-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathrun.html
1246 lines (1060 loc) · 76.3 KB
/
run.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="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta name="robots" content="noindex">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>torchrun (Elastic Launch) — PyTorch 2.4 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/elastic/run.html"/>
<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/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="../_static/sphinx-dropdown.css" type="text/css" />
<link rel="stylesheet" href="../_static/panels-bootstrap.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/jit.css" type="text/css" />
<link rel="index" title="Index" href="../genindex.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="Elastic Agent" href="agent.html" />
<link rel="prev" title="Examples" href="examples.html" />
<!-- Google Tag Manager -->
<script>(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':
new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],
j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src=
'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);
})(window,document,'script','dataLayer','GTM-T8XT4PS');</script>
<!-- End Google Tag Manager -->
<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/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/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.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Learn
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/get-started">
<span class=dropdown-title>Get Started</span>
<p>Run PyTorch locally or get started quickly with one of the supported cloud platforms</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials">
<span class="dropdown-title">Tutorials</span>
<p>Whats new in PyTorch tutorials</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/basics/intro.html">
<span class="dropdown-title">Learn the Basics</span>
<p>Familiarize yourself with PyTorch concepts and modules</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/recipes/recipes_index.html">
<span class="dropdown-title">PyTorch Recipes</span>
<p>Bite-size, ready-to-deploy PyTorch code examples</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/introyt.html">
<span class="dropdown-title">Intro to PyTorch - YouTube Series</span>
<p>Master PyTorch basics with our engaging YouTube tutorial series</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Ecosystem
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem">
<span class="dropdown-title">Tools</span>
<p>Learn about the tools and frameworks in the PyTorch Ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class=dropdown-title>Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class=dropdown-title>Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class=dropdown-title>Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem/contributor-awards-2023">
<span class="dropdown-title">Contributor Awards - 2023</span>
<p>Award winners announced at this year's PyTorch Conference</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Edge
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/edge">
<span class="dropdown-title">About PyTorch Edge</span>
<p>Build innovative and privacy-aware AI experiences for edge devices</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch-overview">
<span class="dropdown-title">ExecuTorch</span>
<p>End-to-end solution for enabling on-device inference capabilities across mobile and edge devices</p>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p>Explore the documentation for comprehensive guidance on how to use PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/pytorch-domains">
<span class="dropdown-title">PyTorch Domains</span>
<p>Read the PyTorch Domains documentation to learn more about domain-specific libraries</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Blogs & News
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/blog/">
<span class="dropdown-title">PyTorch Blog</span>
<p>Catch up on the latest technical news and happenings</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/community-blog">
<span class="dropdown-title">Community Blog</span>
<p>Stories from the PyTorch ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/videos">
<span class="dropdown-title">Videos</span>
<p>Learn about the latest PyTorch tutorials, new, and more </p>
<a class="nav-dropdown-item" href="https://pytorch.org/community-stories">
<span class="dropdown-title">Community Stories</span>
<p>Learn how our community solves real, everyday machine learning problems with PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
About
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn more about the PyTorch Foundation</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/governing-board">
<span class="dropdown-title">Governing Board</span>
<p></p>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div class="no-dropdown">
<a href="https://pytorch.org/join" data-cta="join">
Become a Member
</a>
</div>
</li>
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="github-icon">
</a>
</div>
</li>
<!--- TODO: This block adds the search icon to the nav bar. We will enable it later.
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="search-icon">
</a>
</div>
</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">
<a href='https://pytorch.org/docs/versions.html'>2.4 ▼</a>
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="../search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<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">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../community/build_ci_governance.html">PyTorch Governance | Build + CI</a></li>
<li class="toctree-l1"><a class="reference internal" href="../community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="../community/design.html">PyTorch Design Philosophy</a></li>
<li class="toctree-l1"><a class="reference internal" href="../community/governance.html">PyTorch Governance | Mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="../community/persons_of_interest.html">PyTorch Governance | Maintainers</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Developer Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../notes/amp_examples.html">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/custom_operators.html">PyTorch Custom Operators Landing Page</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/extending.func.html">Extending torch.func with autograd.Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/fsdp.html">FSDP Notes</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/get_start_xpu.html">Pytorch 2.4: Getting Started on Intel GPU</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/mps.html">MPS backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/numerical_accuracy.html">Numerical accuracy</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="../notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
<li class="toctree-l1"><a class="reference internal" href="../deploy.html">torch::deploy</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python API</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="../amp.html">torch.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="../autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="../library.html">torch.library</a></li>
<li class="toctree-l1"><a class="reference internal" href="../cpu.html">torch.cpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="../cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_cuda_memory.html">Understanding CUDA Memory Usage</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_cuda_memory.html#generating-a-snapshot">Generating a Snapshot</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_cuda_memory.html#using-the-visualizer">Using the visualizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_cuda_memory.html#snapshot-api-reference">Snapshot API Reference</a></li>
<li class="toctree-l1"><a class="reference internal" href="../mps.html">torch.mps</a></li>
<li class="toctree-l1"><a class="reference internal" href="../xpu.html">torch.xpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="../mtia.html">torch.mtia</a></li>
<li class="toctree-l1"><a class="reference internal" href="../meta.html">Meta device</a></li>
<li class="toctree-l1"><a class="reference internal" href="../backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="../export.html">torch.export</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1 current"><a class="reference internal" href="../distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fsdp.html">torch.distributed.fsdp</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.pipelining.html">torch.distributed.pipelining</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.tensor.parallel.html">torch.distributed.tensor.parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.checkpoint.html">torch.distributed.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch.compiler.html">torch.compiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="../func.html">torch.func</a></li>
<li class="toctree-l1"><a class="reference internal" href="../futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fx.experimental.html">torch.fx.experimental</a></li>
<li class="toctree-l1"><a class="reference internal" href="../hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="../jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="../linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="../monitor.html">torch.monitor</a></li>
<li class="toctree-l1"><a class="reference internal" href="../signal.html">torch.signal</a></li>
<li class="toctree-l1"><a class="reference internal" href="../special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="../package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="../profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.attention.html">torch.nn.attention</a></li>
<li class="toctree-l1"><a class="reference internal" href="../onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="../optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="../complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="../ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="../quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="../rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="../random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="../masked.html">torch.masked</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nested.html">torch.nested</a></li>
<li class="toctree-l1"><a class="reference internal" href="../size.html">torch.Size</a></li>
<li class="toctree-l1"><a class="reference internal" href="../sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="../storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="../testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="../utils.html">torch.utils</a></li>
<li class="toctree-l1"><a class="reference internal" href="../benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="../bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="../checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="../cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="../data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="../deterministic.html">torch.utils.deterministic</a></li>
<li class="toctree-l1"><a class="reference internal" href="../jit_utils.html">torch.utils.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="../dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="../mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="../model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="../module_tracker.html">torch.utils.module_tracker</a></li>
<li class="toctree-l1"><a class="reference internal" href="../type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="../named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="../name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="../config_mod.html">torch.__config__</a></li>
<li class="toctree-l1"><a class="reference internal" href="../future_mod.html">torch.__future__</a></li>
<li class="toctree-l1"><a class="reference internal" href="../logging.html">torch._logging</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_environment_variables.html">Torch Environment Variables</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/data">TorchData</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/torchrec">TorchRec</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/xla/">PyTorch on XLA Devices</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">
Docs
</a> >
</li>
<li><a href="../distributed.elastic.html">Torch Distributed Elastic</a> ></li>
<li>torchrun (Elastic Launch)</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/elastic/run.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">
<!-- Google Tag Manager (noscript) -->
<noscript><iframe src="https://www.googletagmanager.com/ns.html?id=GTM-T8XT4PS"
height="0" width="0" style="display:none;visibility:hidden"></iframe></noscript>
<!-- End Google Tag Manager (noscript) -->
<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="module-torch.distributed.run">
<span id="torchrun-elastic-launch"></span><span id="launcher-api"></span><h1>torchrun (Elastic Launch)<a class="headerlink" href="#module-torch.distributed.run" title="Permalink to this heading">¶</a></h1>
<p>Superset of <code class="docutils literal notranslate"><span class="pre">torch.distributed.launch</span></code>.</p>
<p><code class="docutils literal notranslate"><span class="pre">torchrun</span></code> provides a superset of the functionality as <code class="docutils literal notranslate"><span class="pre">torch.distributed.launch</span></code>
with the following additional functionalities:</p>
<ol class="arabic simple">
<li><p>Worker failures are handled gracefully by restarting all workers.</p></li>
<li><p>Worker <code class="docutils literal notranslate"><span class="pre">RANK</span></code> and <code class="docutils literal notranslate"><span class="pre">WORLD_SIZE</span></code> are assigned automatically.</p></li>
<li><p>Number of nodes is allowed to change between minimum and maximum sizes (elasticity).</p></li>
</ol>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p><code class="docutils literal notranslate"><span class="pre">torchrun</span></code> is a python
<a class="reference external" href="https://packaging.python.org/en/latest/specifications/entry-points/#use-for-scripts">console script</a>
to the main module
<a class="reference external" href="https://github.com/pytorch/pytorch/blob/master/torch/distributed/run.py">torch.distributed.run</a>
declared in the <code class="docutils literal notranslate"><span class="pre">entry_points</span></code> configuration in
<a class="reference external" href="https://github.com/pytorch/pytorch/blob/master/setup.py">setup.py</a>.
It is equivalent to invoking <code class="docutils literal notranslate"><span class="pre">python</span> <span class="pre">-m</span> <span class="pre">torch.distributed.run</span></code>.</p>
</div>
<div class="section" id="transitioning-from-torch-distributed-launch-to-torchrun">
<h2>Transitioning from torch.distributed.launch to torchrun<a class="headerlink" href="#transitioning-from-torch-distributed-launch-to-torchrun" title="Permalink to this heading">¶</a></h2>
<p><code class="docutils literal notranslate"><span class="pre">torchrun</span></code> supports the same arguments as <code class="docutils literal notranslate"><span class="pre">torch.distributed.launch</span></code> <strong>except</strong>
for <code class="docutils literal notranslate"><span class="pre">--use-env</span></code> which is now deprecated. To migrate from <code class="docutils literal notranslate"><span class="pre">torch.distributed.launch</span></code>
to <code class="docutils literal notranslate"><span class="pre">torchrun</span></code> follow these steps:</p>
<ol class="arabic">
<li><p>If your training script is already reading <code class="docutils literal notranslate"><span class="pre">local_rank</span></code> from the <code class="docutils literal notranslate"><span class="pre">LOCAL_RANK</span></code> environment variable.
Then you need simply omit the <code class="docutils literal notranslate"><span class="pre">--use-env</span></code> flag, e.g.:</p>
<table class="docutils colwidths-auto align-default">
<thead>
<tr class="row-odd"><th class="head"><p><code class="docutils literal notranslate"><span class="pre">torch.distributed.launch</span></code></p></th>
<th class="head"><p><code class="docutils literal notranslate"><span class="pre">torchrun</span></code></p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><div class="highlight-shell-session notranslate"><div class="highlight"><pre><span></span><span class="gp">$ </span>python<span class="w"> </span>-m<span class="w"> </span>torch.distributed.launch<span class="w"> </span>--use-env<span class="w"> </span>train_script.py
</pre></div>
</div>
</td>
<td><div class="highlight-shell-session notranslate"><div class="highlight"><pre><span></span><span class="gp">$ </span>torchrun<span class="w"> </span>train_script.py
</pre></div>
</div>
</td>
</tr>
</tbody>
</table>
</li>
<li><p>If your training script reads local rank from a <code class="docutils literal notranslate"><span class="pre">--local-rank</span></code> cmd argument.
Change your training script to read from the <code class="docutils literal notranslate"><span class="pre">LOCAL_RANK</span></code> environment variable as
demonstrated by the following code snippet:</p>
<table class="docutils colwidths-auto align-default">
<thead>
<tr class="row-odd"><th class="head"><p><code class="docutils literal notranslate"><span class="pre">torch.distributed.launch</span></code></p></th>
<th class="head"><p><code class="docutils literal notranslate"><span class="pre">torchrun</span></code></p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">argparse</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">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s2">"--local-rank"</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">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">local_rank</span> <span class="o">=</span> <span class="n">args</span><span class="o">.</span><span class="n">local_rank</span>
</pre></div>
</div>
</td>
<td><div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">os</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="s2">"LOCAL_RANK"</span><span class="p">])</span>
</pre></div>
</div>
</td>
</tr>
</tbody>
</table>
</li>
</ol>
<div class="versionchanged">
<p><span class="versionmodified changed">Changed in version 2.0.0: </span>The launcher will pass the <code class="docutils literal notranslate"><span class="pre">--local-rank=<rank></span></code> argument to your script.
From PyTorch 2.0.0 onwards, the dashed <code class="docutils literal notranslate"><span class="pre">--local-rank</span></code> is preferred over the
previously used underscored <code class="docutils literal notranslate"><span class="pre">--local_rank</span></code>.</p>
<p>For backward compatibility, it may be necessary for users to handle both
cases in their argument parsing code. This means including both <code class="docutils literal notranslate"><span class="pre">"--local-rank"</span></code>
and <code class="docutils literal notranslate"><span class="pre">"--local_rank"</span></code> in the argument parser. If only <code class="docutils literal notranslate"><span class="pre">"--local_rank"</span></code> is
provided, the launcher will trigger an error: “error: unrecognized arguments:
–local-rank=<rank>”. For training code that only supports PyTorch 2.0.0+,
including <code class="docutils literal notranslate"><span class="pre">"--local-rank"</span></code> should be sufficient.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">argparse</span>
<span class="gp">>>> </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="gp">>>> </span><span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s2">"--local-rank"</span><span class="p">,</span> <span class="s2">"--local_rank"</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="gp">>>> </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>
</pre></div>
</div>
</div>
<p>The aformentioned changes suffice to migrate from <code class="docutils literal notranslate"><span class="pre">torch.distributed.launch</span></code> to <code class="docutils literal notranslate"><span class="pre">torchrun</span></code>.
To take advantage of new features such as elasticity, fault-tolerance, and error reporting of <code class="docutils literal notranslate"><span class="pre">torchrun</span></code>
please refer to:</p>
<ul class="simple">
<li><p><a class="reference internal" href="train_script.html#elastic-train-script"><span class="std std-ref">Train script</span></a> for more information on authoring training scripts that are <code class="docutils literal notranslate"><span class="pre">torchrun</span></code> compliant.</p></li>
<li><p>the rest of this page for more information on the features of <code class="docutils literal notranslate"><span class="pre">torchrun</span></code>.</p></li>
</ul>
<div class="section" id="usage">
<h3>Usage<a class="headerlink" href="#usage" title="Permalink to this heading">¶</a></h3>
<div class="section" id="single-node-multi-worker">
<h4>Single-node multi-worker<a class="headerlink" href="#single-node-multi-worker" title="Permalink to this heading">¶</a></h4>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span>torchrun
--standalone
--nnodes=1
--nproc-per-node=$NUM_TRAINERS
YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)
</pre></div>
</div>
</div>
<div class="section" id="stacked-single-node-multi-worker">
<h4>Stacked single-node multi-worker<a class="headerlink" href="#stacked-single-node-multi-worker" title="Permalink to this heading">¶</a></h4>
<p>To run multiple instances (separate jobs) of single-node, multi-worker on the
same host, we need to make sure that each instance (job) is
setup on different ports to avoid port conflicts (or worse, two jobs being merged
as a single job). To do this you have to run with <code class="docutils literal notranslate"><span class="pre">--rdzv-backend=c10d</span></code>
and specify a different port by setting <code class="docutils literal notranslate"><span class="pre">--rdzv-endpoint=localhost:$PORT_k</span></code>.
For <code class="docutils literal notranslate"><span class="pre">--nodes=1</span></code>, its often convenient to let <code class="docutils literal notranslate"><span class="pre">torchrun</span></code> pick a free random
port automatically instead of manually assigning different ports for each run.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span>torchrun
--rdzv-backend=c10d
--rdzv-endpoint=localhost:0
--nnodes=1
--nproc-per-node=$NUM_TRAINERS
YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)
</pre></div>
</div>
</div>
<div class="section" id="fault-tolerant-fixed-sized-number-of-workers-no-elasticity-tolerates-3-failures">
<h4>Fault tolerant (fixed sized number of workers, no elasticity, tolerates 3 failures)<a class="headerlink" href="#fault-tolerant-fixed-sized-number-of-workers-no-elasticity-tolerates-3-failures" title="Permalink to this heading">¶</a></h4>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span>torchrun
--nnodes=$NUM_NODES
--nproc-per-node=$NUM_TRAINERS
--max-restarts=3
--rdzv-id=$JOB_ID
--rdzv-backend=c10d
--rdzv-endpoint=$HOST_NODE_ADDR
YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)
</pre></div>
</div>
<p><code class="docutils literal notranslate"><span class="pre">HOST_NODE_ADDR</span></code>, in form <host>[:<port>] (e.g. node1.example.com:29400), specifies the node and
the port on which the C10d rendezvous backend should be instantiated and hosted. It can be any
node in your training cluster, but ideally you should pick a node that has a high bandwidth.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>If no port number is specified <code class="docutils literal notranslate"><span class="pre">HOST_NODE_ADDR</span></code> defaults to 29400.</p>
</div>
</div>
<div class="section" id="elastic-min-1-max-4-tolerates-up-to-3-membership-changes-or-failures">
<h4>Elastic (<code class="docutils literal notranslate"><span class="pre">min=1</span></code>, <code class="docutils literal notranslate"><span class="pre">max=4</span></code>, tolerates up to 3 membership changes or failures)<a class="headerlink" href="#elastic-min-1-max-4-tolerates-up-to-3-membership-changes-or-failures" title="Permalink to this heading">¶</a></h4>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span>torchrun
--nnodes=1:4
--nproc-per-node=$NUM_TRAINERS
--max-restarts=3
--rdzv-id=$JOB_ID
--rdzv-backend=c10d
--rdzv-endpoint=$HOST_NODE_ADDR
YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)
</pre></div>
</div>
<p><code class="docutils literal notranslate"><span class="pre">HOST_NODE_ADDR</span></code>, in form <host>[:<port>] (e.g. node1.example.com:29400), specifies the node and
the port on which the C10d rendezvous backend should be instantiated and hosted. It can be any
node in your training cluster, but ideally you should pick a node that has a high bandwidth.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>If no port number is specified <code class="docutils literal notranslate"><span class="pre">HOST_NODE_ADDR</span></code> defaults to 29400.</p>
</div>
</div>
</div>
<div class="section" id="note-on-rendezvous-backend">
<h3>Note on rendezvous backend<a class="headerlink" href="#note-on-rendezvous-backend" title="Permalink to this heading">¶</a></h3>
<p>For multi-node training you need to specify:</p>
<ol class="arabic simple">
<li><p><code class="docutils literal notranslate"><span class="pre">--rdzv-id</span></code>: A unique job id (shared by all nodes participating in the job)</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">--rdzv-backend</span></code>: An implementation of
<a class="reference internal" href="rendezvous.html#torch.distributed.elastic.rendezvous.RendezvousHandler" title="torch.distributed.elastic.rendezvous.RendezvousHandler"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.distributed.elastic.rendezvous.RendezvousHandler</span></code></a></p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">--rdzv-endpoint</span></code>: The endpoint where the rendezvous backend is running; usually in form
<code class="docutils literal notranslate"><span class="pre">host:port</span></code>.</p></li>
</ol>
<p>Currently <code class="docutils literal notranslate"><span class="pre">c10d</span></code> (recommended), <code class="docutils literal notranslate"><span class="pre">etcd-v2</span></code>, and <code class="docutils literal notranslate"><span class="pre">etcd</span></code> (legacy) rendezvous backends are
supported out of the box. To use <code class="docutils literal notranslate"><span class="pre">etcd-v2</span></code> or <code class="docutils literal notranslate"><span class="pre">etcd</span></code>, setup an etcd server with the <code class="docutils literal notranslate"><span class="pre">v2</span></code> api
enabled (e.g. <code class="docutils literal notranslate"><span class="pre">--enable-v2</span></code>).</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p><code class="docutils literal notranslate"><span class="pre">etcd-v2</span></code> and <code class="docutils literal notranslate"><span class="pre">etcd</span></code> rendezvous use etcd API v2. You MUST enable the v2 API on the etcd
server. Our tests use etcd v3.4.3.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>For etcd-based rendezvous we recommend using <code class="docutils literal notranslate"><span class="pre">etcd-v2</span></code> over <code class="docutils literal notranslate"><span class="pre">etcd</span></code> which is functionally
equivalent, but uses a revised implementation. <code class="docutils literal notranslate"><span class="pre">etcd</span></code> is in maintenance mode and will be
removed in a future version.</p>
</div>
</div>
<div class="section" id="definitions">
<h3>Definitions<a class="headerlink" href="#definitions" title="Permalink to this heading">¶</a></h3>
<ol class="arabic simple">
<li><p><code class="docutils literal notranslate"><span class="pre">Node</span></code> - A physical instance or a container; maps to the unit that the job manager works with.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">Worker</span></code> - A worker in the context of distributed training.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">WorkerGroup</span></code> - The set of workers that execute the same function (e.g. trainers).</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">LocalWorkerGroup</span></code> - A subset of the workers in the worker group running on the same node.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">RANK</span></code> - The rank of the worker within a worker group.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">WORLD_SIZE</span></code> - The total number of workers in a worker group.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">LOCAL_RANK</span></code> - The rank of the worker within a local worker group.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">LOCAL_WORLD_SIZE</span></code> - The size of the local worker group.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">rdzv_id</span></code> - A user-defined id that uniquely identifies the worker group for a job. This id is
used by each node to join as a member of a particular worker group.</p></li>
</ol>
<ol class="arabic simple" start="9">
<li><p><code class="docutils literal notranslate"><span class="pre">rdzv_backend</span></code> - The backend of the rendezvous (e.g. <code class="docutils literal notranslate"><span class="pre">c10d</span></code>). This is typically a strongly
consistent key-value store.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">rdzv_endpoint</span></code> - The rendezvous backend endpoint; usually in form <code class="docutils literal notranslate"><span class="pre"><host>:<port></span></code>.</p></li>
</ol>
<p>A <code class="docutils literal notranslate"><span class="pre">Node</span></code> runs <code class="docutils literal notranslate"><span class="pre">LOCAL_WORLD_SIZE</span></code> workers which comprise a <code class="docutils literal notranslate"><span class="pre">LocalWorkerGroup</span></code>. The union of
all <code class="docutils literal notranslate"><span class="pre">LocalWorkerGroups</span></code> in the nodes in the job comprise the <code class="docutils literal notranslate"><span class="pre">WorkerGroup</span></code>.</p>
</div>
<div class="section" id="environment-variables">
<h3>Environment Variables<a class="headerlink" href="#environment-variables" title="Permalink to this heading">¶</a></h3>
<p>The following environment variables are made available to you in your script:</p>
<ol class="arabic simple">
<li><p><code class="docutils literal notranslate"><span class="pre">LOCAL_RANK</span></code> - The local rank.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">RANK</span></code> - The global rank.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">GROUP_RANK</span></code> - The rank of the worker group. A number between 0 and <code class="docutils literal notranslate"><span class="pre">max_nnodes</span></code>. When
running a single worker group per node, this is the rank of the node.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">ROLE_RANK</span></code> - The rank of the worker across all the workers that have the same role. The role
of the worker is specified in the <code class="docutils literal notranslate"><span class="pre">WorkerSpec</span></code>.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">LOCAL_WORLD_SIZE</span></code> - The local world size (e.g. number of workers running locally); equals to
<code class="docutils literal notranslate"><span class="pre">--nproc-per-node</span></code> specified on <code class="docutils literal notranslate"><span class="pre">torchrun</span></code>.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">WORLD_SIZE</span></code> - The world size (total number of workers in the job).</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">ROLE_WORLD_SIZE</span></code> - The total number of workers that was launched with the same role specified
in <code class="docutils literal notranslate"><span class="pre">WorkerSpec</span></code>.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">MASTER_ADDR</span></code> - The FQDN of the host that is running worker with rank 0; used to initialize
the Torch Distributed backend.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">MASTER_PORT</span></code> - The port on the <code class="docutils literal notranslate"><span class="pre">MASTER_ADDR</span></code> that can be used to host the C10d TCP store.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">TORCHELASTIC_RESTART_COUNT</span></code> - The number of worker group restarts so far.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">TORCHELASTIC_MAX_RESTARTS</span></code> - The configured maximum number of restarts.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">TORCHELASTIC_RUN_ID</span></code> - Equal to the rendezvous <code class="docutils literal notranslate"><span class="pre">run_id</span></code> (e.g. unique job id).</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">PYTHON_EXEC</span></code> - System executable override. If provided, the python user script will
use the value of <code class="docutils literal notranslate"><span class="pre">PYTHON_EXEC</span></code> as executable. The <cite>sys.executable</cite> is used by default.</p></li>
</ol>
</div>
<div class="section" id="deployment">
<h3>Deployment<a class="headerlink" href="#deployment" title="Permalink to this heading">¶</a></h3>
<ol class="arabic simple">
<li><p>(Not needed for the C10d backend) Start the rendezvous backend server and get the endpoint (to be
passed as <code class="docutils literal notranslate"><span class="pre">--rdzv-endpoint</span></code> to the launcher script)</p></li>
<li><p>Single-node multi-worker: Start the launcher on the host to start the agent process which
creates and monitors a local worker group.</p></li>
<li><p>Multi-node multi-worker: Start the launcher with the same arguments on all the nodes
participating in training.</p></li>
</ol>
<p>When using a job/cluster manager the entry point command to the multi-node job should be this
launcher.</p>
</div>
<div class="section" id="failure-modes">
<h3>Failure Modes<a class="headerlink" href="#failure-modes" title="Permalink to this heading">¶</a></h3>
<ol class="arabic simple">
<li><p>Worker failure: For a training job with <code class="docutils literal notranslate"><span class="pre">n</span></code> workers, if <code class="docutils literal notranslate"><span class="pre">k<=n</span></code> workers fail all workers
are stopped and restarted up to <code class="docutils literal notranslate"><span class="pre">max_restarts</span></code>.</p></li>
<li><p>Agent failure: An agent failure results in a local worker group failure. It is up to the job
manager to fail the entire job (gang semantics) or attempt to replace the node. Both behaviors
are supported by the agent.</p></li>
<li><p>Node failure: Same as agent failure.</p></li>
</ol>
</div>
<div class="section" id="membership-changes">
<h3>Membership Changes<a class="headerlink" href="#membership-changes" title="Permalink to this heading">¶</a></h3>
<ol class="arabic simple">
<li><p>Node departure (scale-down): The agent is notified of the departure, all existing workers are
stopped, a new <code class="docutils literal notranslate"><span class="pre">WorkerGroup</span></code> is formed, and all workers are started with a new <code class="docutils literal notranslate"><span class="pre">RANK</span></code> and
<code class="docutils literal notranslate"><span class="pre">WORLD_SIZE</span></code>.</p></li>
<li><p>Node arrival (scale-up): The new node is admitted to the job, all existing workers are stopped,
a new <code class="docutils literal notranslate"><span class="pre">WorkerGroup</span></code> is formed, and all workers are started with a new <code class="docutils literal notranslate"><span class="pre">RANK</span></code> and
<code class="docutils literal notranslate"><span class="pre">WORLD_SIZE</span></code>.</p></li>
</ol>
</div>
<div class="section" id="important-notices">
<h3>Important Notices<a class="headerlink" href="#important-notices" title="Permalink to this heading">¶</a></h3>
<ol class="arabic simple">
<li><p>This utility and multi-process distributed (single-node or
multi-node) GPU training currently only achieves the best performance using
the NCCL distributed backend. Thus NCCL backend is the recommended backend to
use for GPU training.</p></li>
<li><p>The environment variables necessary to initialize a Torch process group are provided to you by
this module, no need for you to pass <code class="docutils literal notranslate"><span class="pre">RANK</span></code> manually. To initialize a process group in your
training script, simply run:</p></li>
</ol>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch.distributed</span> <span class="k">as</span> <span class="nn">dist</span>
<span class="gp">>>> </span><span class="n">dist</span><span class="o">.</span><span class="n">init_process_group</span><span class="p">(</span><span class="n">backend</span><span class="o">=</span><span class="s2">"gloo|nccl"</span><span class="p">)</span>
</pre></div>
</div>
<ol class="arabic simple" start="3">
<li><p>In your training program, you can either use regular distributed functions
or use <a class="reference internal" href="../generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel" title="torch.nn.parallel.DistributedDataParallel"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.nn.parallel.DistributedDataParallel()</span></code></a> module. If your
training program uses GPUs for training and you would like to use
<a class="reference internal" href="../generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel" title="torch.nn.parallel.DistributedDataParallel"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.nn.parallel.DistributedDataParallel()</span></code></a> module,
here is how to configure it.</p></li>
</ol>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></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="s2">"LOCAL_RANK"</span><span class="p">])</span>
<span class="n">model</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">parallel</span><span class="o">.</span><span class="n">DistributedDataParallel</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">local_rank</span><span class="p">],</span>
<span class="n">output_device</span><span class="o">=</span><span class="n">local_rank</span><span class="p">)</span>
</pre></div>
</div>
<p>Please ensure that <code class="docutils literal notranslate"><span class="pre">device_ids</span></code> argument is set to be the only GPU device id
that your code will be operating on. This is generally the local rank of the
process. In other words, the <code class="docutils literal notranslate"><span class="pre">device_ids</span></code> needs to be <code class="docutils literal notranslate"><span class="pre">[int(os.environ("LOCAL_RANK"))]</span></code>,
and <code class="docutils literal notranslate"><span class="pre">output_device</span></code> needs to be <code class="docutils literal notranslate"><span class="pre">int(os.environ("LOCAL_RANK"))</span></code> in order to use this
utility</p>
<ol class="arabic simple" start="4">
<li><p>On failures or membership changes ALL surviving workers are killed immediately. Make sure to
checkpoint your progress. The frequency of checkpoints should depend on your job’s tolerance
for lost work.</p></li>
<li><p>This module only supports homogeneous <code class="docutils literal notranslate"><span class="pre">LOCAL_WORLD_SIZE</span></code>. That is, it is assumed that all
nodes run the same number of local workers (per role).</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">RANK</span></code> is NOT stable. Between restarts, the local workers on a node can be assigned a
different range of ranks than before. NEVER hard code any assumptions about the stable-ness of
ranks or some correlation between <code class="docutils literal notranslate"><span class="pre">RANK</span></code> and <code class="docutils literal notranslate"><span class="pre">LOCAL_RANK</span></code>.</p></li>
<li><p>When using elasticity (<code class="docutils literal notranslate"><span class="pre">min_size!=max_size</span></code>) DO NOT hard code assumptions about
<code class="docutils literal notranslate"><span class="pre">WORLD_SIZE</span></code> as the world size can change as nodes are allowed to leave and join.</p></li>
<li><p>It is recommended for your script to have the following structure:</p></li>
</ol>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">main</span><span class="p">():</span>
<span class="n">load_checkpoint</span><span class="p">(</span><span class="n">checkpoint_path</span><span class="p">)</span>
<span class="n">initialize</span><span class="p">()</span>
<span class="n">train</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">train</span><span class="p">():</span>
<span class="k">for</span> <span class="n">batch</span> <span class="ow">in</span> <span class="nb">iter</span><span class="p">(</span><span class="n">dataset</span><span class="p">):</span>
<span class="n">train_step</span><span class="p">(</span><span class="n">batch</span><span class="p">)</span>
<span class="k">if</span> <span class="n">should_checkpoint</span><span class="p">:</span>
<span class="n">save_checkpoint</span><span class="p">(</span><span class="n">checkpoint_path</span><span class="p">)</span>
</pre></div>
</div>
<ol class="arabic simple" start="9">
<li><p>(Recommended) On worker errors, this tool will summarize the details of the error
(e.g. time, rank, host, pid, traceback, etc). On each node, the first error (by timestamp)
is heuristically reported as the “Root Cause” error. To get tracebacks as part of this
error summary print out, you must decorate your main entrypoint function in your
training script as shown in the example below. If not decorated, then the summary
will not include the traceback of the exception and will only contain the exitcode.
For details on torchelastic error handling see: <a class="reference external" href="https://pytorch.org/docs/stable/elastic/errors.html">https://pytorch.org/docs/stable/elastic/errors.html</a></p></li>
</ol>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">torch.distributed.elastic.multiprocessing.errors</span> <span class="kn">import</span> <span class="n">record</span>
<span class="nd">@record</span>
<span class="k">def</span> <span class="nf">main</span><span class="p">():</span>
<span class="c1"># do train</span>
<span class="k">pass</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">main</span><span class="p">()</span>
</pre></div>
</div>
</div>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="agent.html" class="btn btn-neutral float-right" title="Elastic Agent" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="examples.html" class="btn btn-neutral" title="Examples" accesskey="p" rel="prev"><img src="../_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright 2023, PyTorch Contributors.
</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>
var match = window.location.href.match(/\/_[a-zA-Z0-9_]*.html|_dynamo/gi);
var url = window.location.href.lastIndexOf(match[match.length-1]);
if (url)
{
var div = '<div class="admonition note"><p class="admonition-title">Note</p><p><i class="fa fa-exclamation-circle" aria-hidden="true"> </i> This page describes an internal API which is not intended to be used outside of the PyTorch codebase and can be modified or removed without notice.</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="#">torchrun (Elastic Launch)</a><ul>
<li><a class="reference internal" href="#transitioning-from-torch-distributed-launch-to-torchrun">Transitioning from torch.distributed.launch to torchrun</a><ul>
<li><a class="reference internal" href="#usage">Usage</a><ul>
<li><a class="reference internal" href="#single-node-multi-worker">Single-node multi-worker</a></li>
<li><a class="reference internal" href="#stacked-single-node-multi-worker">Stacked single-node multi-worker</a></li>
<li><a class="reference internal" href="#fault-tolerant-fixed-sized-number-of-workers-no-elasticity-tolerates-3-failures">Fault tolerant (fixed sized number of workers, no elasticity, tolerates 3 failures)</a></li>
<li><a class="reference internal" href="#elastic-min-1-max-4-tolerates-up-to-3-membership-changes-or-failures">Elastic (<code class="docutils literal notranslate"><span class="pre">min=1</span></code>, <code class="docutils literal notranslate"><span class="pre">max=4</span></code>, tolerates up to 3 membership changes or failures)</a></li>
</ul>
</li>
<li><a class="reference internal" href="#note-on-rendezvous-backend">Note on rendezvous backend</a></li>
<li><a class="reference internal" href="#definitions">Definitions</a></li>
<li><a class="reference internal" href="#environment-variables">Environment Variables</a></li>
<li><a class="reference internal" href="#deployment">Deployment</a></li>
<li><a class="reference internal" href="#failure-modes">Failure Modes</a></li>
<li><a class="reference internal" href="#membership-changes">Membership Changes</a></li>
<li><a class="reference internal" href="#important-notices">Important Notices</a></li>
</ul>
</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/sphinx_highlight.js"></script>
<script src="../_static/clipboard.min.js"></script>
<script src="../_static/copybutton.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 script type="text/javascript">
var collapsedSections = ['Developer Notes', 'Language Bindings', 'Libraries', 'Community'];
</script>
<img height="1" width="1" style="border-style:none;" alt="" src="https://www.googleadservices.com/pagead/conversion/795629140/?label=txkmCPmdtosBENSssfsC&guid=ON&script=0"/>
<!-- Begin Footer -->
<div class="container-fluid docs-tutorials-resources" id="docs-tutorials-resources">
<div class="container">
<div class="row">