forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdistributed.optim.html
1189 lines (979 loc) · 84.1 KB
/
distributed.optim.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 charset="utf-8">
<meta name="generator" content="Docutils 0.18.1: http://docutils.sourceforge.net/" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Distributed Optimizers — PyTorch master documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/distributed.optim.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="Tensor Parallelism" href="distributed.tensor.parallel.html" />
<link rel="prev" title="FullyShardedDataParallel" href="fsdp.html" />
<!--
Search engines should not index the master version of documentation.
Stable documentation are built without release == 'master'.
-->
<meta name="robots" content="noindex">
<!-- 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/torcharrow">
<span class="dropdown-title">torcharrow</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/data">
<span class="dropdown-title">TorchData</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torchrec">
<span class="dropdown-title">TorchRec</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/torchx/">
<span class="dropdown-title">TorchX</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/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn about the PyTorch foundation</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/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/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/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</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'>master (1.14.0a0+git876b702 ) ▼</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>
<div>
<a style="color:#F05732" href="https://pytorch.org/docs/stable/distributed.optim.html">
You are viewing unstable developer preview docs.
Click here to view docs for latest stable release.
</a>
</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">CUDA 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/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">torch.compile</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="dynamo/index.html">TorchDynamo Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/installation.html">Installing TorchDynamo</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/get-started.html">Getting Started</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/guards-overview.html">Guards Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/custom-backends.html">Custom Backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/deep-dive.html">TorchDynamo Deeper Dive</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/troubleshooting.html">TorchDynamo Troubleshooting</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/faq.html">Frequently Asked Questions</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="cuda.html">torch.cuda</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"><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 current"><a class="current reference internal" href="#">torch.distributed.optim</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="_dynamo.html">torch._dynamo</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="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="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx_diagnostics.html">torch.onnx diagnostics</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="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="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="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="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>
</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="http://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>Distributed Optimizers</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/distributed.optim.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">
<section id="distributed-optimizers">
<h1>Distributed Optimizers<a class="headerlink" href="#distributed-optimizers" title="Permalink to this heading">¶</a></h1>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Distributed optimizer is not currently supported when using CUDA tensors</p>
</div>
<span class="target" id="module-torch.distributed.optim"></span><p><a class="reference internal" href="#module-torch.distributed.optim" title="torch.distributed.optim"><code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.distributed.optim</span></code></a> exposes DistributedOptimizer, which takes a list
of remote parameters (<a class="reference internal" href="rpc.html#torch.distributed.rpc.RRef" title="torch.distributed.rpc.RRef"><code class="xref py py-class docutils literal notranslate"><span class="pre">RRef</span></code></a>) and runs the
optimizer locally on the workers where the parameters live. The distributed
optimizer can use any of the local optimizer <a class="reference internal" href="optim.html#optimizer-algorithms"><span class="std std-ref">Base class</span></a> to
apply the gradients on each worker.</p>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.optim.DistributedOptimizer">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.optim.</span></span><span class="sig-name descname"><span class="pre">DistributedOptimizer</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">optimizer_class</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">params_rref</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/optimizer.html#DistributedOptimizer"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.DistributedOptimizer" title="Permalink to this definition">¶</a></dt>
<dd><p>DistributedOptimizer takes remote references to parameters scattered
across workers and applies the given optimizer locally for each parameter.</p>
<p>This class uses <a class="reference internal" href="rpc.html#torch.distributed.autograd.get_gradients" title="torch.distributed.autograd.get_gradients"><code class="xref py py-meth docutils literal notranslate"><span class="pre">get_gradients()</span></code></a> in order
to retrieve the gradients for specific parameters.</p>
<p>Concurrent calls to
<a class="reference internal" href="#torch.distributed.optim.DistributedOptimizer.step" title="torch.distributed.optim.DistributedOptimizer.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code></a>,
either from the same or different clients, will
be serialized on each worker – as each worker’s optimizer can only work
on one set of gradients at a time. However, there is no guarantee that
the full forward-backward-optimizer sequence will execute for one client
at a time. This means that the gradients being applied may not correspond
to the latest forward pass executed on a given worker. Also, there is no
guaranteed ordering across workers.</p>
<p><cite>DistributedOptimizer</cite> creates the local optimizer with TorchScript enabled
by default, so that optimizer updates are not blocked by the Python Global
Interpreter Lock (GIL) in the case of multithreaded training (e.g. Distributed
Model Parallel). This feature is currently enabled for most optimizers. You
can also follow <a class="reference external" href="https://github.com/pytorch/tutorials/pull/1465">the recipe</a> in PyTorch tutorials to enable TorchScript support
for your own custom optimizers.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>optimizer_class</strong> (<a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.Optimizer"><em>optim.Optimizer</em></a>) – the class of optimizer to
instantiate on each worker.</p></li>
<li><p><strong>params_rref</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#list" title="(in Python v3.11)"><em>list</em></a><em>[</em><a class="reference internal" href="rpc.html#torch.distributed.rpc.RRef" title="torch.distributed.rpc.RRef"><em>RRef</em></a><em>]</em>) – list of RRefs to local or remote parameters
to optimize.</p></li>
<li><p><strong>args</strong> – arguments to pass to the optimizer constructor on each worker.</p></li>
<li><p><strong>kwargs</strong> – arguments to pass to the optimizer constructor on each worker.</p></li>
</ul>
</dd>
</dl>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch.distributed.autograd</span> <span class="k">as</span> <span class="nn">dist_autograd</span>
<span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">optim</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.optim</span> <span class="kn">import</span> <span class="n">DistributedOptimizer</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="k">with</span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">context</span><span class="p">()</span> <span class="k">as</span> <span class="n">context_id</span><span class="p">:</span>
<span class="gp">>>> </span> <span class="c1"># Forward pass.</span>
<span class="gp">>>> </span> <span class="n">rref1</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">remote</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">2</span><span class="p">),</span> <span class="mi">3</span><span class="p">))</span>
<span class="gp">>>> </span> <span class="n">rref2</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">remote</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">2</span><span class="p">),</span> <span class="mi">1</span><span class="p">))</span>
<span class="gp">>>> </span> <span class="n">loss</span> <span class="o">=</span> <span class="n">rref1</span><span class="o">.</span><span class="n">to_here</span><span class="p">()</span> <span class="o">+</span> <span class="n">rref2</span><span class="o">.</span><span class="n">to_here</span><span class="p">()</span>
<span class="go">>>></span>
<span class="gp">>>> </span> <span class="c1"># Backward pass.</span>
<span class="gp">>>> </span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">backward</span><span class="p">(</span><span class="n">context_id</span><span class="p">,</span> <span class="p">[</span><span class="n">loss</span><span class="o">.</span><span class="n">sum</span><span class="p">()])</span>
<span class="go">>>></span>
<span class="gp">>>> </span> <span class="c1"># Optimizer.</span>
<span class="gp">>>> </span> <span class="n">dist_optim</span> <span class="o">=</span> <span class="n">DistributedOptimizer</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="p">[</span><span class="n">rref1</span><span class="p">,</span> <span class="n">rref2</span><span class="p">],</span>
<span class="gp">>>> </span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.05</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="p">)</span>
<span class="gp">>>> </span> <span class="n">dist_optim</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">context_id</span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.DistributedOptimizer.step">
<span class="sig-name descname"><span class="pre">step</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">context_id</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/optimizer.html#DistributedOptimizer.step"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.DistributedOptimizer.step" title="Permalink to this definition">¶</a></dt>
<dd><p>Performs a single optimization step.</p>
<p>This will call <a class="reference internal" href="generated/torch.optim.Optimizer.step.html#torch.optim.Optimizer.step" title="torch.optim.Optimizer.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.optim.Optimizer.step()</span></code></a> on each worker
containing parameters to be optimized, and will block until all workers
return. The provided <code class="docutils literal notranslate"><span class="pre">context_id</span></code> will be used to retrieve the
corresponding <a class="reference internal" href="rpc.html#torch.distributed.autograd.context" title="torch.distributed.autograd.context"><code class="xref py py-class docutils literal notranslate"><span class="pre">context</span></code></a> that
contains the gradients that should be applied to the parameters.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>context_id</strong> – the autograd context id for which we should run the
optimizer step.</p>
</dd>
</dl>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.optim.PostLocalSGDOptimizer">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.optim.</span></span><span class="sig-name descname"><span class="pre">PostLocalSGDOptimizer</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">optim</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">averager</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/post_localSGD_optimizer.html#PostLocalSGDOptimizer"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.PostLocalSGDOptimizer" title="Permalink to this definition">¶</a></dt>
<dd><p>Wraps an arbitrary <a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.Optimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.optim.Optimizer</span></code></a> and runs <a class="reference external" href="https://arxiv.org/abs/1808.07217">post-local SGD</a>,
This optimizer runs local optimizer at every step.
After the warm-up stage, it averages parameters periodically afer the local optimizer is applied.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>optim</strong> (<a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.optimizer.Optimizer"><em>Optimizer</em></a>) – The local optimizer.</p></li>
<li><p><strong>averager</strong> (<em>ModelAverager</em>) – A model averager instance to run post-localSGD algorithm.</p></li>
</ul>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch</span>
<span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch.distributed</span> <span class="k">as</span> <span class="nn">dist</span>
<span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch.distributed.algorithms.model_averaging.averagers</span> <span class="k">as</span> <span class="nn">averagers</span>
<span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch.nn</span> <span class="k">as</span> <span class="nn">nn</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.optim</span> <span class="kn">import</span> <span class="n">PostLocalSGDOptimizer</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.algorithms.ddp_comm_hooks.post_localSGD_hook</span> <span class="kn">import</span> <span class="p">(</span>
<span class="gp">>>> </span> <span class="n">PostLocalSGDState</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="n">post_localSGD_hook</span><span class="p">,</span>
<span class="gp">>>> </span><span class="p">)</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="n">model</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">parallel</span><span class="o">.</span><span class="n">DistributedDataParallel</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">module</span><span class="p">,</span> <span class="n">device_ids</span><span class="o">=</span><span class="p">[</span><span class="n">rank</span><span class="p">],</span> <span class="n">output_device</span><span class="o">=</span><span class="n">rank</span>
<span class="gp">>>> </span><span class="p">)</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># Register a post-localSGD communication hook.</span>
<span class="gp">>>> </span><span class="n">state</span> <span class="o">=</span> <span class="n">PostLocalSGDState</span><span class="p">(</span><span class="n">process_group</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="n">subgroup</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="n">start_localSGD_iter</span><span class="o">=</span><span class="mi">100</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">model</span><span class="o">.</span><span class="n">register_comm_hook</span><span class="p">(</span><span class="n">state</span><span class="p">,</span> <span class="n">post_localSGD_hook</span><span class="p">)</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># Create a post-localSGD optimizer that wraps a local optimizer.</span>
<span class="gp">>>> </span><span class="c1"># Note that ``warmup_steps`` used in ``PostLocalSGDOptimizer`` must be the same as</span>
<span class="gp">>>> </span><span class="c1"># ``start_localSGD_iter`` used in ``PostLocalSGDState``.</span>
<span class="gp">>>> </span><span class="n">local_optim</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">params</span><span class="o">=</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.01</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">opt</span> <span class="o">=</span> <span class="n">PostLocalSGDOptimizer</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">optim</span><span class="o">=</span><span class="n">local_optim</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="n">averager</span><span class="o">=</span><span class="n">averagers</span><span class="o">.</span><span class="n">PeriodicModelAverager</span><span class="p">(</span><span class="n">period</span><span class="o">=</span><span class="mi">4</span><span class="p">,</span> <span class="n">warmup_steps</span><span class="o">=</span><span class="mi">100</span><span class="p">)</span>
<span class="gp">>>> </span><span class="p">)</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># In the first 100 steps, DDP runs global gradient averaging at every step.</span>
<span class="gp">>>> </span><span class="c1"># After 100 steps, DDP runs gradient averaging within each subgroup (intra-node by default),</span>
<span class="gp">>>> </span><span class="c1"># and post-localSGD optimizer runs global model averaging every 4 steps after applying the local optimizer.</span>
<span class="gp">>>> </span><span class="k">for</span> <span class="n">step</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="mi">200</span><span class="p">):</span>
<span class="gp">>>> </span> <span class="n">opt</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="gp">>>> </span> <span class="n">loss</span> <span class="o">=</span> <span class="n">loss_fn</span><span class="p">(</span><span class="n">output</span><span class="p">,</span> <span class="n">labels</span><span class="p">)</span>
<span class="gp">>>> </span> <span class="n">loss</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="gp">>>> </span> <span class="n">opt</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.PostLocalSGDOptimizer.load_state_dict">
<span class="sig-name descname"><span class="pre">load_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/post_localSGD_optimizer.html#PostLocalSGDOptimizer.load_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.PostLocalSGDOptimizer.load_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>This is the same as <a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.Optimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.optim.Optimizer</span></code></a> <a class="reference internal" href="#torch.distributed.optim.PostLocalSGDOptimizer.load_state_dict" title="torch.distributed.optim.PostLocalSGDOptimizer.load_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">load_state_dict()</span></code></a>,
but also restores model averager’s step value to the one
saved in the provided <code class="docutils literal notranslate"><span class="pre">state_dict</span></code>.</p>
<p>If there is no <code class="docutils literal notranslate"><span class="pre">"step"</span></code> entry in <code class="docutils literal notranslate"><span class="pre">state_dict</span></code>,
it will raise a warning and initialize the model averager’s step to 0.</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.PostLocalSGDOptimizer.state_dict">
<span class="sig-name descname"><span class="pre">state_dict</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/post_localSGD_optimizer.html#PostLocalSGDOptimizer.state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.PostLocalSGDOptimizer.state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>This is the same as <a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.Optimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.optim.Optimizer</span></code></a> <a class="reference internal" href="#torch.distributed.optim.PostLocalSGDOptimizer.state_dict" title="torch.distributed.optim.PostLocalSGDOptimizer.state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">state_dict()</span></code></a>,
but adds an extra entry to record model averager’s step to the checkpoint
to ensure reload does not cause unnecessary warm up again.</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.PostLocalSGDOptimizer.step">
<span class="sig-name descname"><span class="pre">step</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/post_localSGD_optimizer.html#PostLocalSGDOptimizer.step"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.PostLocalSGDOptimizer.step" title="Permalink to this definition">¶</a></dt>
<dd><p>Performs a single optimization step (parameter update).</p>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.optim.ZeroRedundancyOptimizer">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.optim.</span></span><span class="sig-name descname"><span class="pre">ZeroRedundancyOptimizer</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">params</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">optimizer_class</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">process_group</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">parameters_as_bucket_view</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">overlap_with_ddp</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">defaults</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/zero_redundancy_optimizer.html#ZeroRedundancyOptimizer"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.ZeroRedundancyOptimizer" title="Permalink to this definition">¶</a></dt>
<dd><p>This class wraps an arbitrary <a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.Optimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">optim.Optimizer</span></code></a> and shards its states across ranks in the group as
described by <a class="reference external" href="https://arxiv.org/abs/1910.02054">ZeRO</a>. The local optimizer instance in each rank is only
responsible for updating approximately <code class="docutils literal notranslate"><span class="pre">1</span> <span class="pre">/</span> <span class="pre">world_size</span></code> parameters and
hence only needs to keep <code class="docutils literal notranslate"><span class="pre">1</span> <span class="pre">/</span> <span class="pre">world_size</span></code> optimizer states. After
parameters are updated locally, each rank will broadcast its parameters to
all other peers to keep all model replicas in the same state.
<code class="docutils literal notranslate"><span class="pre">ZeroRedundancyOptimizer</span></code> can be used in conjunction with
<a class="reference internal" href="generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel" title="torch.nn.parallel.DistributedDataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.parallel.DistributedDataParallel</span></code></a> to reduce per-rank peak
memory consumption.</p>
<p><code class="docutils literal notranslate"><span class="pre">ZeroRedundancyOptimizer</span></code> uses a sorted-greedy algorithm to pack a number
of parameters at each rank. Each parameter belongs to a single rank and is
not divided among ranks. The partition is arbitrary and might not match the
the parameter registration or usage order.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>params</strong> (<code class="docutils literal notranslate"><span class="pre">Iterable</span></code>) – an <code class="docutils literal notranslate"><span class="pre">Iterable</span></code> of <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a> s
or <a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a> s giving all parameters, which will be sharded
across ranks.</p>
</dd>
<dt class="field-even">Keyword Arguments<span class="colon">:</span></dt>
<dd class="field-even"><ul class="simple">
<li><p><strong>optimizer_class</strong> (<code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.Optimizer</span></code>) – the class of the local
optimizer.</p></li>
<li><p><strong>process_group</strong> (<code class="docutils literal notranslate"><span class="pre">ProcessGroup</span></code>, optional) – <code class="docutils literal notranslate"><span class="pre">torch.distributed</span></code>
<code class="docutils literal notranslate"><span class="pre">ProcessGroup</span></code> (default: <code class="docutils literal notranslate"><span class="pre">dist.group.WORLD</span></code> initialized by
<a class="reference internal" href="distributed.html#torch.distributed.init_process_group" title="torch.distributed.init_process_group"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.distributed.init_process_group()</span></code></a>).</p></li>
<li><p><strong>parameters_as_bucket_view</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.11)"><em>bool</em></a><em>, </em><em>optional</em>) – if <code class="docutils literal notranslate"><span class="pre">True</span></code>, parameters are
packed into buckets to speed up communication, and <code class="docutils literal notranslate"><span class="pre">param.data</span></code>
fields point to bucket views at different offsets; if <code class="docutils literal notranslate"><span class="pre">False</span></code>,
each individual parameter is communicated separately, and each
<code class="docutils literal notranslate"><span class="pre">params.data</span></code> stays intact (default: <code class="docutils literal notranslate"><span class="pre">False</span></code>).</p></li>
<li><p><strong>overlap_with_ddp</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.11)"><em>bool</em></a><em>, </em><em>optional</em>) – if <code class="docutils literal notranslate"><span class="pre">True</span></code>, <a class="reference internal" href="#torch.distributed.optim.ZeroRedundancyOptimizer.step" title="torch.distributed.optim.ZeroRedundancyOptimizer.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code></a> is
overlapped with <code class="xref py py-class docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code> ‘s gradient
synchronization; this requires (1) either a functional optimizer
for the <code class="docutils literal notranslate"><span class="pre">optimizer_class</span></code> argument or one with a functional
equivalent and (2) registering a DDP communication hook
constructed from one of the functions in <code class="docutils literal notranslate"><span class="pre">ddp_zero_hook.py</span></code>;
parameters are packed into buckets matching those in
<code class="xref py py-class docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code>, meaning that the
<code class="docutils literal notranslate"><span class="pre">parameters_as_bucket_view</span></code> argument is ignored.
If <code class="docutils literal notranslate"><span class="pre">False</span></code>, <a class="reference internal" href="#torch.distributed.optim.ZeroRedundancyOptimizer.step" title="torch.distributed.optim.ZeroRedundancyOptimizer.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code></a> runs disjointly after the backward pass
(per normal).
(default: <code class="docutils literal notranslate"><span class="pre">False</span></code>)</p></li>
<li><p><strong>**defaults</strong> – any trailing arguments, which are forwarded to the local
optimizer.</p></li>
</ul>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch.nn</span> <span class="k">as</span> <span class="nn">nn</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.optim</span> <span class="kn">import</span> <span class="n">ZeroRedundancyOptimizer</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.nn.parallel</span> <span class="kn">import</span> <span class="n">DistributedDataParallel</span> <span class="k">as</span> <span class="n">DDP</span>
<span class="gp">>>> </span><span class="n">model</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Sequential</span><span class="p">(</span><span class="o">*</span><span class="p">[</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">2000</span><span class="p">,</span> <span class="mi">2000</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">rank</span><span class="p">)</span> <span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">20</span><span class="p">)])</span>
<span class="gp">>>> </span><span class="n">ddp</span> <span class="o">=</span> <span class="n">DDP</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">device_ids</span><span class="o">=</span><span class="p">[</span><span class="n">rank</span><span class="p">])</span>
<span class="gp">>>> </span><span class="n">opt</span> <span class="o">=</span> <span class="n">ZeroRedundancyOptimizer</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">ddp</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span>
<span class="gp">>>> </span> <span class="n">optimizer_class</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">optim</span><span class="o">.</span><span class="n">Adam</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.01</span>
<span class="gp">>>> </span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">ddp</span><span class="p">(</span><span class="n">inputs</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">opt</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Currently, <code class="docutils literal notranslate"><span class="pre">ZeroRedundancyOptimizer</span></code> requires that all of the
passed-in parameters are the same dense type.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>If you pass <code class="docutils literal notranslate"><span class="pre">overlap_with_ddp=True</span></code>, be wary of the following: Given
the way that overlapping <code class="xref py py-class docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code> with
<a class="reference internal" href="#torch.distributed.optim.ZeroRedundancyOptimizer" title="torch.distributed.optim.ZeroRedundancyOptimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">ZeroRedundancyOptimizer</span></code></a> is currently implemented, the first
two or three training iterations do not perform parameter updates in
the optimizer step, depending on if <code class="docutils literal notranslate"><span class="pre">static_graph=False</span></code> or
<code class="docutils literal notranslate"><span class="pre">static_graph=True</span></code>, respectively. This is because it needs
information about the gradient bucketing strategy used by
<code class="xref py py-class docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code>, which is not finalized until the
second forward pass if <code class="docutils literal notranslate"><span class="pre">static_graph=False</span></code> or until the third
forward pass if <code class="docutils literal notranslate"><span class="pre">static_graph=True</span></code>. To adjust for this, one option
is to prepend dummy inputs.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>ZeroRedundancyOptimizer is experimental and subject to change.</p>
</div>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.ZeroRedundancyOptimizer.add_param_group">
<span class="sig-name descname"><span class="pre">add_param_group</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">param_group</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/zero_redundancy_optimizer.html#ZeroRedundancyOptimizer.add_param_group"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.ZeroRedundancyOptimizer.add_param_group" title="Permalink to this definition">¶</a></dt>
<dd><p>Add a parameter group to the <code class="xref py py-class docutils literal notranslate"><span class="pre">Optimizer</span></code> ‘s <code class="docutils literal notranslate"><span class="pre">param_groups</span></code>.</p>
<p>This can be useful when fine tuning a pre-trained network, as frozen
layers can be made trainable and added to the <code class="xref py py-class docutils literal notranslate"><span class="pre">Optimizer</span></code> as
training progresses.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>param_group</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><em>dict</em></a>) – specifies the parameters to be optimized and
group-specific optimization options.</p>
</dd>
</dl>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>This method handles updating the shards on all partitions
but needs to be called on all ranks. Calling this on a subset of
the ranks will cause the training to hang because communication
primitives are called depending on the managed parameters and
expect all the ranks to participate on the same set of parameters.</p>
</div>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.ZeroRedundancyOptimizer.consolidate_state_dict">
<span class="sig-name descname"><span class="pre">consolidate_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">to</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/zero_redundancy_optimizer.html#ZeroRedundancyOptimizer.consolidate_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.ZeroRedundancyOptimizer.consolidate_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Consolidate a list of <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> s (one per rank) on the target
rank.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>to</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.11)"><em>int</em></a>) – the rank that receives the optimizer states (default: 0).</p>
</dd>
<dt class="field-even">Raises<span class="colon">:</span></dt>
<dd class="field-even"><p><a class="reference external" href="https://docs.python.org/3/library/exceptions.html#RuntimeError" title="(in Python v3.11)"><strong>RuntimeError</strong></a> – if <code class="docutils literal notranslate"><span class="pre">overlap_with_ddp=True</span></code> and this method is
called before this <a class="reference internal" href="#torch.distributed.optim.ZeroRedundancyOptimizer" title="torch.distributed.optim.ZeroRedundancyOptimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">ZeroRedundancyOptimizer</span></code></a> instance
has been fully initialized, which happens once
<code class="xref py py-class docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code> gradient buckets have been
rebuilt.</p>
</dd>
</dl>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>This needs to be called on all ranks.</p>
</div>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.ZeroRedundancyOptimizer.join_hook">
<span class="sig-name descname"><span class="pre">join_hook</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/zero_redundancy_optimizer.html#ZeroRedundancyOptimizer.join_hook"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.ZeroRedundancyOptimizer.join_hook" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns the ZeRO join hook, which enables training on uneven inputs by
shadowing the collective communications in the optimizer step.</p>
<p>Gradients must be properly set before this hook is called.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>kwargs</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><em>dict</em></a>) – a <a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a> containing any keyword arguments
to modify the behavior of the join hook at run time; all
<code class="xref py py-class docutils literal notranslate"><span class="pre">Joinable</span></code> instances sharing the same join context
manager are forwarded the same value for <code class="docutils literal notranslate"><span class="pre">kwargs</span></code>.</p>
</dd>
</dl>
<p>This hook does not support any keyword arguments; i.e. <code class="docutils literal notranslate"><span class="pre">kwargs</span></code> is
unused.</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.ZeroRedundancyOptimizer.load_state_dict">
<span class="sig-name descname"><span class="pre">load_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/zero_redundancy_optimizer.html#ZeroRedundancyOptimizer.load_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.ZeroRedundancyOptimizer.load_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Load the state pertaining to the given rank from the input
<code class="docutils literal notranslate"><span class="pre">state_dict</span></code>, updating the local optimizer as needed.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>state_dict</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><em>dict</em></a>) – optimizer state; should be an object returned
from a call to <a class="reference internal" href="#torch.distributed.optim.ZeroRedundancyOptimizer.state_dict" title="torch.distributed.optim.ZeroRedundancyOptimizer.state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">state_dict()</span></code></a>.</p>
</dd>
<dt class="field-even">Raises<span class="colon">:</span></dt>
<dd class="field-even"><p><a class="reference external" href="https://docs.python.org/3/library/exceptions.html#RuntimeError" title="(in Python v3.11)"><strong>RuntimeError</strong></a> – if <code class="docutils literal notranslate"><span class="pre">overlap_with_ddp=True</span></code> and this method is
called before this <a class="reference internal" href="#torch.distributed.optim.ZeroRedundancyOptimizer" title="torch.distributed.optim.ZeroRedundancyOptimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">ZeroRedundancyOptimizer</span></code></a> instance
has been fully initialized, which happens once
<code class="xref py py-class docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code> gradient buckets have been
rebuilt.</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.ZeroRedundancyOptimizer.state_dict">
<span class="sig-name descname"><span class="pre">state_dict</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/zero_redundancy_optimizer.html#ZeroRedundancyOptimizer.state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.ZeroRedundancyOptimizer.state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns the last global optimizer state known to this rank.</p>
<dl class="field-list simple">
<dt class="field-odd">Raises<span class="colon">:</span></dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/exceptions.html#RuntimeError" title="(in Python v3.11)"><strong>RuntimeError</strong></a> – if <code class="docutils literal notranslate"><span class="pre">overlap_with_ddp=True</span></code> and this method is
called before this <a class="reference internal" href="#torch.distributed.optim.ZeroRedundancyOptimizer" title="torch.distributed.optim.ZeroRedundancyOptimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">ZeroRedundancyOptimizer</span></code></a> instance
has been fully initialized, which happens once
<code class="xref py py-class docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code> gradient buckets have been
rebuilt; or if this method is called without a preceding call
to <a class="reference internal" href="#torch.distributed.optim.ZeroRedundancyOptimizer.consolidate_state_dict" title="torch.distributed.optim.ZeroRedundancyOptimizer.consolidate_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">consolidate_state_dict()</span></code></a>.</p>
</dd>
<dt class="field-even">Return type<span class="colon">:</span></dt>
<dd class="field-even"><p><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Dict" title="(in Python v3.11)"><em>Dict</em></a>[<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.11)">str</a>, <a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Any" title="(in Python v3.11)"><em>Any</em></a>]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.optim.ZeroRedundancyOptimizer.step">
<span class="sig-name descname"><span class="pre">step</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">closure</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="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/zero_redundancy_optimizer.html#ZeroRedundancyOptimizer.step"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.optim.ZeroRedundancyOptimizer.step" title="Permalink to this definition">¶</a></dt>
<dd><p>Performs a single optimizer step and syncs parameters across all ranks.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>closure</strong> (<em>Callable</em>) – a closure that re-evaluates the model and
returns the loss; optional for most optimizers.</p>
</dd>
<dt class="field-even">Returns<span class="colon">:</span></dt>
<dd class="field-even"><p>Optional loss depending on the underlying local optimizer.</p>
</dd>
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Optional" title="(in Python v3.11)"><em>Optional</em></a>[<a class="reference external" href="https://docs.python.org/3/library/functions.html#float" title="(in Python v3.11)">float</a>]</p>
</dd>
</dl>
</dd></dl>
</dd></dl>
</section>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="distributed.tensor.parallel.html" class="btn btn-neutral float-right" title="Tensor Parallelism" accesskey="n" rel="next">Next <img src="_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="fsdp.html" class="btn btn-neutral" title="FullyShardedDataParallel" accesskey="p" rel="prev"><img src="_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright 2022, PyTorch Contributors.
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
</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="#">Distributed Optimizers</a></li>
</ul>
</div>
</div>
</div>
</section>
</div>
<script type="text/javascript" id="documentation_options" data-url_root="./" src="_static/documentation_options.js"></script>
<script data-url_root="./" id="documentation_options" src="_static/documentation_options.js"></script>
<script src="_static/jquery.js"></script>
<script src="_static/underscore.js"></script>
<script src="_static/_sphinx_javascript_frameworks_compat.js"></script>
<script src="_static/doctools.js"></script>
<script src="_static/clipboard.min.js"></script>
<script src="_static/copybutton.js"></script>
<script type="text/javascript" src="_static/js/vendor/popper.min.js"></script>
<script type="text/javascript" src="_static/js/vendor/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/list.js/1.5.0/list.min.js"></script>
<script type="text/javascript" src="_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<script script type="text/javascript">
var collapsedSections = ['Developer Notes', 'Language Bindings', 'Libraries', 'Community'];
</script>
<img height="1" width="1" style="border-style:none;" alt="" src="https://www.googleadservices.com/pagead/conversion/795629140/?label=txkmCPmdtosBENSssfsC&guid=ON&script=0"/>
<!-- Begin Footer -->
<div class="container-fluid docs-tutorials-resources" id="docs-tutorials-resources">
<div class="container">
<div class="row">
<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">
<ul>
<li class="list-title">Stay up to date</li>
<li><a href="https://www.facebook.com/pytorch" target="_blank">Facebook</a></li>
<li><a href="https://twitter.com/pytorch" target="_blank">Twitter</a></li>
<li><a href="https://www.youtube.com/pytorch" target="_blank">YouTube</a></li>
<li><a href="https://www.linkedin.com/company/pytorch" target="_blank">LinkedIn</a></li>
</ul>
</div>