forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathddp_comm_hooks.html
1137 lines (931 loc) · 76.4 KB
/
ddp_comm_hooks.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>DDP Communication Hooks — PyTorch 1.10.1 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/ddp_comm_hooks.html"/>
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="_static/css/jit.css" type="text/css" />
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="Pipeline Parallelism" href="pipeline.html" />
<link rel="prev" title="Complex Numbers" href="complex_numbers.html" />
<!-- Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-117752657-2"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-117752657-2');
</script>
<!-- End Google Analytics -->
<script src="_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<a href="https://pytorch.org/mobile">Mobile</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="active docs-active">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-orange-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/audio/stable/index.html">
<span class="dropdown-title">torchaudio</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/text/stable/index.html">
<span class="dropdown-title">torchtext</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/vision/stable/index.html">
<span class="dropdown-title">torchvision</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/elastic/">
<span class="dropdown-title">TorchElastic</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/serve/">
<span class="dropdown-title">TorchServe</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/xla">
<span class="dropdown-title">PyTorch on XLA Devices</span>
<p></p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-arrow">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class="dropdown-title">About</span>
<p>Learn about PyTorch’s features and capabilities</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class="dropdown-title">Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered.</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class="dropdown-title">Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class="dropdown-title">Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/hub">
<span class="dropdown-title">Models (Beta)</span>
<p>Discover, publish, and reuse pre-trained models</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">GitHub</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>1.10.1 ▼</a>
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption"><span class="caption-text">Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="notes/amp_examples.html">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
</ul>
<p class="caption"><span class="caption-text">Python API</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="amp.html">torch.cuda.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="pipeline.html">Pipeline Parallelism</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="__config__.html">torch.__config__</a></li>
</ul>
<p class="caption"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="http://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</ul>
<p class="caption"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/governance.html">PyTorch Governance</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/persons_of_interest.html">PyTorch Governance | Persons of Interest</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="index.html">
Docs
</a> >
</li>
<li>DDP Communication Hooks</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/ddp_comm_hooks.rst.txt" rel="nofollow"><img src="_static/images/view-page-source-icon.svg"></a>
</li>
</ul>
</div>
</div>
<div class="pytorch-shortcuts-wrapper" id="pytorch-shortcuts-wrapper">
Shortcuts
</div>
</div>
<section data-toggle="wy-nav-shift" id="pytorch-content-wrap" class="pytorch-content-wrap">
<div class="pytorch-content-left">
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<div class="section" id="ddp-communication-hooks">
<h1>DDP Communication Hooks<a class="headerlink" href="#ddp-communication-hooks" title="Permalink to this headline">¶</a></h1>
<p>DDP communication hook is a generic interface to control how to communicate
gradients across workers by overriding the vanilla allreduce in
<a class="reference external" href="https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel.">DistributedDataParallel</a>.
A few built-in communication hooks are provided,
and users can easily apply any of these hooks to optimize communication.
Besides, the hook interface can also support user-defined communication
strategies for more advanced use cases.</p>
<div class="section" id="how-to-use-a-communication-hook">
<h2>How to Use a Communication Hook?<a class="headerlink" href="#how-to-use-a-communication-hook" title="Permalink to this headline">¶</a></h2>
<p>To use a communication hook, the user just needs to let the DDP model register
the hook before the training loop as below.</p>
<p><a class="reference internal" href="generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel.register_comm_hook" title="torch.nn.parallel.DistributedDataParallel.register_comm_hook"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.nn.parallel.DistributedDataParallel.register_comm_hook()</span></code></a></p>
</div>
<div class="section" id="what-does-a-communication-hook-operate-on">
<h2>What Does a Communication Hook Operate On?<a class="headerlink" href="#what-does-a-communication-hook-operate-on" title="Permalink to this headline">¶</a></h2>
<p>Communication hook provides a flexible way to allreduce gradients.
Therefore, it mainly operates on the gradients on each replica before allreduce,
which are bucketized to increase the overlap between communication and computation.
Particularly, <a class="reference internal" href="#torch.distributed.GradBucket" title="torch.distributed.GradBucket"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.distributed.GradBucket</span></code></a> represents a bucket of gradient tensors to be allreduced.</p>
<dl class="py class">
<dt id="torch.distributed.GradBucket">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.</span></code><code class="sig-name descname"><span class="pre">GradBucket</span></code><a class="headerlink" href="#torch.distributed.GradBucket" title="Permalink to this definition">¶</a></dt>
<dd><p>This class mainly passes a flattened gradient tensor
(returned by <a class="reference internal" href="#torch.distributed.GradBucket.buffer" title="torch.distributed.GradBucket.buffer"><code class="xref py py-meth docutils literal notranslate"><span class="pre">buffer()</span></code></a>)
to DDP communication hook.
This tensor can be further decomposed into a list of per-parameter tensors within this bucket
(returned by <code class="xref py py-meth docutils literal notranslate"><span class="pre">get_per_parameter_tensors()</span></code>)
to apply layer-wise operations.</p>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.GradBucket.index">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.GradBucket.</span></code><code class="sig-name descname"><span class="pre">index</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">self</span></span><span class="p"><span class="pre">:</span></span> <span class="n"><span class="pre">torch._C._distributed_c10d.GradBucket</span></span></em><span class="sig-paren">)</span> → <a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><span class="pre">int</span></a><a class="headerlink" href="#torch.distributed.GradBucket.index" title="Permalink to this definition">¶</a></dt>
<dd><div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Since the buckets are rebuilt after the first iteration, should not rely on the indices at the beginning of training.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>The index of a bucket that stores gradients of a few contiguous layers.
All the gradients are bucketized.</p>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.GradBucket.buffer">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.GradBucket.</span></code><code class="sig-name descname"><span class="pre">buffer</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">self</span></span><span class="p"><span class="pre">:</span></span> <span class="n"><span class="pre">torch._C._distributed_c10d.GradBucket</span></span></em><span class="sig-paren">)</span> → <span class="pre">at::Tensor</span><a class="headerlink" href="#torch.distributed.GradBucket.buffer" title="Permalink to this definition">¶</a></dt>
<dd><dl class="field-list simple">
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>A flattened 1D <code class="docutils literal notranslate"><span class="pre">torch.Tensor</span></code> buffer,
which can be further decomposed into a list of per-parameter tensors within this bucket.</p>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.GradBucket.gradients">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.GradBucket.</span></code><code class="sig-name descname"><span class="pre">gradients</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">self</span></span><span class="p"><span class="pre">:</span></span> <span class="n"><span class="pre">torch._C._distributed_c10d.GradBucket</span></span></em><span class="sig-paren">)</span> → <span class="pre">List[at::Tensor]</span><a class="headerlink" href="#torch.distributed.GradBucket.gradients" title="Permalink to this definition">¶</a></dt>
<dd><dl class="field-list simple">
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>A list of <code class="docutils literal notranslate"><span class="pre">torch.Tensor</span></code>. Each tensor in the list corresponds to a gradient.</p>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.GradBucket.is_last">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.GradBucket.</span></code><code class="sig-name descname"><span class="pre">is_last</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">self</span></span><span class="p"><span class="pre">:</span></span> <span class="n"><span class="pre">torch._C._distributed_c10d.GradBucket</span></span></em><span class="sig-paren">)</span> → <a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.10)"><span class="pre">bool</span></a><a class="headerlink" href="#torch.distributed.GradBucket.is_last" title="Permalink to this definition">¶</a></dt>
<dd><dl class="field-list simple">
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>Whether this bucket is the last bucket to allreduce in an iteration.
This also means that this bucket corresponds to the first few layers in the forward pass.</p>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.GradBucket.set_buffer">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.GradBucket.</span></code><code class="sig-name descname"><span class="pre">set_buffer</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="pre">self:</span> <span class="pre">torch._C._distributed_c10d.GradBucket</span></em>, <em class="sig-param"><span class="pre">buffer:</span> <span class="pre">at::Tensor</span></em><span class="sig-paren">)</span> → <a class="reference external" href="https://docs.python.org/3/library/constants.html#None" title="(in Python v3.10)"><span class="pre">None</span></a><a class="headerlink" href="#torch.distributed.GradBucket.set_buffer" title="Permalink to this definition">¶</a></dt>
<dd><p>Replaces the tensor in the bucket with the input tensor buffer.</p>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.GradBucket.parameters">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.GradBucket.</span></code><code class="sig-name descname"><span class="pre">parameters</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">self</span></span><span class="p"><span class="pre">:</span></span> <span class="n"><span class="pre">torch._C._distributed_c10d.GradBucket</span></span></em><span class="sig-paren">)</span> → <span class="pre">List[at::Tensor]</span><a class="headerlink" href="#torch.distributed.GradBucket.parameters" title="Permalink to this definition">¶</a></dt>
<dd><dl class="field-list simple">
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>A list of <code class="docutils literal notranslate"><span class="pre">torch.Tensor</span></code>. Each tensor in the list corresponds to a model
parameter.</p>
</dd>
</dl>
</dd></dl>
</div>
<div class="section" id="default-communication-hooks">
<h2>Default Communication Hooks<a class="headerlink" href="#default-communication-hooks" title="Permalink to this headline">¶</a></h2>
<p>Default communication hooks are simple <strong>stateless</strong> hooks, so the input state
in <code class="docutils literal notranslate"><span class="pre">register_comm_hook</span></code> is either a process group or <code class="docutils literal notranslate"><span class="pre">None</span></code>.
The input <code class="docutils literal notranslate"><span class="pre">bucket</span></code> is a <a class="reference internal" href="#torch.distributed.GradBucket" title="torch.distributed.GradBucket"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.distributed.GradBucket</span></code></a> object.</p>
<dl class="py function">
<dt id="torch.distributed.algorithms.ddp_comm_hooks.default_hooks.allreduce_hook">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.algorithms.ddp_comm_hooks.default_hooks.</span></code><code class="sig-name descname"><span class="pre">allreduce_hook</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">process_group</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">bucket</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/algorithms/ddp_comm_hooks/default_hooks.html#allreduce_hook"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.allreduce_hook" title="Permalink to this definition">¶</a></dt>
<dd><p>This DDP communication hook just calls <code class="docutils literal notranslate"><span class="pre">allreduce</span></code> using <code class="docutils literal notranslate"><span class="pre">GradBucket</span></code>
tensors. Once gradient tensors are aggregated across all workers, its <code class="docutils literal notranslate"><span class="pre">then</span></code>
callback takes the mean and returns the result. If user registers this hook,
DDP results is expected to be same as the case where no hook was registered.
Hence, this won’t change behavior of DDP and user can use this as a reference
or modify this hook to log useful information or any other purposes while
unaffecting DDP behavior.</p>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">ddp_model</span><span class="o">.</span><span class="n">register_comm_hook</span><span class="p">(</span><span class="n">process_group</span><span class="p">,</span> <span class="n">allreduce_hook</span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.algorithms.ddp_comm_hooks.default_hooks.fp16_compress_hook">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.algorithms.ddp_comm_hooks.default_hooks.</span></code><code class="sig-name descname"><span class="pre">fp16_compress_hook</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">process_group</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">bucket</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/algorithms/ddp_comm_hooks/default_hooks.html#fp16_compress_hook"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.fp16_compress_hook" title="Permalink to this definition">¶</a></dt>
<dd><p>This DDP communication hook implements a simple gradient compression
approach that casts <code class="docutils literal notranslate"><span class="pre">GradBucket</span></code> tensor to half-precision floating-point format (<code class="docutils literal notranslate"><span class="pre">torch.float16</span></code>)
and then divides it by the process group size.
It allreduces those <code class="docutils literal notranslate"><span class="pre">float16</span></code> gradient tensors. Once compressed gradient
tensors are allreduced, the chained callback <code class="docutils literal notranslate"><span class="pre">decompress</span></code> casts it back to the input data type (such as <code class="docutils literal notranslate"><span class="pre">float32</span></code>).</p>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">ddp_model</span><span class="o">.</span><span class="n">register_comm_hook</span><span class="p">(</span><span class="n">process_group</span><span class="p">,</span> <span class="n">fp16_compress_hook</span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.algorithms.ddp_comm_hooks.default_hooks.bf16_compress_hook">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.algorithms.ddp_comm_hooks.default_hooks.</span></code><code class="sig-name descname"><span class="pre">bf16_compress_hook</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">process_group</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">bucket</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/algorithms/ddp_comm_hooks/default_hooks.html#bf16_compress_hook"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.bf16_compress_hook" title="Permalink to this definition">¶</a></dt>
<dd><p>Warning: This API is experimental, and it requires NCCL version later than 2.9.6.</p>
<p>This DDP communication hook implements a simple gradient compression
approach that casts <code class="docutils literal notranslate"><span class="pre">GradBucket</span></code> tensor to half-precision
<a class="reference external" href="https://en.wikipedia.org/wiki/Bfloat16_floating-point_format">Brain floating point format</a> (<code class="docutils literal notranslate"><span class="pre">torch.bfloat16</span></code>)
and then divides it by the process group size.
It allreduces those <code class="docutils literal notranslate"><span class="pre">bfloat16</span></code> gradient tensors. Once compressed gradient
tensors are allreduced, the chained callback <code class="docutils literal notranslate"><span class="pre">decompress</span></code> casts it back to the input data type (such as <code class="docutils literal notranslate"><span class="pre">float32</span></code>).</p>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">ddp_model</span><span class="o">.</span><span class="n">register_comm_hook</span><span class="p">(</span><span class="n">process_group</span><span class="p">,</span> <span class="n">bf16_compress_hook</span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
<p>Additionally, a communication hook wraper is provided to support <a class="reference internal" href="#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.fp16_compress_hook" title="torch.distributed.algorithms.ddp_comm_hooks.default_hooks.fp16_compress_hook"><code class="xref py py-meth docutils literal notranslate"><span class="pre">fp16_compress_hook()</span></code></a> or <a class="reference internal" href="#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.bf16_compress_hook" title="torch.distributed.algorithms.ddp_comm_hooks.default_hooks.bf16_compress_hook"><code class="xref py py-meth docutils literal notranslate"><span class="pre">bf16_compress_hook()</span></code></a> as a wrapper,
which can be combined with other communication hooks.</p>
<dl class="py function">
<dt id="torch.distributed.algorithms.ddp_comm_hooks.default_hooks.fp16_compress_wrapper">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.algorithms.ddp_comm_hooks.default_hooks.</span></code><code class="sig-name descname"><span class="pre">fp16_compress_wrapper</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">hook</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/algorithms/ddp_comm_hooks/default_hooks.html#fp16_compress_wrapper"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.fp16_compress_wrapper" title="Permalink to this definition">¶</a></dt>
<dd><p>This wrapper casts the input gradient tensor of a given DDP communication hook to half-precision
floating point format (<code class="docutils literal notranslate"><span class="pre">torch.float16</span></code>), and casts the resulting tensor of the given hook back to
the input data type, such as <code class="docutils literal notranslate"><span class="pre">float32</span></code>.</p>
<p>Therefore, <code class="docutils literal notranslate"><span class="pre">fp16_compress_hook</span></code> is equivalent to <code class="docutils literal notranslate"><span class="pre">fp16_compress_wrapper(allreduce_hook)</span></code>.</p>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">state</span> <span class="o">=</span> <span class="n">PowerSGDState</span><span class="p">(</span><span class="n">process_group</span><span class="o">=</span><span class="n">process_group</span><span class="p">,</span> <span class="n">matrix_approximation_rank</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">start_powerSGD_iter</span><span class="o">=</span><span class="mi">10</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">ddp_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">fp16_compress_wrapper</span><span class="p">(</span><span class="n">powerSGD_hook</span><span class="p">))</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.algorithms.ddp_comm_hooks.default_hooks.bf16_compress_wrapper">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.algorithms.ddp_comm_hooks.default_hooks.</span></code><code class="sig-name descname"><span class="pre">bf16_compress_wrapper</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">hook</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/algorithms/ddp_comm_hooks/default_hooks.html#bf16_compress_wrapper"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.bf16_compress_wrapper" title="Permalink to this definition">¶</a></dt>
<dd><p>Warning: This API is experimental, and it requires NCCL version later than 2.9.6.</p>
<p>This wrapper casts the input gradient tensor of a given DDP communication hook to half-precision
<cite>Brain floating point format <https://en.wikipedia.org/wiki/Bfloat16_floating-point_format> `_ (``torch.bfloat16`</cite>),
and casts the resulting tensor of the given hook back to the input data type, such as <code class="docutils literal notranslate"><span class="pre">float32</span></code>.</p>
<p>Therefore, <code class="docutils literal notranslate"><span class="pre">bf16_compress_hook</span></code> is equivalent to <code class="docutils literal notranslate"><span class="pre">bf16_compress_wrapper(allreduce_hook)</span></code>.</p>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">state</span> <span class="o">=</span> <span class="n">PowerSGDState</span><span class="p">(</span><span class="n">process_group</span><span class="o">=</span><span class="n">process_group</span><span class="p">,</span> <span class="n">matrix_approximation_rank</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">start_powerSGD_iter</span><span class="o">=</span><span class="mi">10</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">ddp_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">bf16_compress_wrapper</span><span class="p">(</span><span class="n">powerSGD_hook</span><span class="p">))</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
</div>
<div class="section" id="powersgd-communication-hook">
<h2>PowerSGD Communication Hook<a class="headerlink" href="#powersgd-communication-hook" title="Permalink to this headline">¶</a></h2>
<p>PowerSGD (<a class="reference external" href="https://arxiv.org/abs/1905.13727">Vogels et al., NeurIPS 2019</a>)
is a gradient compression algorithm, which can provide very high compression
rates and accelerate bandwidth-bound distributed training.
This algorithm needs to maintain both some hyperparameters and the internal
state. Therefore, PowerSGD communication hook is a <strong>stateful</strong> hook,
and the user needs to provide a state object defined as below.</p>
<div class="section" id="powersgd-state">
<h3>PowerSGD State<a class="headerlink" href="#powersgd-state" title="Permalink to this headline">¶</a></h3>
<dl class="py class">
<dt id="torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.PowerSGDState">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.</span></code><code class="sig-name descname"><span class="pre">PowerSGDState</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">process_group</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">matrix_approximation_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">start_powerSGD_iter</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">1000</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">min_compression_rate</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">2</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">use_error_feedback</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">warm_start</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">orthogonalization_epsilon</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">random_seed</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">compression_stats_logging_frequency</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">10000</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/algorithms/ddp_comm_hooks/powerSGD_hook.html#PowerSGDState"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.PowerSGDState" title="Permalink to this definition">¶</a></dt>
<dd><p>Stores both the algorithm’s hyperparameters and the internal state for all the gradients during the training.
Particularly, <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> and <code class="docutils literal notranslate"><span class="pre">start_powerSGD_iter</span></code> are the main hyperparameters that should be tuned by the user.
For performance, we suggest to keep binary hyperparameters <code class="docutils literal notranslate"><span class="pre">use_error_feedback</span></code> and <code class="docutils literal notranslate"><span class="pre">warm_start</span></code> on.</p>
<ol class="arabic">
<li><p><code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> controls the size of compressed low-rank tensors, which determines the compression rate. The lower the rank, the stronger the compression.</p>
<blockquote>
<div><p>1.1. If <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> is too low, the full model quality will need more training steps to reach or will never reach and yield loss in accuracy.</p>
<p>1.2. The increase of <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> can substantially increase the computation costs of the compression, and the accuracy may not be futher improved beyond a certain <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> threshold.</p>
</div></blockquote>
</li>
</ol>
<p>To tune <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code>, we suggest to start from 1 and increase by factors of 2 (like an expoential grid search, 1, 2, 4, …), until a satisfactory accuracy is reached. Typically only a small value 1-4 is used. For some NLP tasks (as shown in Appendix D of the original paper), this value has been increased to 32.</p>
<ol class="arabic simple" start="2">
<li><p><code class="docutils literal notranslate"><span class="pre">start_powerSGD_iter</span></code> defers PowerSGD compression until step <code class="docutils literal notranslate"><span class="pre">start_powerSGD_iter</span></code>, and vanilla allreduce runs prior to step <code class="docutils literal notranslate"><span class="pre">start_powerSGD_iter</span></code>. This hybrid scheme of <strong>vanilla allreduce + PowerSGD</strong> can effectively improve the accuracy, even a relatively small <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> is used. This is because that, the beginning of training phase is usually very sensitive to inaccurate gradients, and compressing gradients too early may make the training quickly take a suboptimal trajectory, which can result in an irrecoverable impact on the accuracy.</p></li>
</ol>
<p>To tune <code class="docutils literal notranslate"><span class="pre">start_powerSGD_iter</span></code>, we suggest to start with 10% of total training steps, and increase it until a satisfactory accuracy is reached. If there is a warm-up stage in the training, <code class="docutils literal notranslate"><span class="pre">start_powerSGD_iter</span></code> typically should be no less than the number of warm-up steps.</p>
<ol class="arabic simple" start="3">
<li><p><code class="docutils literal notranslate"><span class="pre">min_compression_rate</span></code> is the minimum compression rate required when a layer is compressed. Due to the computation overheads incurred by the compression, a tensor is worth compressing only if there can be sufficient saving in bandwidth, where <code class="docutils literal notranslate"><span class="pre">(num_rows</span> <span class="pre">+</span> <span class="pre">num_cols)</span> <span class="pre">*</span> <span class="pre">matrix_approximation_rank</span> <span class="pre">*</span> <span class="pre">min_compression_rate</span> <span class="pre"><</span> <span class="pre">num_rows</span> <span class="pre">*</span> <span class="pre">num_cols</span></code>. If the specified compression rate threshold cannot be satisfied, the tensor will be directly allreduced without compression.</p></li>
</ol>
<p>Compression statistics are logged every <code class="docutils literal notranslate"><span class="pre">compression_stats_logging_frequency</span></code> iterations once PowerSGD compression starts.</p>
<ol class="arabic simple" start="4">
<li><p><code class="docutils literal notranslate"><span class="pre">orthogonalization_epsilon</span></code> can be a very small value (e.g., 1e-8) added to every normalized matrix column in orthogonalization step, to prevent div-by-zero error if any column has all 0s. If this can already be prevented (e.g., by batch normalization), an epsilon of 0 is recommended for accuracy.</p></li>
</ol>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>If error feedback or warm-up is enabled, the minimum value of <code class="docutils literal notranslate"><span class="pre">start_powerSGD_iter</span></code> allowed in DDP is 2.
This is because there is another internal optimization that rebuilds buckets at iteration 1 in DDP,
and this can conflict with any tensor memorized before the rebuild process.</p>
</div>
</dd></dl>
</div>
<div class="section" id="powersgd-hooks">
<h3>PowerSGD Hooks<a class="headerlink" href="#powersgd-hooks" title="Permalink to this headline">¶</a></h3>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>PowerSGD typically requires extra memory of the same size as the model’s
gradients to enable error feedback, which can compensate for biased
compressed communication and improve accuracy.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>PowerSGD hooks may conflict with <a class="reference external" href="https://github.com/NVIDIA/apex">Apex automatic mixed precision package</a>.
Please use PyTorch <a class="reference external" href="https://pytorch.org/docs/stable/amp.html">native automatic mixed precision package</a>
instead.</p>
</div>
<dl class="py function">
<dt id="torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.powerSGD_hook">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.</span></code><code class="sig-name descname"><span class="pre">powerSGD_hook</span></code><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">bucket</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/algorithms/ddp_comm_hooks/powerSGD_hook.html#powerSGD_hook"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.powerSGD_hook" title="Permalink to this definition">¶</a></dt>
<dd><p>This DDP communication hook implements PowerSGD gradient compression
algorithm described in the <a class="reference external" href="https://arxiv.org/abs/1905.13727">paper</a>.
Once gradient tensors are aggregated across all workers, this hook applies
compression as follows:</p>
<ol class="arabic">
<li><p>Views the input flattened 1D gradient tensor as a list of per-parameter tensors, and divides all the tensors into two groups:</p>
<blockquote>
<div><p>1.1 The tensors that should be compressed before allreduce, because the compression can give enough saving in bandwidth.</p>
<p>1.2 Rest of the tensors will be directly allreduced without compression, including all the vector tensors (for biases).</p>
</div></blockquote>
</li>
<li><p>Handles uncompressed tensors:</p>
<blockquote>
<div><p>2.1. Allocate contiguous memory for those uncompressed tensors, and allreduces all the uncompressed tensors as a batch, without compression;</p>
<p>2.2. Copies the individual uncompressed tensors from the contiguous memory back to the input tensor.</p>
</div></blockquote>
</li>
<li><p>Handles the tensors that should be compressed by PowerSGD compression:</p>
<blockquote>
<div><p>3.1. For each tensor M, creates two low-rank tensors P and Q for decomposing M,
such that M = PQ^T, where Q is initialized from a standard normal distribution and orthogonalized;</p>
<p>3.2. Computes each P in Ps, which is equal to MQ;</p>
<p>3.3. Allreduces Ps as a batch;</p>
<p>3.4. Orthogonalizes each P in Ps;</p>
<p>3.5. Computes each Q in Qs, which is approximately equal to M^TP;</p>
<p>3.6. Allreduces Qs as a batch;</p>
<p>3.7. Computes each M among all the compressed tensors, which is approximately equal to PQ^T.</p>
</div></blockquote>
</li>
</ol>
<p>Note that this communication hook enforces vanilla allreduce for the first <code class="docutils literal notranslate"><span class="pre">state.start_powerSGD_iter</span></code> iterations.
This not only gives the user more control over the tradeoff between speedup and accuracy,
but also helps abstract away some complexity of the internal optimization of DDP for future communication hook developers.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>state</strong> (<a class="reference internal" href="#torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.PowerSGDState" title="torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.PowerSGDState"><em>PowerSGDState</em></a>) – State information to configure the compression rate and support error feedback, warm start, etc.
To tune the compression configs, mainly need to tune <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code>, <code class="docutils literal notranslate"><span class="pre">start_powerSGD_iter</span></code>
and <code class="docutils literal notranslate"><span class="pre">min_compression_rate</span></code>.</p></li>
<li><p><strong>bucket</strong> (<em>dist.GradBucket</em>) – Bucket that stores a 1D flattened gradient tensor that batches multiple per-variable tensors.
Note that since DDP comm hook only supports single process single device mode,
only exactly one tensor is stored in this bucket.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>Future handler of the communication, which updates the gradients in place.</p>
</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="n">state</span> <span class="o">=</span> <span class="n">PowerSGDState</span><span class="p">(</span><span class="n">process_group</span><span class="o">=</span><span class="n">process_group</span><span class="p">,</span> <span class="n">matrix_approximation_rank</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
<span class="go"> start_powerSGD_iter=10, min_compression_rate=0.5)</span>
<span class="gp">>>> </span><span class="n">ddp_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">powerSGD_hook</span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.batched_powerSGD_hook">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.</span></code><code class="sig-name descname"><span class="pre">batched_powerSGD_hook</span></code><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">bucket</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/algorithms/ddp_comm_hooks/powerSGD_hook.html#batched_powerSGD_hook"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.batched_powerSGD_hook" title="Permalink to this definition">¶</a></dt>
<dd><p>This DDP communication hook implements a simplified PowerSGD gradient compression
algorithm described in the <a class="reference external" href="https://arxiv.org/abs/1905.13727">paper</a>.
This variant does not compress the gradients layer by layer,
but instead compresses the flattened input tensor that batches all the gradients.
Therefore, it is <strong>faster</strong> than <a class="reference internal" href="#torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.powerSGD_hook" title="torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.powerSGD_hook"><code class="xref py py-meth docutils literal notranslate"><span class="pre">powerSGD_hook()</span></code></a>,
but usually results in a <strong>much lower accuracy</strong>, unless <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> is 1.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Increasing <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> here may not necessarily increase the accuracy,
because batching per-parameter tensors without column/row alignment can destroy low-rank structure.
Therefore, the user should always consider <a class="reference internal" href="#torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.powerSGD_hook" title="torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.powerSGD_hook"><code class="xref py py-meth docutils literal notranslate"><span class="pre">powerSGD_hook()</span></code></a> first,
and only consider this variant when a satisfactory accuracy can be achieved when <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> is 1.</p>
</div>
<p>Once gradient tensors are aggregated across all workers, this hook applies
compression as follows:</p>
<ol class="arabic simple">
<li><p>Views the input flattened 1D gradient tensor as a square-shaped tensor M with 0 paddings;</p></li>
<li><p>Creates two low-rank tensors P and Q for decomposing M, such that M = PQ^T, where Q is initialized from a standard normal distribution and orthogonalized;</p></li>
<li><p>Computes P, which is equal to MQ;</p></li>
<li><p>Allreduces P;</p></li>
<li><p>Orthogonalizes P;</p></li>
<li><p>Computes Q, which is approximately equal to M^TP;</p></li>
<li><p>Allreduces Q;</p></li>
<li><p>Computes M, which is approximately equal to PQ^T.</p></li>
<li><p>Truncates the input tensor to the original length.</p></li>
</ol>
<p>Note that this communication hook enforces vanilla allreduce for the first <code class="docutils literal notranslate"><span class="pre">state.start_powerSGD_iter</span></code> iterations.
This not only gives the user more control over the tradeoff between speedup and accuracy,
but also helps abstract away some complexity of the internal optimization of DDP for future communication hook developers.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>state</strong> (<a class="reference internal" href="#torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.PowerSGDState" title="torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook.PowerSGDState"><em>PowerSGDState</em></a>) – State information to configure the compression rate and support error feedback, warm start, etc.
To tune the compression configs, mainly need to tune <code class="docutils literal notranslate"><span class="pre">matrix_approximation_rank</span></code> and <code class="docutils literal notranslate"><span class="pre">start_powerSGD_iter</span></code>.</p></li>
<li><p><strong>bucket</strong> (<em>dist.GradBucket</em>) – Bucket that stores a 1D flattened gradient tensor that batches multiple per-variable tensors.
Note that since DDP comm hook only supports single process single device mode,
only exactly one tensor is stored in this bucket.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>Future handler of the communication, which updates the gradients in place.</p>
</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="n">state</span> <span class="o">=</span> <span class="n">PowerSGDState</span><span class="p">(</span><span class="n">process_group</span><span class="o">=</span><span class="n">process_group</span><span class="p">,</span> <span class="n">matrix_approximation_rank</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">ddp_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">batched_powerSGD_hook</span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
</div>
</div>
<div class="section" id="debugging-communication-hooks">
<h2>Debugging Communication Hooks<a class="headerlink" href="#debugging-communication-hooks" title="Permalink to this headline">¶</a></h2>
<p>As the name implies, debugging communication hooks are <strong>only</strong> used for debugging and performance optimization purpose.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Debugging communication hooks do not necessarily output the correct results.</p>
</div>
<dl class="py function">
<dt id="torch.distributed.algorithms.ddp_comm_hooks.debugging_hooks.noop_hook">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.algorithms.ddp_comm_hooks.debugging_hooks.</span></code><code class="sig-name descname"><span class="pre">noop_hook</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">_</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">bucket</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/algorithms/ddp_comm_hooks/debugging_hooks.html#noop_hook"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.algorithms.ddp_comm_hooks.debugging_hooks.noop_hook" title="Permalink to this definition">¶</a></dt>
<dd><p>This DDP communication hook returns the a future that wraps the input,
so it is a noop that does not incur any communication overheads.</p>
<p>This hook should <strong>only</strong> be used for headroom analysis of allreduce optimization,
instead of the normal gradient synchronization.
For example, if only less than 10% speedup of training time can be observed after this hook is registered,
it usually implies that allreduce is not a performance bottleneck for this case.
Such instrumentation can be particularly useful
if GPU traces cannot be easily retrieved or the trace analysis is complicated
some factors such as the overlap between allreduce and computation or the desynchronization across ranks.</p>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">ddp_model</span><span class="o">.</span><span class="n">register_comm_hook</span><span class="p">(</span><span class="kc">None</span><span class="p">,</span> <span class="n">noop_hook</span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
</div>
<div class="section" id="acknowledgements">
<h2>Acknowledgements<a class="headerlink" href="#acknowledgements" title="Permalink to this headline">¶</a></h2>
<p>Many thanks to PowerSGD paper author <strong>Thijs Vogels</strong> for the code review on
PowerSGD communication hook, as well as the
<a class="reference external" href="https://observablehq.com/@tvogels/powersgd-benchmark">comparison experiments</a>,
which show that the performance of PowerSGD communication hook is on par with
the implementation in the original <a class="reference external" href="https://arxiv.org/abs/1905.13727">paper</a>.</p>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="pipeline.html" class="btn btn-neutral float-right" title="Pipeline Parallelism" accesskey="n" rel="next">Next <img src="_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="complex_numbers.html" class="btn btn-neutral" title="Complex Numbers" accesskey="p" rel="prev"><img src="_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright 2019, Torch Contributors.
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
</div>
<div class="pytorch-content-right" id="pytorch-content-right">
<div class="pytorch-right-menu" id="pytorch-right-menu">
<div class="pytorch-side-scroll" id="pytorch-side-scroll-right">
<ul>
<li><a class="reference internal" href="#">DDP Communication Hooks</a><ul>
<li><a class="reference internal" href="#how-to-use-a-communication-hook">How to Use a Communication Hook?</a></li>
<li><a class="reference internal" href="#what-does-a-communication-hook-operate-on">What Does a Communication Hook Operate On?</a></li>
<li><a class="reference internal" href="#default-communication-hooks">Default Communication Hooks</a></li>
<li><a class="reference internal" href="#powersgd-communication-hook">PowerSGD Communication Hook</a><ul>
<li><a class="reference internal" href="#powersgd-state">PowerSGD State</a></li>
<li><a class="reference internal" href="#powersgd-hooks">PowerSGD Hooks</a></li>
</ul>
</li>
<li><a class="reference internal" href="#debugging-communication-hooks">Debugging Communication Hooks</a></li>
<li><a class="reference internal" href="#acknowledgements">Acknowledgements</a></li>
</ul>
</li>
</ul>
</div>
</div>
</div>
</section>
</div>
<script type="text/javascript" id="documentation_options" data-url_root="./" src="_static/documentation_options.js"></script>
<script src="_static/jquery.js"></script>
<script src="_static/underscore.js"></script>
<script src="_static/doctools.js"></script>
<script src="_static/clipboard.min.js"></script>
<script src="_static/copybutton.js"></script>
<script type="text/javascript" src="_static/js/vendor/popper.min.js"></script>
<script type="text/javascript" src="_static/js/vendor/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/list.js/1.5.0/list.min.js"></script>
<script type="text/javascript" src="_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<script script type="text/javascript">
var collapsedSections = ['Notes', 'Language Bindings', 'Libraries', 'Community'];
</script>
<img height="1" width="1" style="border-style:none;" alt="" src="https://www.googleadservices.com/pagead/conversion/795629140/?label=txkmCPmdtosBENSssfsC&guid=ON&script=0"/>
<!-- Begin Footer -->
<div class="container-fluid docs-tutorials-resources" id="docs-tutorials-resources">
<div class="container">
<div class="row">
<div class="col-md-4 text-center">
<h2>Docs</h2>
<p>Access comprehensive developer documentation for PyTorch</p>
<a class="with-right-arrow" href="https://pytorch.org/docs/stable/index.html">View Docs</a>
</div>
<div class="col-md-4 text-center">
<h2>Tutorials</h2>
<p>Get in-depth tutorials for beginners and advanced developers</p>
<a class="with-right-arrow" href="https://pytorch.org/tutorials">View Tutorials</a>
</div>
<div class="col-md-4 text-center">
<h2>Resources</h2>
<p>Find development resources and get your questions answered</p>
<a class="with-right-arrow" href="https://pytorch.org/resources">View Resources</a>
</div>
</div>
</div>
</div>
<footer class="site-footer">
<div class="container footer-container">
<div class="footer-logo-wrapper">
<a href="https://pytorch.org/" class="footer-logo"></a>
</div>
<div class="footer-links-wrapper">
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/">PyTorch</a></li>
<li><a href="https://pytorch.org/get-started">Get Started</a></li>
<li><a href="https://pytorch.org/features">Features</a></li>
<li><a href="https://pytorch.org/ecosystem">Ecosystem</a></li>
<li><a href="https://pytorch.org/blog/">Blog</a></li>
<li><a href="https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md">Contributing</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/resources">Resources</a></li>
<li><a href="https://pytorch.org/tutorials">Tutorials</a></li>
<li><a href="https://pytorch.org/docs/stable/index.html">Docs</a></li>
<li><a href="https://discuss.pytorch.org" target="_blank">Discuss</a></li>
<li><a href="https://github.com/pytorch/pytorch/issues" target="_blank">Github Issues</a></li>
<li><a href="https://pytorch.org/assets/brand-guidelines/PyTorch-Brand-Guidelines.pdf" target="_blank">Brand Guidelines</a></li>
</ul>
</div>
<div class="footer-links-col follow-us-col">
<ul>
<li class="list-title">Stay Connected</li>
<li>
<div id="mc_embed_signup">
<form
action="https://twitter.us14.list-manage.com/subscribe/post?u=75419c71fe0a935e53dfa4a3f&id=91d0dccd39"
method="post"
id="mc-embedded-subscribe-form"
name="mc-embedded-subscribe-form"
class="email-subscribe-form validate"
target="_blank"
novalidate>
<div id="mc_embed_signup_scroll" class="email-subscribe-form-fields-wrapper">
<div class="mc-field-group">
<label for="mce-EMAIL" style="display:none;">Email Address</label>
<input type="email" value="" name="EMAIL" class="required email" id="mce-EMAIL" placeholder="Email Address">
</div>
<div id="mce-responses" class="clear">
<div class="response" id="mce-error-response" style="display:none"></div>
<div class="response" id="mce-success-response" style="display:none"></div>
</div> <!-- real people should not fill this in and expect good things - do not remove this or risk form bot signups-->
<div style="position: absolute; left: -5000px;" aria-hidden="true"><input type="text" name="b_75419c71fe0a935e53dfa4a3f_91d0dccd39" tabindex="-1" value=""></div>
<div class="clear">
<input type="submit" value="" name="subscribe" id="mc-embedded-subscribe" class="button email-subscribe-button">
</div>
</div>
</form>
</div>
</li>
</ul>
<div class="footer-social-icons">
<a href="https://www.facebook.com/pytorch" target="_blank" class="facebook"></a>
<a href="https://twitter.com/pytorch" target="_blank" class="twitter"></a>
<a href="https://www.youtube.com/pytorch" target="_blank" class="youtube"></a>
</div>
</div>
</div>
</div>
</footer>
<div class="cookie-banner-wrapper">
<div class="container">