-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathagent.html
1451 lines (1242 loc) · 109 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 2.4 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/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="Multiprocessing" href="multiprocessing.html" />
<link rel="prev" title="torchrun (Elastic Launch)" href="run.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>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">
<!-- 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.elastic.agent">
<span id="elastic-agent"></span><h1>Elastic Agent<a class="headerlink" href="#module-torch.distributed.elastic.agent" title="Permalink to this heading">¶</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 heading">¶</a></h2>
<p>The elastic agent is the control plane of torchelastic.</p>
<p>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 heading">¶</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 class="sig sig-object py" id="torch.distributed.elastic.agent.server.ElasticAgent">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></span><span class="sig-name descname"><span class="pre">ElasticAgent</span></span><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>An agent process responsible for managing one or more worker processes.</p>
<p>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">logger</span><span class="o">.</span><span class="n">exception</span><span class="p">(</span><span class="s2">"worker 0 failed with exit code : </span><span class="si">%s</span><span class="s2">"</span><span class="p">,</span> <span class="n">failure</span><span class="o">.</span><span class="n">exit_code</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 class="sig sig-object py" id="torch.distributed.elastic.agent.server.ElasticAgent.get_worker_group">
<em class="property"><span class="pre">abstract</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">get_worker_group</span></span><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><p>Return 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>.</p>
<p>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>
<dl class="field-list simple">
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p><a class="reference internal" href="#torch.distributed.elastic.agent.server.WorkerGroup" title="torch.distributed.elastic.agent.server.api.WorkerGroup"><em>WorkerGroup</em></a></p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.ElasticAgent.run">
<em class="property"><span class="pre">abstract</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">run</span></span><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>Run the agent.</p>
<p>Supports 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>
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p><a class="reference internal" href="#torch.distributed.elastic.agent.server.api.RunResult" title="torch.distributed.elastic.agent.server.api.RunResult"><em>RunResult</em></a></p>
</dd>
</dl>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.WorkerSpec">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></span><span class="sig-name descname"><span class="pre">WorkerSpec</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">role</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">local_world_size</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">rdzv_handler</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">fn</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">entrypoint</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">args</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">()</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">max_restarts</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">3</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">monitor_interval</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0.1</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">master_port</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">master_addr</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">local_addr</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/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>Blueprint information about a particular type of worker.</p>
<p>For a given role, there must only exist a single worker spec.
Worker spec is expected to be homogeneous 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> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.12)"><em>str</em></a>) – user-defined role for the workers with this spec</p></li>
<li><p><strong>local_world_size</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.12)"><em>int</em></a>) – number local workers to run</p></li>
<li><p><strong>fn</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Optional" title="(in Python v3.12)"><em>Optional</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Callable" title="(in Python v3.12)"><em>Callable</em></a><em>]</em>) – (deprecated use entrypoint instead)</p></li>
<li><p><strong>entrypoint</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Optional" title="(in Python v3.12)"><em>Optional</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Union" title="(in Python v3.12)"><em>Union</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Callable" title="(in Python v3.12)"><em>Callable</em></a><em>, </em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.12)"><em>str</em></a><em>]</em><em>]</em>) – worker function or command</p></li>
<li><p><strong>args</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Tuple" title="(in Python v3.12)"><em>Tuple</em></a>) – arguments to pass to <code class="docutils literal notranslate"><span class="pre">entrypoint</span></code></p></li>
<li><p><strong>rdzv_handler</strong> (<a class="reference internal" href="rendezvous.html#torch.distributed.elastic.rendezvous.RendezvousHandler" title="torch.distributed.elastic.rendezvous.api.RendezvousHandler"><em>RendezvousHandler</em></a>) – handles rdzv for this set of workers</p></li>
<li><p><strong>max_restarts</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.12)"><em>int</em></a>) – number of max retries for the workers</p></li>
<li><p><strong>monitor_interval</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#float" title="(in Python v3.12)"><em>float</em></a>) – 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> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Optional" title="(in Python v3.12)"><em>Optional</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.12)"><em>int</em></a><em>]</em>) – 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> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Optional" title="(in Python v3.12)"><em>Optional</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.12)"><em>str</em></a><em>]</em>) – 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 class="sig sig-object py" id="torch.distributed.elastic.agent.server.WorkerSpec.get_entrypoint_name">
<span class="sig-name descname"><span class="pre">get_entrypoint_name</span></span><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>Get the entry point name.</p>
<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 class="sig sig-object py" id="torch.distributed.elastic.agent.server.WorkerState">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></span><span class="sig-name descname"><span class="pre">WorkerState</span></span><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>A state of the <code class="docutils literal notranslate"><span class="pre">WorkerGroup</span></code>.</p>
<p>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 (interrupted) 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 class="sig sig-object py" id="torch.distributed.elastic.agent.server.WorkerState.is_running">
<em class="property"><span class="pre">static</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">is_running</span></span><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><p>Return the state of the Worker.</p>
<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>
<dt class="field-even">Return type</dt>
<dd class="field-even"><p><a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.12)">bool</a></p>
</dd>
</dl>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.Worker">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></span><span class="sig-name descname"><span class="pre">Worker</span></span><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">-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">-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">-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">-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>A worker instance.</p>
<p>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 v3.12)"><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 v3.12)"><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 v3.12)"><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 v3.12)"><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 v3.12)"><em>int</em></a>) – number of workers that have the same role</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.WorkerGroup">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></span><span class="sig-name descname"><span class="pre">WorkerGroup</span></span><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>A set of <code class="docutils literal notranslate"><span class="pre">Worker</span></code> instances.</p>
<p>The class defines a 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>
<dl class="field-list simple">
</dl>
</dd></dl>
</div>
<div class="section" id="implementations">
<h2>Implementations<a class="headerlink" href="#implementations" title="Permalink to this heading">¶</a></h2>
<p>Below are the agent implementations provided by torchelastic.</p>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.local_elastic_agent.LocalElasticAgent">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.local_elastic_agent.</span></span><span class="sig-name descname"><span class="pre">LocalElasticAgent</span></span><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">logs_specs</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_line_prefix_template</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.</p>
<p>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>A named pipe based watchdog can be enabled in <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code> if an
environment variable <code class="docutils literal notranslate"><span class="pre">TORCHELASTIC_ENABLE_FILE_TIMER</span></code> with value 1 has
been defined in the <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code> process.
Optionally, another environment variable <code class="docutils literal notranslate"><span class="pre">`TORCHELASTIC_TIMER_FILE`</span></code>
can be set with a unique file name for the named pipe. If the environment
variable <code class="docutils literal notranslate"><span class="pre">`TORCHELASTIC_TIMER_FILE`</span></code> is not set, <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code>
will internally create a unique file name and set it to the environment
variable <code class="docutils literal notranslate"><span class="pre">`TORCHELASTIC_TIMER_FILE`</span></code>, and this environment variable will
be propagated to the worker processes to allow them to connect to the same
named pipe that <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code> uses.</p>
<p>Logs are written to the specified log directory. Each log line will be by default
prefixed by <code class="docutils literal notranslate"><span class="pre">[${role_name}${local_rank}]:</span></code> (e.g. <code class="docutils literal notranslate"><span class="pre">[trainer0]:</span> <span class="pre">foobar</span></code>).
Log prefixes can be customized by passing a <a class="reference external" href="https://docs.python.org/3/library/string.html#template-strings">template string</a> as the
<code class="docutils literal notranslate"><span class="pre">log_line_prefix_template</span></code> argument.
The following macros (identifiers) are substituted at runtime:
<code class="docutils literal notranslate"><span class="pre">${role_name},</span> <span class="pre">${local_rank},</span> <span class="pre">${rank}</span></code>. For example, to prefix each log line with
global rank instead of the local rank, set <code class="docutils literal notranslate"><span class="pre">log_line_prefix_template</span> <span class="pre">=</span> <span class="pre">"[${rank}]:</span></code>.</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>
<dl class="field-list simple">
</dl>
</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 heading">¶</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 class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.</span></span><span class="sig-name descname"><span class="pre">SimpleElasticAgent</span></span><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 one particular type of worker role.</p>
<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>
such as one particular type of worker role.</p>
<dl class="field-list simple">
</dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent._assign_worker_ranks">
<span class="sig-name descname"><span class="pre">_assign_worker_ranks</span></span><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>Determine proper ranks for worker processes.</p>
<p>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>The rank 0 agent reads all the role_info from the store and
determines each agents worker ranks.</p></li>
<li><p>Determine the global rank: the global rank of the workers is computed
by cumulative sum of the local_world_size for all workers in front of it.
For efficiency reasons each worker is assigned a base global rank
such that it’s workers are in the range [base_global_rank,
base_global_rank + 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 ranks are calculated with
respect to the role name.</p></li>
<li><p>The rank 0 agent writes the assigned ranks to the store.</p></li>
<li><p>Each agent reads the assigned ranks from the store.</p></li>
</ol>
<p>Time complexity: each worker O(1), rank0 O(n), overall O(n)</p>
<dl class="field-list simple">
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.List" title="(in Python v3.12)"><em>List</em></a>[<a class="reference internal" href="#torch.distributed.elastic.agent.server.Worker" title="torch.distributed.elastic.agent.server.api.Worker"><em>Worker</em></a>]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent._exit_barrier">
<span class="sig-name descname"><span class="pre">_exit_barrier</span></span><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>Define a barrier that keeps the agent process alive until all workers finish.</p>
<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.</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent._initialize_workers">
<span class="sig-name descname"><span class="pre">_initialize_workers</span></span><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>Start a fresh set of workers for the worker_group.</p>
<p>Essentially, a rendezvous followed by a <code class="docutils literal notranslate"><span class="pre">start_workers</span></code>.
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>
<dl class="field-list simple">
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent._monitor_workers">
<em class="property"><span class="pre">abstract</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">_monitor_workers</span></span><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>Check on the workers for the <code class="docutils literal notranslate"><span class="pre">worker_group</span></code>.</p>
<p>This function also returns the new state of the worker group.</p>
<dl class="field-list simple">
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p><a class="reference internal" href="#torch.distributed.elastic.agent.server.api.RunResult" title="torch.distributed.elastic.agent.server.api.RunResult"><em>RunResult</em></a></p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent._rendezvous">
<span class="sig-name descname"><span class="pre">_rendezvous</span></span><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>Run rendezvous for the workers specified by the worker spec.</p>
<p>Assigns workers a new global rank and world size.
Updates the rendezvous store for the worker group.</p>
<dl class="field-list simple">
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent._restart_workers">
<span class="sig-name descname"><span class="pre">_restart_workers</span></span><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>Restart (stops, rendezvous, starts) all local workers in the group.</p>
<dl class="field-list simple">
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent._shutdown">
<em class="property"><span class="pre">abstract</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">_shutdown</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">death_sig</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">Signals.SIGTERM</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">is_restart</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></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>Clean 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> (<a class="reference external" href="https://docs.python.org/3/library/signal.html#signal.Signals" title="(in Python v3.12)"><em>Signals</em></a>) – Signal to send to the child process, SIGTERM is default</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent._start_workers">
<em class="property"><span class="pre">abstract</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">_start_workers</span></span><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>Start <code class="docutils literal notranslate"><span class="pre">worker_group.spec.local_world_size</span></code> number of workers.</p>
<p>This is according to worker spec for the worker group .
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>
<dl class="field-list simple">
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Dict" title="(in Python v3.12)"><em>Dict</em></a>[<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.12)">int</a>, <a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Any" title="(in Python v3.12)"><em>Any</em></a>]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.SimpleElasticAgent._stop_workers">
<em class="property"><span class="pre">abstract</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">_stop_workers</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">worker_group</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">is_restart</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</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>Stop all workers in the given worker group.</p>
<p>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>
<dl class="field-list simple">
</dl>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.elastic.agent.server.api.RunResult">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.elastic.agent.server.api.</span></span><span class="sig-name descname"><span class="pre">RunResult</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">return_values=<factory></span></span></em>, <em class="sig-param"><span class="n"><span class="pre">failures=<factory></span></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>Return results of the worker executions.</p>
<p>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>
<dl class="field-list simple">
</dl>
</dd></dl>
</div>
<div class="section" id="watchdog-in-the-agent">
<h2>Watchdog in the Agent<a class="headerlink" href="#watchdog-in-the-agent" title="Permalink to this heading">¶</a></h2>
<p>A named pipe based watchdog can be enabled in <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code> if an
environment variable <code class="docutils literal notranslate"><span class="pre">TORCHELASTIC_ENABLE_FILE_TIMER</span></code> with value 1 has
been defined in the <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code> process.
Optionally, another environment variable <code class="docutils literal notranslate"><span class="pre">`TORCHELASTIC_TIMER_FILE`</span></code>
can be set with a unique file name for the named pipe. If the environment
variable <code class="docutils literal notranslate"><span class="pre">`TORCHELASTIC_TIMER_FILE`</span></code> is not set, <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code>
will internally create a unique file name and set it to the environment
variable <code class="docutils literal notranslate"><span class="pre">`TORCHELASTIC_TIMER_FILE`</span></code>, and this environment variable will
be propagated to the worker processes to allow them to connect to the same
named pipe that <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code> uses.</p>
</div>
<div class="section" id="health-check-server">
<h2>Health Check Server<a class="headerlink" href="#health-check-server" title="Permalink to this heading">¶</a></h2>
<p>A health check monitoring server can be enabled in <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code>
if an environment variable <code class="docutils literal notranslate"><span class="pre">TORCHELASTIC_HEALTH_CHECK_PORT</span></code> has been defined
in the <code class="docutils literal notranslate"><span class="pre">`LocalElasticAgent`</span></code> process.