forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathagent.html
1136 lines (921 loc) · 75.1 KB
/
agent.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>Elastic Agent — PyTorch 1.10.1 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/elastic/agent.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/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/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="Multiprocessing" href="multiprocessing.html" />
<link rel="prev" title="torchrun (Elastic Launch)" href="run.html" />
<!-- Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-117752657-2"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-117752657-2');
</script>
<!-- End Google Analytics -->
<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>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<a href="https://pytorch.org/mobile">Mobile</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="active docs-active">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-orange-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/audio/stable/index.html">
<span class="dropdown-title">torchaudio</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/text/stable/index.html">
<span class="dropdown-title">torchtext</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/vision/stable/index.html">
<span class="dropdown-title">torchvision</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/elastic/">
<span class="dropdown-title">TorchElastic</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/serve/">
<span class="dropdown-title">TorchServe</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/xla">
<span class="dropdown-title">PyTorch on XLA Devices</span>
<p></p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-arrow">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class="dropdown-title">About</span>
<p>Learn about PyTorch’s features and capabilities</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://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://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/hub">
<span class="dropdown-title">Models (Beta)</span>
<p>Discover, publish, and reuse pre-trained models</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">GitHub</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">
<a href='https://pytorch.org/docs/versions.html'>1.10.1 ▼</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"><span class="caption-text">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/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/faq.html">Frequently Asked Questions</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/multiprocessing.html">Multiprocessing best practices</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"><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>
</ul>
<p class="caption"><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="../autograd.html">torch.autograd</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="../amp.html">torch.cuda.amp</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="../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="../distributed.optim.html">torch.distributed.optim</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="../fft.html">torch.fft</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="../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="../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="../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="../pipeline.html">Pipeline Parallelism</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="../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="../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="../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="../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__.html">torch.__config__</a></li>
</ul>
<p class="caption"><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/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/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="http://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</ul>
<p class="caption"><span class="caption-text">Community</span></p>
<ul>
<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/governance.html">PyTorch Governance</a></li>
<li class="toctree-l1"><a class="reference internal" href="../community/persons_of_interest.html">PyTorch Governance | Persons of Interest</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>Elastic Agent</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/elastic/agent.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="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.elastic.agent">
<span id="elastic-agent"></span><h1>Elastic Agent<a class="headerlink" href="#module-torch.distributed.elastic.agent" title="Permalink to this headline">¶</a></h1>
<div class="section" id="module-torch.distributed.elastic.agent.server">
<span id="server"></span><h2>Server<a class="headerlink" href="#module-torch.distributed.elastic.agent.server" title="Permalink to this headline">¶</a></h2>
<p>The elastic agent is the control plane of torchelastic. It is a process
that launches and manages underlying worker processes. The agent is
responsible for:</p>
<ol class="arabic simple">
<li><p>Working with distributed torch: the workers are started with all the
necessary information to successfully and trivially call
<code class="docutils literal notranslate"><span class="pre">torch.distributed.init_process_group()</span></code>.</p></li>
<li><p>Fault tolerance: monitors workers and upon detecting worker failures
or unhealthiness, tears down all workers and restarts everyone.</p></li>
<li><p>Elasticity: Reacts to membership changes and restarts workers with the new
members.</p></li>
</ol>
<p>The simplest agents are deployed per node and works with local processes.
A more advanced agent can launch and manage workers remotely. Agents can
be completely decentralized, making decisions based on the workers it manages.
Or can be coordinated, communicating to other agents (that manage workers
in the same job) to make a collective decision.</p>
<p>Below is a diagram of an agent that manages a local group of workers.</p>
<img alt="../_images/agent_diagram.jpg" src="../_images/agent_diagram.jpg" />
</div>
<div class="section" id="concepts">
<h2>Concepts<a class="headerlink" href="#concepts" title="Permalink to this headline">¶</a></h2>
<p>This section describes the high-level classes and concepts that
are relevant to understanding the role of the <code class="docutils literal notranslate"><span class="pre">agent</span></code> in torchelastic.</p>
<dl class="py class">
<dt id="torch.distributed.elastic.agent.server.ElasticAgent">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></code><code class="sig-name descname"><span class="pre">ElasticAgent</span></code><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#ElasticAgent"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.ElasticAgent" title="Permalink to this definition">¶</a></dt>
<dd><p>Agent process responsible for managing one or more worker processes.
The worker processes are assumed to be regular distributed PyTorch scripts.
When the worker process is created by the agent, the agent provides the
necessary information for the worker processes to properly initialize
a torch process group.</p>
<p>The exact deployment topology and ratio of agent-to-worker is dependent
on the specific implementation of the agent and the user’s job placement
preferences. For instance, to run a distributed training job on GPU with
8 trainers (one per GPU) one can:</p>
<ol class="arabic simple">
<li><p>Use 8 x single GPU instances, place an agent per instance, managing
1 worker per agent.</p></li>
<li><p>Use 4 x double GPU instances, place an agent per instance, managing
2 workers per agent.</p></li>
<li><p>Use 2 x quad GPU instances, place an agent per instance, managing
4 workers per agent.</p></li>
<li><p>Use 1 x 8 GPU instance, place an agent per instance, managing
8 workers per agent.</p></li>
</ol>
<p>Usage</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">group_result</span> <span class="o">=</span> <span class="n">agent</span><span class="o">.</span><span class="n">run</span><span class="p">()</span>
<span class="k">if</span> <span class="n">group_result</span><span class="o">.</span><span class="n">is_failed</span><span class="p">():</span>
<span class="c1"># workers failed</span>
<span class="n">failure</span> <span class="o">=</span> <span class="n">group_result</span><span class="o">.</span><span class="n">failures</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
<span class="n">log</span><span class="o">.</span><span class="n">exception</span><span class="p">(</span><span class="sa">f</span><span class="s2">"worker 0 failed with exit code : </span><span class="si">{</span><span class="n">failure</span><span class="o">.</span><span class="n">exit_code</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="n">group_result</span><span class="o">.</span><span class="n">return_values</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="c1"># return rank 0's results</span>
</pre></div>
</div>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.ElasticAgent.get_worker_group">
<em class="property"><span class="pre">abstract</span> </em><code class="sig-name descname"><span class="pre">get_worker_group</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">role</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">'default'</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#ElasticAgent.get_worker_group"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.ElasticAgent.get_worker_group" title="Permalink to this definition">¶</a></dt>
<dd><dl class="field-list simple">
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>The <code class="docutils literal notranslate"><span class="pre">WorkerGroup</span></code> for the given <code class="docutils literal notranslate"><span class="pre">role</span></code>.
Note that the worker group is a mutable object and hence in a
multi-threaded/process environment it may change state.
Implementors are encouraged (but not required) to return
a defensive read-only copy.</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.ElasticAgent.run">
<em class="property"><span class="pre">abstract</span> </em><code class="sig-name descname"><span class="pre">run</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">role</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">'default'</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#ElasticAgent.run"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.ElasticAgent.run" title="Permalink to this definition">¶</a></dt>
<dd><p>Runs the agent, retrying the worker group on failures up to
<code class="docutils literal notranslate"><span class="pre">max_restarts</span></code>.</p>
<dl class="field-list simple">
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>The result of the execution, containing the return values or
failure details for each worker mapped by the worker’s global rank.</p>
</dd>
<dt class="field-even">Raises</dt>
<dd class="field-even"><p><strong>Exception - any other failures NOT related to worker process</strong> – </p>
</dd>
</dl>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt id="torch.distributed.elastic.agent.server.WorkerSpec">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></code><code class="sig-name descname"><span class="pre">WorkerSpec</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="pre">role</span></em>, <em class="sig-param"><span class="pre">local_world_size</span></em>, <em class="sig-param"><span class="pre">rdzv_handler</span></em>, <em class="sig-param"><span class="pre">fn=None</span></em>, <em class="sig-param"><span class="pre">entrypoint=None</span></em>, <em class="sig-param"><span class="pre">args=()</span></em>, <em class="sig-param"><span class="pre">max_restarts=3</span></em>, <em class="sig-param"><span class="pre">monitor_interval=30.0</span></em>, <em class="sig-param"><span class="pre">master_port=None</span></em>, <em class="sig-param"><span class="pre">master_addr=None</span></em>, <em class="sig-param"><span class="pre">redirects=<Std.NONE:</span> <span class="pre">0></span></em>, <em class="sig-param"><span class="pre">tee=<Std.NONE:</span> <span class="pre">0></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#WorkerSpec"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.WorkerSpec" title="Permalink to this definition">¶</a></dt>
<dd><p>Contains blueprint information about a particular type of worker.
For a given role, there must only exist a single worker spec.
Worker spec is expected to be homogenous across all nodes (machine),
that is each node runs the same number of workers for a particular spec.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>role</strong> – user-defined role for the workers with this spec</p></li>
<li><p><strong>local_world_size</strong> – number local workers to run</p></li>
<li><p><strong>fn</strong> – (deprecated use entrypoint instead)</p></li>
<li><p><strong>entrypoint</strong> – worker function or command</p></li>
<li><p><strong>args</strong> – arguments to pass to <code class="docutils literal notranslate"><span class="pre">entrypoint</span></code></p></li>
<li><p><strong>rdzv_handler</strong> – handles rdzv for this set of workers</p></li>
<li><p><strong>max_restarts</strong> – number of max retries for the workers</p></li>
<li><p><strong>monitor_interval</strong> – monitor status of workers every <code class="docutils literal notranslate"><span class="pre">n</span></code> seconds</p></li>
<li><p><strong>master_port</strong> – fixed port to run the c10d store on rank 0
if not specified then will chose a random free port</p></li>
<li><p><strong>master_addr</strong> – fixed master_addr to run the c10d store on rank 0
if not specified then will chose hostname on agent rank 0</p></li>
<li><p><strong>redirects</strong> – redirect std streams to a file,
selectively redirect for a particular
local rank by passing a map</p></li>
<li><p><strong>tee</strong> – tees the specified std stream(s) to console + file,
selectively tee for a particular local rank by passing a map,
takes precedence over <code class="docutils literal notranslate"><span class="pre">redirects</span></code> settings.</p></li>
</ul>
</dd>
</dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.WorkerSpec.get_entrypoint_name">
<code class="sig-name descname"><span class="pre">get_entrypoint_name</span></code><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#WorkerSpec.get_entrypoint_name"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.WorkerSpec.get_entrypoint_name" title="Permalink to this definition">¶</a></dt>
<dd><p>If the entrypoint is a function (e.g. <code class="docutils literal notranslate"><span class="pre">Callable</span></code>) returns its <code class="docutils literal notranslate"><span class="pre">__qualname__</span></code>,
else if the entrypoint is a binary (e.g. <code class="docutils literal notranslate"><span class="pre">str</span></code>), returns the binary name.</p>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt id="torch.distributed.elastic.agent.server.WorkerState">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></code><code class="sig-name descname"><span class="pre">WorkerState</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">value</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#WorkerState"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.WorkerState" title="Permalink to this definition">¶</a></dt>
<dd><p>State of the <code class="docutils literal notranslate"><span class="pre">WorkerGroup</span></code>. Workers in a worker group change state as a unit.
If a single worker in a worker group fails the entire set is considered
failed:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span>UNKNOWN - agent lost track of worker group state, unrecoverable
INIT - worker group object created not yet started
HEALTHY - workers running and healthy
UNHEALTHY - workers running and unhealthy
STOPPED - workers stopped (interruped) by the agent
SUCCEEDED - workers finished running (exit 0)
FAILED - workers failed to successfully finish (exit !0)
</pre></div>
</div>
<p>A worker group starts from an initial <code class="docutils literal notranslate"><span class="pre">INIT</span></code> state,
then progresses to <code class="docutils literal notranslate"><span class="pre">HEALTHY</span></code> or <code class="docutils literal notranslate"><span class="pre">UNHEALTHY</span></code> states,
and finally reaches a terminal <code class="docutils literal notranslate"><span class="pre">SUCCEEDED</span></code> or <code class="docutils literal notranslate"><span class="pre">FAILED</span></code> state.</p>
<p>Worker groups can be interrupted and temporarily put into <code class="docutils literal notranslate"><span class="pre">STOPPED</span></code> state
by the agent. Workers in <code class="docutils literal notranslate"><span class="pre">STOPPED</span></code> state are scheduled to be restarted
in the near future by the agent. Some examples of workers being put into
<code class="docutils literal notranslate"><span class="pre">STOPPED</span></code> state are:</p>
<ol class="arabic simple">
<li><p>Worker group failure|unhealthy observed</p></li>
<li><p>Membership change detected</p></li>
</ol>
<p>When actions (start, stop, rdzv, retry, etc) on worker group fails
and results in the action being partially applied to the worker group
the state will be <code class="docutils literal notranslate"><span class="pre">UNKNOWN</span></code>. Typically this happens on uncaught/unhandled
exceptions during state change events on the agent. The agent is not
expected to recover worker groups in <code class="docutils literal notranslate"><span class="pre">UNKNOWN</span></code> state and is better off
self terminating and allowing the job manager to retry the node.</p>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.WorkerState.is_running">
<em class="property"><span class="pre">static</span> </em><code class="sig-name descname"><span class="pre">is_running</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#WorkerState.is_running"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.WorkerState.is_running" title="Permalink to this definition">¶</a></dt>
<dd><dl class="field-list simple">
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>True if the worker state represents workers still running
(e.g. that the process exists but not necessarily healthy).</p>
</dd>
</dl>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt id="torch.distributed.elastic.agent.server.Worker">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></code><code class="sig-name descname"><span class="pre">Worker</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">local_rank</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">global_rank</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">-</span> <span class="pre">1</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">role_rank</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">-</span> <span class="pre">1</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">world_size</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">-</span> <span class="pre">1</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">role_world_size</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">-</span> <span class="pre">1</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#Worker"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.Worker" title="Permalink to this definition">¶</a></dt>
<dd><p>Represents a worker instance. Contrast this with <code class="docutils literal notranslate"><span class="pre">WorkerSpec</span></code> that
represents the specifications of a worker. A <code class="docutils literal notranslate"><span class="pre">Worker</span></code> is created from
a <code class="docutils literal notranslate"><span class="pre">WorkerSpec</span></code>. A <code class="docutils literal notranslate"><span class="pre">Worker</span></code> is to a <code class="docutils literal notranslate"><span class="pre">WorkerSpec</span></code> as an object is to
a class.</p>
<p>The <code class="docutils literal notranslate"><span class="pre">id</span></code> of the worker is interpreted
by the specific implementation of <code class="docutils literal notranslate"><span class="pre">ElasticAgent</span></code>. For a local
agent, it could be the <code class="docutils literal notranslate"><span class="pre">pid</span> <span class="pre">(int)</span></code> of the worker, for a remote
agent it could be encoded as <code class="docutils literal notranslate"><span class="pre">host:port</span> <span class="pre">(string)</span></code>.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>id</strong> (<em>Any</em>) – uniquely identifies a worker (interpreted by the agent)</p></li>
<li><p><strong>local_rank</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a>) – local rank of the worker</p></li>
<li><p><strong>global_rank</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a>) – global rank of the worker</p></li>
<li><p><strong>role_rank</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a>) – rank of the worker across all workers that have the same role</p></li>
<li><p><strong>world_size</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a>) – number of workers (globally)</p></li>
<li><p><strong>role_world_size</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a>) – number of workers that have the same role</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="py class">
<dt id="torch.distributed.elastic.agent.server.WorkerGroup">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></code><code class="sig-name descname"><span class="pre">WorkerGroup</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">spec</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#WorkerGroup"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.WorkerGroup" title="Permalink to this definition">¶</a></dt>
<dd><p>Represents the set of <code class="docutils literal notranslate"><span class="pre">Worker</span></code> instances for the given <code class="docutils literal notranslate"><span class="pre">WorkerSpec</span></code>
managed by <code class="docutils literal notranslate"><span class="pre">ElasticAgent</span></code>. Whether the worker group contains cross
instance workers or not depends on the implementation of the agent.</p>
</dd></dl>
</div>
<div class="section" id="implementations">
<h2>Implementations<a class="headerlink" href="#implementations" title="Permalink to this headline">¶</a></h2>
<p>Below are the agent implementations provided by torchelastic.</p>
<dl class="py class">
<dt id="torch.distributed.elastic.agent.server.local_elastic_agent.LocalElasticAgent">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.local_elastic_agent.</span></code><code class="sig-name descname"><span class="pre">LocalElasticAgent</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">spec</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">start_method</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">'spawn'</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">exit_barrier_timeout</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">300</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">log_dir</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/local_elastic_agent.html#LocalElasticAgent"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.local_elastic_agent.LocalElasticAgent" title="Permalink to this definition">¶</a></dt>
<dd><p>An implementation of <code class="xref py py-class docutils literal notranslate"><span class="pre">torchelastic.agent.server.ElasticAgent</span></code>
that handles host-local workers.
This agent is deployed per host and is configured to spawn <code class="docutils literal notranslate"><span class="pre">n</span></code> workers.
When using GPUs, <code class="docutils literal notranslate"><span class="pre">n</span></code> maps to the number of GPUs available on the host.</p>
<p>The local agent does not communicate to other local agents deployed on
other hosts, even if the workers may communicate inter-host. The worker id
is interpreted to be a local process. The agent starts and stops all worker
processes as a single unit.</p>
<p>The worker function and argument passed to the worker function must be
python multiprocessing compatible. To pass multiprocessing data structures
to the workers you may create the data structure in the same multiprocessing
context as the specified <code class="docutils literal notranslate"><span class="pre">start_method</span></code> and pass it as a function argument.</p>
<p>The <code class="docutils literal notranslate"><span class="pre">exit_barrier_timeout</span></code> specifies the amount of time (in seconds) to wait
for other agents to finish. This acts as a safety net to handle cases where
workers finish at different times, to prevent agents from viewing workers
that finished early as a scale-down event. It is strongly advised that the
user code deal with ensuring that workers are terminated in a synchronous
manner rather than relying on the exit_barrier_timeout.</p>
<p>Example launching function</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">trainer</span><span class="p">(</span><span class="n">args</span><span class="p">)</span> <span class="o">-></span> <span class="nb">str</span><span class="p">:</span>
<span class="k">return</span> <span class="s2">"do train"</span>
<span class="k">def</span> <span class="nf">main</span><span class="p">():</span>
<span class="n">start_method</span><span class="o">=</span><span class="s2">"spawn"</span>
<span class="n">shared_queue</span><span class="o">=</span> <span class="n">multiprocessing</span><span class="o">.</span><span class="n">get_context</span><span class="p">(</span><span class="n">start_method</span><span class="p">)</span><span class="o">.</span><span class="n">Queue</span><span class="p">()</span>
<span class="n">spec</span> <span class="o">=</span> <span class="n">WorkerSpec</span><span class="p">(</span>
<span class="n">role</span><span class="o">=</span><span class="s2">"trainer"</span><span class="p">,</span>
<span class="n">local_world_size</span><span class="o">=</span><span class="n">nproc_per_process</span><span class="p">,</span>
<span class="n">entrypoint</span><span class="o">=</span><span class="n">trainer</span><span class="p">,</span>
<span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="s2">"foobar"</span><span class="p">,),</span>
<span class="o">...<</span><span class="n">OTHER_PARAMS</span><span class="o">...></span><span class="p">)</span>
<span class="n">agent</span> <span class="o">=</span> <span class="n">LocalElasticAgent</span><span class="p">(</span><span class="n">spec</span><span class="p">,</span> <span class="n">start_method</span><span class="p">)</span>
<span class="n">results</span> <span class="o">=</span> <span class="n">agent</span><span class="o">.</span><span class="n">run</span><span class="p">()</span>
<span class="k">if</span> <span class="n">results</span><span class="o">.</span><span class="n">is_failed</span><span class="p">():</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"trainer failed"</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"rank 0 return value: </span><span class="si">{</span><span class="n">results</span><span class="o">.</span><span class="n">return_values</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="c1"># prints -> rank 0 return value: do train</span>
</pre></div>
</div>
<p>Example launching binary</p>
<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">spec</span> <span class="o">=</span> <span class="n">WorkerSpec</span><span class="p">(</span>
<span class="n">role</span><span class="o">=</span><span class="s2">"trainer"</span><span class="p">,</span>
<span class="n">local_world_size</span><span class="o">=</span><span class="n">nproc_per_process</span><span class="p">,</span>
<span class="n">entrypoint</span><span class="o">=</span><span class="s2">"/usr/local/bin/trainer"</span><span class="p">,</span>
<span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="s2">"--trainer_args"</span><span class="p">,</span> <span class="s2">"foobar"</span><span class="p">),</span>
<span class="o">...<</span><span class="n">OTHER_PARAMS</span><span class="o">...></span><span class="p">)</span>
<span class="n">agent</span> <span class="o">=</span> <span class="n">LocalElasticAgent</span><span class="p">(</span><span class="n">spec</span><span class="p">)</span>
<span class="n">results</span> <span class="o">=</span> <span class="n">agent</span><span class="o">.</span><span class="n">run</span><span class="p">()</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">results</span><span class="o">.</span><span class="n">is_failed</span><span class="p">():</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"binary launches do not have return values"</span><span class="p">)</span>
</pre></div>
</div>
</dd></dl>
</div>
<div class="section" id="extending-the-agent">
<h2>Extending the Agent<a class="headerlink" href="#extending-the-agent" title="Permalink to this headline">¶</a></h2>
<p>To extend the agent you can implement <code class="docutils literal notranslate"><span class="pre">`ElasticAgent</span></code> directly, however
we recommend you extend <code class="docutils literal notranslate"><span class="pre">SimpleElasticAgent</span></code> instead, which provides
most of the scaffolding and leaves you with a few specific abstract methods
to implement.</p>
<dl class="py class">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></code><code class="sig-name descname"><span class="pre">SimpleElasticAgent</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">spec</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">exit_barrier_timeout</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">300</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent" title="Permalink to this definition">¶</a></dt>
<dd><p>An <code class="docutils literal notranslate"><span class="pre">ElasticAgent</span></code> that manages workers (<code class="docutils literal notranslate"><span class="pre">WorkerGroup</span></code>)
for a single <code class="docutils literal notranslate"><span class="pre">WorkerSpec</span></code> (e.g. one particular type of worker role).</p>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent._assign_worker_ranks">
<code class="sig-name descname"><span class="pre">_assign_worker_ranks</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">store</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">group_rank</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">group_world_size</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">spec</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent._assign_worker_ranks"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent._assign_worker_ranks" title="Permalink to this definition">¶</a></dt>
<dd><p>Determines proper ranks for worker processes. The rank assignment
is done according to the following algorithm:</p>
<ol class="arabic simple">
<li><p>Each agent writes its configuration(group_rank, group_world_size
, num_workers) to the common store.</p></li>
<li><p>Each agent retrieves configuration for all agents
and performs two level sort using role and rank.</p></li>
<li><p>Determine the global rank: the global rank of the workers for the current
agent is the offset of the infos array up to group_rank of the agent.
The offset is computed as a sum of local_world_size of all agents that
have rank less than the group_rank. The workers would have the ranks:
[offset, offset+local_world_size)</p></li>
<li><p>Determine the role rank: The role rank is determined using the algorithms
in the point 3 with the exception that the offset is done from the first
agent that has the same role as current one and has the minimum group rank.</p></li>
</ol>
</dd></dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent._exit_barrier">
<code class="sig-name descname"><span class="pre">_exit_barrier</span></code><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent._exit_barrier"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent._exit_barrier" title="Permalink to this definition">¶</a></dt>
<dd><p>Wait for <code class="docutils literal notranslate"><span class="pre">exit_barrier_timeout</span></code> seconds for all agents to finish
executing their local workers (either successfully or not). This
acts as a safety guard against user scripts that terminate at different
times. This barrier keeps the agent process alive until all workers finish.</p>
</dd></dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent._initialize_workers">
<code class="sig-name descname"><span class="pre">_initialize_workers</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">worker_group</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent._initialize_workers"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent._initialize_workers" title="Permalink to this definition">¶</a></dt>
<dd><p>Starts a fresh set of workers for the worker_group.
Essentially a rendezvous followed by a start_workers.</p>
<p>The caller should first call <code class="docutils literal notranslate"><span class="pre">_stop_workers()</span></code> to stop running workers
prior to calling this method.</p>
<p>Optimistically sets the state of the worker group that
just started as <code class="docutils literal notranslate"><span class="pre">HEALTHY</span></code> and delegates the actual monitoring
of state to <code class="docutils literal notranslate"><span class="pre">_monitor_workers()</span></code> method</p>
</dd></dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent._monitor_workers">
<em class="property"><span class="pre">abstract</span> </em><code class="sig-name descname"><span class="pre">_monitor_workers</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">worker_group</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent._monitor_workers"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent._monitor_workers" title="Permalink to this definition">¶</a></dt>
<dd><p>Checks on the workers for the <code class="docutils literal notranslate"><span class="pre">worker_group</span></code> and returns
the new state of the worker group.</p>
</dd></dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent._rendezvous">
<code class="sig-name descname"><span class="pre">_rendezvous</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">worker_group</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent._rendezvous"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent._rendezvous" title="Permalink to this definition">¶</a></dt>
<dd><p>Runs rendezvous for the workers specified by worker spec.
Assigns workers a new global rank and world size.
Updates the rendezvous store for the worker group.</p>
</dd></dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent._restart_workers">
<code class="sig-name descname"><span class="pre">_restart_workers</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">worker_group</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent._restart_workers"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent._restart_workers" title="Permalink to this definition">¶</a></dt>
<dd><p>Restarts (stops, rendezvous, starts) all local workers in the group.</p>
</dd></dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent._shutdown">
<em class="property"><span class="pre">abstract</span> </em><code class="sig-name descname"><span class="pre">_shutdown</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="pre">death_sig=<Signals.SIGTERM:</span> <span class="pre">15></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent._shutdown"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent._shutdown" title="Permalink to this definition">¶</a></dt>
<dd><p>Cleans up any resources that were allocated during the agent’s work.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>death_sig</strong> – Signal to send to the child process, SIGTERM is default</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent._start_workers">
<em class="property"><span class="pre">abstract</span> </em><code class="sig-name descname"><span class="pre">_start_workers</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">worker_group</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent._start_workers"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent._start_workers" title="Permalink to this definition">¶</a></dt>
<dd><p>Starts <code class="docutils literal notranslate"><span class="pre">worker_group.spec.local_world_size</span></code> number of workers
according to worker spec for the worker group .</p>
<p>Returns a map of <code class="docutils literal notranslate"><span class="pre">local_rank</span></code> to worker <code class="docutils literal notranslate"><span class="pre">id</span></code>.</p>
</dd></dl>
<dl class="py method">
<dt id="torch.distributed.elastic.agent.server.SimpleElasticAgent._stop_workers">
<em class="property"><span class="pre">abstract</span> </em><code class="sig-name descname"><span class="pre">_stop_workers</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">worker_group</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#SimpleElasticAgent._stop_workers"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.SimpleElasticAgent._stop_workers" title="Permalink to this definition">¶</a></dt>
<dd><p>Stops all workers in the given worker group. Implementors
must deal with workers in all states defined by <code class="docutils literal notranslate"><span class="pre">WorkerState</span></code>.
That is, it must gracefully handle stopping non-existent workers,
unhealthy (stuck) workers, etc.</p>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt id="torch.distributed.elastic.agent.server.api.RunResult">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.api.</span></code><code class="sig-name descname"><span class="pre">RunResult</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="pre">state</span></em>, <em class="sig-param"><span class="pre">return_values=<factory></span></em>, <em class="sig-param"><span class="pre">failures=<factory></span></em><span class="sig-paren">)</span><a class="reference internal" href="../_modules/torch/distributed/elastic/agent/server/api.html#RunResult"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.elastic.agent.server.api.RunResult" title="Permalink to this definition">¶</a></dt>
<dd><p>Results returned by the worker executions. Run results follow an “all-or-nothing” policy
where the run is successful if and only if ALL local workers managed by this agent
complete successfully.</p>
<p>If the result is successful (e.g. <code class="docutils literal notranslate"><span class="pre">is_failed()</span> <span class="pre">=</span> <span class="pre">False</span></code>) then the <code class="docutils literal notranslate"><span class="pre">return_values</span></code>
field contains the outputs (return values) of the workers managed by THIS agent mapped
by their GLOBAL ranks. That is <code class="docutils literal notranslate"><span class="pre">result.return_values[0]</span></code> is the return value of
global rank 0.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p><code class="docutils literal notranslate"><span class="pre">return_values</span></code> are only meaningful for when the worker entrypoint
is a function. Workers specified as a binary entrypoint do not canonically
have a return value and the <code class="docutils literal notranslate"><span class="pre">return_values</span></code> field is meaningless and
may be empty.</p>
</div>
<p>If <code class="docutils literal notranslate"><span class="pre">is_failed()</span></code> returns <code class="docutils literal notranslate"><span class="pre">True</span></code> then the <code class="docutils literal notranslate"><span class="pre">failures</span></code> field contains the
failure information, again, mapped by the GLOBAL rank of the worker that failed.</p>
<p>The keys in <code class="docutils literal notranslate"><span class="pre">return_values</span></code> and <code class="docutils literal notranslate"><span class="pre">failures</span></code> are mutually exclusive, that is,
a worker’s final state can only be one of: succeeded, failed. Workers intentionally
terminated by the agent according to the agent’s restart policy, are not represented
in either <code class="docutils literal notranslate"><span class="pre">return_values</span></code> nor <code class="docutils literal notranslate"><span class="pre">failures</span></code>.</p>
</dd></dl>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="multiprocessing.html" class="btn btn-neutral float-right" title="Multiprocessing" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="run.html" class="btn btn-neutral" title="torchrun (Elastic Launch)" accesskey="p" rel="prev"><img src="../_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright 2019, Torch 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>
</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="#">Elastic Agent</a><ul>
<li><a class="reference internal" href="#module-torch.distributed.elastic.agent.server">Server</a></li>
<li><a class="reference internal" href="#concepts">Concepts</a></li>
<li><a class="reference internal" href="#implementations">Implementations</a></li>
<li><a class="reference internal" href="#extending-the-agent">Extending the Agent</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 src="../_static/jquery.js"></script>
<script src="../_static/underscore.js"></script>
<script src="../_static/doctools.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 = ['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">
<div class="col-md-4 text-center">
<h2>Docs</h2>
<p>Access comprehensive developer documentation for PyTorch</p>
<a class="with-right-arrow" href="https://pytorch.org/docs/stable/index.html">View Docs</a>
</div>
<div class="col-md-4 text-center">
<h2>Tutorials</h2>
<p>Get in-depth tutorials for beginners and advanced developers</p>
<a class="with-right-arrow" href="https://pytorch.org/tutorials">View Tutorials</a>
</div>
<div class="col-md-4 text-center">
<h2>Resources</h2>
<p>Find development resources and get your questions answered</p>
<a class="with-right-arrow" href="https://pytorch.org/resources">View Resources</a>
</div>
</div>
</div>
</div>
<footer class="site-footer">
<div class="container footer-container">
<div class="footer-logo-wrapper">
<a href="https://pytorch.org/" class="footer-logo"></a>
</div>
<div class="footer-links-wrapper">
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/">PyTorch</a></li>
<li><a href="https://pytorch.org/get-started">Get Started</a></li>
<li><a href="https://pytorch.org/features">Features</a></li>
<li><a href="https://pytorch.org/ecosystem">Ecosystem</a></li>
<li><a href="https://pytorch.org/blog/">Blog</a></li>
<li><a href="https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md">Contributing</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/resources">Resources</a></li>
<li><a href="https://pytorch.org/tutorials">Tutorials</a></li>
<li><a href="https://pytorch.org/docs/stable/index.html">Docs</a></li>
<li><a href="https://discuss.pytorch.org" target="_blank">Discuss</a></li>
<li><a href="https://github.com/pytorch/pytorch/issues" target="_blank">Github Issues</a></li>
<li><a href="https://pytorch.org/assets/brand-guidelines/PyTorch-Brand-Guidelines.pdf" target="_blank">Brand Guidelines</a></li>
</ul>
</div>
<div class="footer-links-col follow-us-col">
<ul>
<li class="list-title">Stay Connected</li>
<li>
<div id="mc_embed_signup">
<form
action="https://twitter.us14.list-manage.com/subscribe/post?u=75419c71fe0a935e53dfa4a3f&id=91d0dccd39"
method="post"
id="mc-embedded-subscribe-form"
name="mc-embedded-subscribe-form"
class="email-subscribe-form validate"
target="_blank"
novalidate>
<div id="mc_embed_signup_scroll" class="email-subscribe-form-fields-wrapper">
<div class="mc-field-group">
<label for="mce-EMAIL" style="display:none;">Email Address</label>
<input type="email" value="" name="EMAIL" class="required email" id="mce-EMAIL" placeholder="Email Address">
</div>
<div id="mce-responses" class="clear">
<div class="response" id="mce-error-response" style="display:none"></div>
<div class="response" id="mce-success-response" style="display:none"></div>
</div> <!-- real people should not fill this in and expect good things - do not remove this or risk form bot signups-->
<div style="position: absolute; left: -5000px;" aria-hidden="true"><input type="text" name="b_75419c71fe0a935e53dfa4a3f_91d0dccd39" tabindex="-1" value=""></div>
<div class="clear">
<input type="submit" value="" name="subscribe" id="mc-embedded-subscribe" class="button email-subscribe-button">
</div>
</div>
</form>
</div>
</li>
</ul>
<div class="footer-social-icons">
<a href="https://www.facebook.com/pytorch" target="_blank" class="facebook"></a>
<a href="https://twitter.com/pytorch" target="_blank" class="twitter"></a>
<a href="https://www.youtube.com/pytorch" target="_blank" class="youtube"></a>
</div>
</div>
</div>
</div>
</footer>
<div class="cookie-banner-wrapper">
<div class="container">
<p class="gdpr-notice">To analyze traffic and optimize your experience, we serve cookies on this site. By clicking or navigating, you agree to allow our usage of cookies. As the current maintainers of this site, Facebook’s Cookies Policy applies. Learn more, including about available controls: <a href="https://www.facebook.com/policies/cookies/">Cookies Policy</a>.</p>