginner/saving_loading_models.html">SAVING AND LOADING MODELS</a>๋ฅผ ์ฐธ๊ณ ํ์ธ์.
DDP๋ฅผ ์ฌ์ฉํ ๋, ์ต์ ์ ๋ฐฉ๋ฒ์ ๋ชจ๋ธ์ ํ ์์
์๋ง ์ ์ฅํ๊ณ
๊ทธ ๋ชจ๋ธ์ ๋ชจ๋ ์์
์ ์ฐ๊ธฐ ๊ณผ๋ถํ(write overhead)๋ฅผ ์ค์ด๋ฉฐ ์ฝ์ด์ค๋ ๊ฒ์
๋๋ค.
์ด๋ ๋ชจ๋ ์์
์ด ๊ฐ์ ๋งค๊ฐ๋ณ์๋ก๋ถํฐ ์์๋๊ณ ๋ณํ๋๋
์ญ์ ํ ์ ๋ฌ๋ก ๋๊ธฐํ๋๋ฏ๋ก ์ตํฐ๋ง์ด์ (optimizer)๋
๋งค๊ฐ๋ณ์๋ฅผ ๋์ผํ ๊ฐ์ผ๋ก ๊ณ์ ์ค์ ํด์ผ ํ๊ธฐ ๋๋ฌธ์ ์ ํํฉ๋๋ค. ์ด๋ฌํ ์ต์ ํ๋ฅผ ์ฌ์ฉํ๋ ๊ฒฝ์ฐ,
์ ์ฅ์ด ์๋ฃ๋๊ธฐ ์ ์ ๋ถ๋ฌ์ค๋ ์ด๋ ํ ์์
๋ ์์ํ์ง ์๋๋ก ํด์ผ ํฉ๋๋ค. ๋๋ถ์ด, ๋ชจ๋์ ์ฝ์ด์ฌ ๋
์์
์ด ๋ค๋ฅธ ๊ธฐ๊ธฐ์ ์ ๊ทผํ์ง ์๋๋ก ์ ์ ํ <code class="docutils literal notranslate"><span class="pre">map_location</span></code> ์ธ์๋ฅผ ์ ๊ณตํด์ผํฉ๋๋ค.
<code class="docutils literal notranslate"><span class="pre">map_location</span></code>๊ฐ์ด ์์ ๊ฒฝ์ฐ, <code class="docutils literal notranslate"><span class="pre">torch.load</span></code>๋ ๋จผ์ ๋ชจ๋์ CPU์ ์ฝ์ด์จ ๋ค์ ๊ฐ ๋งค๊ฐ๋ณ์๊ฐ
์ ์ฅ๋ ์์น๋ก ๋ณต์ฌํ์ฌ ๋์ผํ ์ฅ์น๋ฅผ ์ฌ์ฉํ๋ ๋์ผํ ๊ธฐ๊ธฐ์์ ๋ชจ๋ ์์
์ ๋ฐ์์ํต๋๋ค.
๋ ์ถ๊ฐ์ ์ธ ์คํจ ๋ณต๊ตฌ์ ์๋ผ์คํฑ(elasticity support)์ <a class="reference external" href="https://pytorch.org/elastic">TorchElastic</a>์ ์ฐธ๊ณ ํ์ธ์.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">demo_checkpoint</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="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Running DDP checkpoint example on rank </span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s2">."</span><span class="p">)</span>
<span class="n">setup</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">model</span> <span class="o">=</span> <span class="n">ToyModel</span><span class="p">()</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">rank</span><span class="p">)</span>
<span class="n">ddp_model</span> <span class="o">=</span> <span class="n">DDP</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">device_ids</span><span class="o">=</span><span class="p">[</span><span class="n">rank</span><span class="p">])</span>
<span class="n">CHECKPOINT_PATH</span> <span class="o">=</span> <span class="n">tempfile</span><span class="o">.</span><span class="n">gettempdir</span><span class="p">()</span> <span class="o">+</span> <span class="s2">"/model.checkpoint"</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="c1"># ๋ชจ๋ ์์
์ ๊ฐ์ ๋งค๊ฐ๋ณ์๋ก๋ถํฐ ์์๋๋ค๊ณ ์๊ฐํด์ผ ํฉ๋๋ค.</span>
<span class="c1"># ๋ฌด์์์ ๋งค๊ฐ๋ณ์์ ๋ณํ๋๋ ์ญ์ ํ ์ ๋ฌ๋ก ๋๊ธฐํ๋ฉ๋๋ค.</span>
<span class="c1"># ๊ทธ๋ผ์ผ๋ก, ํ๋์ ์์
์ ๋ชจ๋ธ์ ์ ์ฅํ๊ธฐ์ ์ถฉ๋ถํฉ๋๋ค.</span>
<span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">ddp_model</span><span class="o">.</span><span class="n">state_dict</span><span class="p">(),</span> <span class="n">CHECKPOINT_PATH</span><span class="p">)</span>
<span class="c1"># ์์
0์ด ์ ์ฅํ ํ ์์
1์ด ๋ชจ๋ธ์ ์ฝ์ด์ค๋๋ก barrier()๋ฅผ ์ฌ์ฉํฉ๋๋ค.</span>
<span class="n">dist</span><span class="o">.</span><span class="n">barrier</span><span class="p">()</span>
<span class="c1"># configure map_location properly</span>
<span class="n">map_location</span> <span class="o">=</span> <span class="p">{</span><span class="s1">'cuda:</span><span class="si">%d</span><span class="s1">'</span> <span class="o">%</span> <span class="mi">0</span><span class="p">:</span> <span class="s1">'cuda:</span><span class="si">%d</span><span class="s1">'</span> <span class="o">%</span> <span class="n">rank</span><span class="p">}</span>
<span class="n">ddp_model</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span>
<span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="n">CHECKPOINT_PATH</span><span class="p">,</span> <span class="n">map_location</span><span class="o">=</span><span class="n">map_location</span><span class="p">))</span>
<span class="n">loss_fn</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">MSELoss</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">SGD</span><span class="p">(</span><span class="n">ddp_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="mf">0.001</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">outputs</span> <span class="o">=</span> <span class="n">ddp_model</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">20</span><span class="p">,</span> <span class="mi">10</span><span class="p">))</span>
<span class="n">labels</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">20</span><span class="p">,</span> <span class="mi">5</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">rank</span><span class="p">)</span>
<span class="n">loss_fn</span><span class="p">(</span><span class="n">outputs</span><span class="p">,</span> <span class="n">labels</span><span class="p">)</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="c1"># ํ์ผ์ญ์ ๋ฅผ ๋ณดํธํ๊ธฐ ์ํด ์๋์ dist.barrier()๋ฅผ ์ฌ์ฉํ ํ์๋ ์์ต๋๋ค.</span>
<span class="c1"># DDP์ ์ญ์ ํ ์ ๋ฌ ๊ณผ์ ์ ์๋ AllReduce ์ต์ค(ops)๊ฐ ๋๊ธฐํ ๊ธฐ๋ฅ์ ์ํํ๊ธฐ ๋๋ฌธ์</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">os</span><span class="o">.</span><span class="n">remove</span><span class="p">(</span><span class="n">CHECKPOINT_PATH</span><span class="p">)</span>
<span class="n">cleanup</span><span class="p">()</span>
</pre></div>
</div>
</div>
<div class="section" id="ddp">
<h2>๋ชจ๋ธ ๋ณ๋ ฌ ์ฒ๋ฆฌ๋ฅผ ํ์ฉํ DDP<a class="headerlink" href="#ddp" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>DDP๋ ๋ค์ค GPU ๋ชจ๋ธ์์๋ ์๋ํฉ๋๋ค.
๋ค์ค GPU ๋ชจ๋ธ์ ํ์ฉํ DDP๋ ๋์ฉ๋์ ๋ฐ์ดํฐ๋ฅผ ๊ฐ์ง ๋์ฉ๋ ๋ชจ๋ธ์ ํ์ต์ํฌ ๋ ํนํ ์ ์ฉํฉ๋๋ค.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">ToyMpModel</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">dev0</span><span class="p">,</span> <span class="n">dev1</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">ToyMpModel</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">dev0</span> <span class="o">=</span> <span class="n">dev0</span>
<span class="bp">self</span><span class="o">.</span><span class="n">dev1</span> <span class="o">=</span> <span class="n">dev1</span>
<span class="bp">self</span><span class="o">.</span><span class="n">net1</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">10</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dev0</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">relu</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">net2</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">5</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dev1</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">x</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">dev0</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">net1</span><span class="p">(</span><span class="n">x</span><span class="p">))</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">x</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">dev1</span><span class="p">)</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">net2</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
</pre></div>
</div>
<p>๋ค์ค GPU ๋ชจ๋ธ์ DDP๋ก ์ ๋ฌํ ๋, <code class="docutils literal notranslate"><span class="pre">device_ids</span></code>์ <code class="docutils literal notranslate"><span class="pre">output_device</span></code>๋ฅผ ์ค์ ํ์ง ์์์ผ ํฉ๋๋ค.
์
๋ ฅ ๋ฐ ์ถ๋ ฅ ๋ฐ์ดํฐ๋ ์ดํ๋ฆฌ์ผ์ด์
๋๋ ๋ชจ๋ธ <code class="docutils literal notranslate"><span class="pre">forward()</span></code>์ ์ํด ์ ์ ํ ์ฅ์น์ ๋ฐฐ์น๋ฉ๋๋ค.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">demo_model_parallel</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="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Running DDP with model parallel example on rank </span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s2">."</span><span class="p">)</span>
<span class="n">setup</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="c1"># ์์
์ ์ํ mp_model ๋ฐ ์ฅ์น ์ค์ </span>
<span class="n">dev0</span> <span class="o">=</span> <span class="n">rank</span> <span class="o">*</span> <span class="mi">2</span>
<span class="n">dev1</span> <span class="o">=</span> <span class="n">rank</span> <span class="o">*</span> <span class="mi">2</span> <span class="o">+</span> <span class="mi">1</span>
<span class="n">mp_model</span> <span class="o">=</span> <span class="n">ToyMpModel</span><span class="p">(</span><span class="n">dev0</span><span class="p">,</span> <span class="n">dev1</span><span class="p">)</span>
<span class="n">ddp_mp_model</span> <span class="o">=</span> <span class="n">DDP</span><span class="p">(</span><span class="n">mp_model</span><span class="p">)</span>
<span class="n">loss_fn</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">MSELoss</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">SGD</span><span class="p">(</span><span class="n">ddp_mp_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="mf">0.001</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="c1"># ์ถ๋ ฅ๊ฐ์ dev1์ ์ ์ฅ</span>
<span class="n">outputs</span> <span class="o">=</span> <span class="n">ddp_mp_model</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">20</span><span class="p">,</span> <span class="mi">10</span><span class="p">))</span>
<span class="n">labels</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">20</span><span class="p">,</span> <span class="mi">5</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dev1</span><span class="p">)</span>
<span class="n">loss_fn</span><span class="p">(</span><span class="n">outputs</span><span class="p">,</span> <span class="n">labels</span><span class="p">)</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">cleanup</span><span class="p">()</span>
<span class="k">if</span> <span class="vm">__name__</span> <span class="o">==</span> <span class="s2">"__main__"</span><span class="p">:</span>
<span class="n">n_gpus</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">device_count</span><span class="p">()</span>
<span class="k">assert</span> <span class="n">n_gpus</span> <span class="o">>=</span> <span class="mi">2</span><span class="p">,</span> <span class="sa">f</span><span class="s2">"Requires at least 2 GPUs to run, but got </span><span class="si">{</span><span class="n">n_gpus</span><span class="si">}</span><span class="s2">"</span>
<span class="n">world_size</span> <span class="o">=</span> <span class="n">n_gpus</span>
<span class="n">run_demo</span><span class="p">(</span><span class="n">demo_basic</span><span class="p">,</span> <span class="n">world_size</span><span class="p">)</span>
<span class="n">run_demo</span><span class="p">(</span><span class="n">demo_checkpoint</span><span class="p">,</span> <span class="n">world_size</span><span class="p">)</span>
<span class="n">run_demo</span><span class="p">(</span><span class="n">demo_model_parallel</span><span class="p">,</span> <span class="n">world_size</span><span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="initialize-ddp-with-torch-distributed-run-torchrun">
<h2>Initialize DDP with torch.distributed.run/torchrun<a class="headerlink" href="#initialize-ddp-with-torch-distributed-run-torchrun" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>We can leverage PyTorch Elastic to simplify the DDP code and initialize the job more easily.
Letโs still use the Toymodel example and create a file named <code class="docutils literal notranslate"><span class="pre">elastic_ddp.py</span></code>.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</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.nn</span> <span class="k">as</span> <span class="nn">nn</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">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="k">class</span> <span class="nc">ToyModel</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">ToyModel</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">net1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">10</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">relu</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">net2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">5</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">net2</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">net1</span><span class="p">(</span><span class="n">x</span><span class="p">)))</span>
<span class="k">def</span> <span class="nf">demo_basic</span><span class="p">():</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="n">rank</span> <span class="o">=</span> <span class="n">dist</span><span class="o">.</span><span class="n">get_rank</span><span class="p">()</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Start running basic DDP example on rank </span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s2">."</span><span class="p">)</span>
<span class="c1"># create model and move it to GPU with id rank</span>
<span class="n">device_id</span> <span class="o">=</span> <span class="n">rank</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">device_count</span><span class="p">()</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">ToyModel</span><span class="p">()</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device_id</span><span class="p">)</span>
<span class="n">ddp_model</span> <span class="o">=</span> <span class="n">DDP</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">device_ids</span><span class="o">=</span><span class="p">[</span><span class="n">device_id</span><span class="p">])</span>
<span class="n">loss_fn</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">MSELoss</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">SGD</span><span class="p">(</span><span class="n">ddp_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="mf">0.001</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">outputs</span> <span class="o">=</span> <span class="n">ddp_model</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">20</span><span class="p">,</span> <span class="mi">10</span><span class="p">))</span>
<span class="n">labels</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">20</span><span class="p">,</span> <span class="mi">5</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device_id</span><span class="p">)</span>
<span class="n">loss_fn</span><span class="p">(</span><span class="n">outputs</span><span class="p">,</span> <span class="n">labels</span><span class="p">)</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">dist</span><span class="o">.</span><span class="n">destroy_process_group</span><span class="p">()</span>
<span class="k">if</span> <span class="vm">__name__</span> <span class="o">==</span> <span class="s2">"__main__"</span><span class="p">:</span>
<span class="n">demo_basic</span><span class="p">()</span>
</pre></div>
</div>
<p>One can then run a <a class="reference external" href="https://pytorch.org/docs/stable/elastic/quickstart.html">torch elastic/torchrun</a> command
on all nodes to initialize the DDP job created above:</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span>torchrun<span class="w"> </span>--nnodes<span class="o">=</span><span class="m">2</span><span class="w"> </span>--nproc_per_node<span class="o">=</span><span class="m">8</span><span class="w"> </span>--rdzv_id<span class="o">=</span><span class="m">100</span><span class="w"> </span>--rdzv_backend<span class="o">=</span>c10d<span class="w"> </span>--rdzv_endpoint<span class="o">=</span><span class="nv">$MASTER_ADDR</span>:29400<span class="w"> </span>elastic_ddp.py
</pre></div>
</div>
<p>We are running the DDP script on two hosts, and each host we run with 8 processes, aka, we
are running it on 16 GPUs. Note that <code class="docutils literal notranslate"><span class="pre">$MASTER_ADDR</span></code> must be the same across all nodes.</p>
<p>Here torchrun will launch 8 process and invoke <code class="docutils literal notranslate"><span class="pre">elastic_ddp.py</span></code>
on each process on the node it is launched on, but user also needs to apply cluster
management tools like slurm to actually run this command on 2 nodes.</p>
<p>For example, on a SLURM enabled cluster, we can write a script to run the command above
and set <code class="docutils literal notranslate"><span class="pre">MASTER_ADDR</span></code> as:</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span><span class="nb">export</span><span class="w"> </span><span class="nv">MASTER_ADDR</span><span class="o">=</span><span class="k">$(</span>scontrol<span class="w"> </span>show<span class="w"> </span>hostname<span class="w"> </span><span class="si">${</span><span class="nv">SLURM_NODELIST</span><span class="si">}</span><span class="w"> </span><span class="p">|</span><span class="w"> </span>head<span class="w"> </span>-n<span class="w"> </span><span class="m">1</span><span class="k">)</span>
</pre></div>
</div>
<p>Then we can just run this script using the SLURM command: <code class="docutils literal notranslate"><span class="pre">srun</span> <span class="pre">--nodes=2</span> <span class="pre">./torchrun_script.sh</span></code>.
Of course, this is just an example; you can choose your own cluster scheduling tools
to initiate the torchrun job.</p>
<p>For more information about Elastic run, one can check this
<a class="reference external" href="https://pytorch.org/docs/stable/elastic/quickstart.html">quick start document</a> to learn more.</p>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="dist_tuto.html" class="btn btn-neutral float-right" title="PyTorch๋ก ๋ถ์ฐ ์ดํ๋ฆฌ์ผ์ด์
๊ฐ๋ฐํ๊ธฐ" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="model_parallel_tutorial.html" class="btn btn-neutral" title="๋จ์ผ ๋จธ์ ์ ์ฌ์ฉํ ๋ชจ๋ธ ๋ณ๋ ฌํ ๋ชจ๋ฒ ์ฌ๋ก" accesskey="p" rel="prev"><img src="../_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr class="community-hr hr-top" />
<div class="community-container">
<div class="community-prompt">๋ ๊ถ๊ธํ์๊ฑฐ๋ ๊ฐ์ ํ ๋ด์ฉ์ด ์์ผ์ ๊ฐ์? ์ปค๋ฎค๋ํฐ์ ์ฐธ์ฌํด๋ณด์ธ์!</div>
<div class="community-link"><a href="https://discuss.pytorch.kr/" aria-label="PyTorchKoreaCommunity">ํ๊ตญ์ด ์ปค๋ฎค๋ํฐ ๋ฐ๋ก๊ฐ๊ธฐ</a></div>
</div>
<hr class="community-hr hr-bottom"/>
<hr class="rating-hr hr-top" />
<div class="rating-container">
<div class="rating-prompt">์ด ํํ ๋ฆฌ์ผ์ด ์ด๋ ์
จ๋์? ํ๊ฐํด์ฃผ์๋ฉด ์ดํ ๊ฐ์ ์ ์ฐธ๊ณ ํ๊ฒ ์ต๋๋ค! :)</div>
<div class="stars-outer">
<i class="far fa-star" title="1 Star" data-behavior="tutorial-rating" data-count="1"></i>
<i class="far fa-star" title="2 Stars" data-behavior="tutorial-rating" data-count="2"></i>
<i class="far fa-star" title="3 Stars" data-behavior="tutorial-rating" data-count="3"></i>
<i class="far fa-star" title="4 Stars" data-behavior="tutorial-rating" data-count="4"></i>
<i class="far fa-star" title="5 Stars" data-behavior="tutorial-rating" data-count="5"></i>
</div>
</div>
<hr class="rating-hr hr-bottom"/>
<div role="contentinfo">
<p>
© Copyright 2018-2024, PyTorch & ํ์ดํ ์น ํ๊ตญ ์ฌ์ฉ์ ๋ชจ์(PyTorch Korea User Group).
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
<script>
if((window.location.href.indexOf("/prototype/")!= -1) && (window.location.href.indexOf("/prototype/prototype_index")< 1))
{
var div = '<div class="admonition note"><p class="admonition-title">Note</p><p><i class="fa fa-flask" aria-hidden="true"> </i> ์ด ํํ ๋ฆฌ์ผ์ ํ๋กํ ํ์
(prototype) ๊ธฐ๋ฅ๋ค์ ๋ํด์ ์ค๋ช
ํ๊ณ ์์ต๋๋ค. ํ๋กํ ํ์
๊ธฐ๋ฅ์ ์ผ๋ฐ์ ์ผ๋ก ํผ๋๋ฐฑ ๋ฐ ํ
์คํธ์ฉ์ผ๋ก, ๋ฐํ์ ํ๋๊ทธ ์์ด๋ PyPI๋ Conda๋ก ๋ฐฐํฌ๋๋ ๋ฐ์ด๋๋ฆฌ์์๋ ์ฌ์ฉํ ์ ์์ต๋๋ค.</p></div>'
document.getElementById("pytorch-article").insertAdjacentHTML('afterBegin', div)
}
</script>
</div>
<div class="pytorch-content-right" id="pytorch-content-right">
<div class="pytorch-right-menu" id="pytorch-right-menu">
<div class="pytorch-side-scroll" id="pytorch-side-scroll-right">
<ul>
<li><a class="reference internal" href="#">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ์์ํ๊ธฐ</a><ul>
<li><a class="reference internal" href="#id3">๊ธฐ๋ณธ์ ์ธ ์ฌ์ฉ๋ฒ</a></li>
<li><a class="reference internal" href="#id4">๋น๋์นญ ์์
์๋</a></li>
<li><a class="reference internal" href="#id5">์ฒดํฌํฌ์ธํธ๋ฅผ ์ ์ฅํ๊ณ ์ฝ์ด์ค๊ธฐ</a></li>
<li><a class="reference internal" href="#ddp">๋ชจ๋ธ ๋ณ๋ ฌ ์ฒ๋ฆฌ๋ฅผ ํ์ฉํ DDP</a></li>
<li><a class="reference internal" href="#initialize-ddp-with-torch-distributed-run-torchrun">Initialize DDP with torch.distributed.run/torchrun</a></li>
</ul>
</li>
</ul>
</div>
</div>
</div>
</section>
</div>
<script type="text/javascript" id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script data-url_root="../" id="documentation_options" src="../_static/documentation_options.js"></script>
<script src="../_static/jquery.js"></script>
<script src="../_static/underscore.js"></script>
<script src="../_static/_sphinx_javascript_frameworks_compat.js"></script>
<script src="../_static/doctools.js"></script>
<script src="../_static/clipboard.min.js"></script>
<script src="../_static/copybutton.js"></script>
<script src="../_static/translations.js"></script>
<script src="../_static/katex.min.js"></script>
<script src="../_static/auto-render.min.js"></script>
<script src="../_static/katex_autorenderer.js"></script>
<script src="../_static/design-tabs.js"></script>
<script type="text/javascript" src="../_static/js/vendor/popper.min.js"></script>
<script type="text/javascript" src="../_static/js/vendor/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/list.js/1.5.0/list.min.js"></script>
<script type="text/javascript" src="../_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<script>
// Helper function to make it easier to call dataLayer.push()
function gtag(){window.dataLayer.push(arguments);}
//add microsoft link
if(window.location.href.indexOf("/beginner/basics/")!= -1)
{
var url="https://docs.microsoft.com/learn/paths/pytorch-fundamentals/?wt.mc_id=aiml-7486-cxa";
switch(window.location.pathname.split("/").pop().replace('.html',''))
{
case"quickstart_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/9-quickstart?WT.mc_id=aiml-7486-cxa";
break;
case"tensorqs_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/2-tensors?WT.mc_id=aiml-7486-cxa";
break;
case"data_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/3-data?WT.mc_id=aiml-7486-cxa";
break;
case"transforms_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/4-transforms?WT.mc_id=aiml-7486-cxa";
break;
case"buildmodel_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/5-model?WT.mc_id=aiml-7486-cxa";
break;
case"autogradqs_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/6-autograd?WT.mc_id=aiml-7486-cxa";
break;
case"optimization_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/7-optimization?WT.mc_id=aiml-7486-cxa";
break;
case"saveloadrun_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/8-inference?WT.mc_id=aiml-7486-cxa";
}
$(".pytorch-call-to-action-links").children().first().before("<a href="+url+' data-behavior="call-to-action-event" data-response="Run in Microsoft Learn" target="_blank"><div id="microsoft-learn-link" style="padding-bottom: 0.625rem;border-bottom: 1px solid #f3f4f7;padding-right: 2.5rem;display: -webkit-box; display: -ms-flexbox; display: flex; -webkit-box-align: center;-ms-flex-align: center;align-items: center;"><img class="call-to-action-img" src="../../_static/images/microsoft-logo.svg"/><div class="call-to-action-desktop-view">Run in Microsoft Learn</div><div class="call-to-action-mobile-view">Learn</div></div></a>')
}
</script>
<script async src="https://www.googletagmanager.com/gtag/js?id=G-LZRD6GXDLF"></script>
<script data-cfasync="false">
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'G-LZRD6GXDLF'); // GA4
gtag('config', 'UA-71919972-3'); // UA
</script>
<script data-cfasync="false">
$("[data-behavior='call-to-action-event']").on('click', function(){
ga('send', {
hitType: 'event',
eventCategory: $(this).attr("data-response"),
eventAction: 'click',
eventLabel: window.location.href
});
gtag('event', 'click', {
'event_category': $(this).attr("data-response"),
'event_label': $("h1").first().text(),
'tutorial_link': window.location.href
});
});
$("[data-behavior='tutorial-rating']").on('click', function(){
gtag('event', 'click', {
'event_category': 'Tutorial Rating',
'event_label': $("h1").first().text(),
'value': $(this).attr("data-count"),
'customEvent:Rating': $(this).attr("data-count") // send to GA custom dimension customEvent:Rating.
});
});
if (location.pathname == "/") {
$(".rating-container").hide();
$(".hr-bottom").hide();
}
</script>
<script type="text/javascript">
var collapsedSections = ['ํ์ดํ ์น(PyTorch) ๋ ์ํผ', 'ํ์ดํ ์น(PyTorch) ๋ฐฐ์ฐ๊ธฐ', 'Introduction to PyTorch on YouTube', '์ด๋ฏธ์ง/๋น๋์ค', '์ค๋์ค', 'ํ
์คํธ', '๋ฐฑ์๋', '๊ฐํํ์ต', 'PyTorch ๋ชจ๋ธ์ ํ๋ก๋์
ํ๊ฒฝ์ ๋ฐฐํฌํ๊ธฐ', 'PyTorch ํ๋กํ์ผ๋ง', 'Code Transforms with FX', 'ํ๋ก ํธ์๋ API', 'PyTorch ํ์ฅํ๊ธฐ', '๋ชจ๋ธ ์ต์ ํ', '๋ณ๋ ฌ ๋ฐ ๋ถ์ฐ ํ์ต', 'Edge with ExecuTorch', '์ถ์ฒ ์์คํ
', 'Multimodality'];
</script>
<!-- Begin Footer -->
<div class="container-fluid docs-tutorials-resources" id="docs-tutorials-resources">
<div class="container">
<div class="row">
<div class="col-md-4 text-center">
<h2>PyTorchKorea @ GitHub</h2>
<p>ํ์ดํ ์น ํ๊ตญ ์ฌ์ฉ์ ๋ชจ์์ GitHub์์ ๋ง๋๋ณด์ธ์.</p>
<a class="with-right-arrow" href="https://github.com/PyTorchKorea" target="_blank">GitHub๋ก ์ด๋</a>
</div>
<div class="col-md-4 text-center">
<h2>ํ๊ตญ์ด ํํ ๋ฆฌ์ผ</h2>
<p>ํ๊ตญ์ด๋ก ๋ฒ์ญ ์ค์ธ PyTorch ํํ ๋ฆฌ์ผ์
๋๋ค.</p>
<a class="with-right-arrow" href="https://tutorials.pytorch.kr/">ํํ ๋ฆฌ์ผ๋ก ์ด๋</a>
</div>
<div class="col-md-4 text-center">
<h2>์ปค๋ฎค๋ํฐ</h2>
<p>๋ค๋ฅธ ์ฌ์ฉ์๋ค๊ณผ ์๊ฒฌ์ ๋๋๊ณ , ๋์์ฃผ์ธ์!</p>
<a class="with-right-arrow" href="https://discuss.pytorch.kr/">์ปค๋ฎค๋ํฐ๋ก ์ด๋</a>
</div>
</div>
</div>
</div>
<footer class="site-footer">
<div class="container footer-container">
<div class="footer-logo-wrapper">
<a href="https://pytorch.kr/" class="footer-logo"></a>
</div>
<div class="footer-links-wrapper">
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.kr/">ํ์ดํ ์น ํ๊ตญ ์ฌ์ฉ์ ๋ชจ์</a></li>