forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpipeline.html
1047 lines (846 loc) · 67.5 KB
/
pipeline.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>Pipeline Parallelism — PyTorch 1.10 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/pipeline.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="Quantization" href="quantization.html" />
<link rel="prev" title="DDP Communication Hooks" href="ddp_comm_hooks.html" />
<script src="_static/js/modernizr.min.js"></script>
<!-- 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>
<!-- 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 ▼</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"><a class="reference internal" href="ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">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>Pipeline Parallelism</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/pipeline.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="pipeline-parallelism">
<span id="id1"></span><h1>Pipeline Parallelism<a class="headerlink" href="#pipeline-parallelism" title="Permalink to this headline">¶</a></h1>
<p>Pipeline parallelism was original introduced in the
<a class="reference external" href="https://arxiv.org/abs/1811.06965">Gpipe</a> paper and is an efficient
technique to train large models on multiple GPUs.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Pipeline Parallelism is experimental and subject to change.</p>
</div>
<div class="section" id="model-parallelism-using-multiple-gpus">
<h2>Model Parallelism using multiple GPUs<a class="headerlink" href="#model-parallelism-using-multiple-gpus" title="Permalink to this headline">¶</a></h2>
<p>Typically for large models which don’t fit on a single GPU, model parallelism
is employed where certain parts of the model are placed on different GPUs.
Although, if this is done naively for sequential models, the training process
suffers from GPU under utilization since only one GPU is active at one time as
shown in the figure below:</p>
<div class="figure align-default" id="id2">
<img alt="_images/no_pipe.png" src="_images/no_pipe.png" />
<p class="caption"><span class="caption-text">The figure represents a model with 4 layers placed on 4 different GPUs
(vertical axis). The horizontal axis represents training this model through
time demonstrating that only 1 GPU is utilized at a time
(<a class="reference external" href="https://arxiv.org/abs/1811.06965">image source</a>).</span><a class="headerlink" href="#id2" title="Permalink to this image">¶</a></p>
</div>
</div>
<div class="section" id="pipelined-execution">
<h2>Pipelined Execution<a class="headerlink" href="#pipelined-execution" title="Permalink to this headline">¶</a></h2>
<p>To alleviate this problem, pipeline parallelism splits the input minibatch into
multiple microbatches and pipelines the execution of these microbatches across
multiple GPUs. This is outlined in the figure below:</p>
<div class="figure align-default" id="id3">
<img alt="_images/pipe.png" src="_images/pipe.png" />
<p class="caption"><span class="caption-text">The figure represents a model with 4 layers placed on 4 different GPUs
(vertical axis). The horizontal axis represents training this model through
time demonstrating that the GPUs are utilized much more efficiently.
However, there still exists a bubble (as demonstrated in the figure) where
certain GPUs are not utilized.
(<a class="reference external" href="https://arxiv.org/abs/1811.06965">image source</a>).</span><a class="headerlink" href="#id3" title="Permalink to this image">¶</a></p>
</div>
</div>
<div class="section" id="pipe-apis-in-pytorch">
<h2>Pipe APIs in PyTorch<a class="headerlink" href="#pipe-apis-in-pytorch" title="Permalink to this headline">¶</a></h2>
<dl class="py class">
<dt id="torch.distributed.pipeline.sync.Pipe">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.pipeline.sync.</span></code><code class="sig-name descname"><span class="pre">Pipe</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">module</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">chunks</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">checkpoint</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">'except_last'</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">deferred_batch_norm</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/pipeline/sync/pipe.html#Pipe"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.pipeline.sync.Pipe" title="Permalink to this definition">¶</a></dt>
<dd><p>Wraps an arbitrary <a class="reference internal" href="generated/torch.nn.Sequential.html#torch.nn.Sequential" title="torch.nn.Sequential"><code class="xref py py-class docutils literal notranslate"><span class="pre">nn.Sequential</span></code></a> module
to train on using synchronous pipeline parallelism. If the module requires
lots of memory and doesn’t fit on a single GPU, pipeline parallelism is a
useful technique to employ for training.</p>
<p>The implementation is based on the <a class="reference external" href="https://arxiv.org/abs/2004.09910">torchgpipe</a> paper.</p>
<p>Pipe combines pipeline parallelism with checkpointing to reduce peak
memory required to train while minimizing device under-utilization.</p>
<p>You should place all the modules on the appropriate devices and wrap them
into an <a class="reference internal" href="generated/torch.nn.Sequential.html#torch.nn.Sequential" title="torch.nn.Sequential"><code class="xref py py-class docutils literal notranslate"><span class="pre">nn.Sequential</span></code></a> module defining the
desired order of execution. If a module does not contain any
parameters/buffers, it is assumed this module should be executed on CPU
and appropriate input tensors to the module are moved to CPU before
execution. This behavior can be overridden by the <code class="xref py py-class docutils literal notranslate"><span class="pre">WithDevice</span></code>
wrapper which can be used to explicitly specify which device a module
should run on.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>module</strong> (<a class="reference internal" href="generated/torch.nn.Sequential.html#torch.nn.Sequential" title="torch.nn.Sequential"><code class="xref py py-class docutils literal notranslate"><span class="pre">nn.Sequential</span></code></a>) – sequential module to be parallelized using pipelining. Each module
in the sequence has to have all of its parameters on a single
device. Each module in the sequence has to either be an nn.Module
or <a class="reference internal" href="generated/torch.nn.Sequential.html#torch.nn.Sequential" title="torch.nn.Sequential"><code class="xref py py-class docutils literal notranslate"><span class="pre">nn.Sequential</span></code></a> (to combine multiple
sequential modules on a single device)</p></li>
<li><p><strong>chunks</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a>) – number of micro-batches (default: <code class="docutils literal notranslate"><span class="pre">1</span></code>)</p></li>
<li><p><strong>checkpoint</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.10)"><em>str</em></a>) – when to enable checkpointing, one of <code class="docutils literal notranslate"><span class="pre">'always'</span></code>,
<code class="docutils literal notranslate"><span class="pre">'except_last'</span></code>, or <code class="docutils literal notranslate"><span class="pre">'never'</span></code> (default: <code class="docutils literal notranslate"><span class="pre">'except_last'</span></code>).
<code class="docutils literal notranslate"><span class="pre">'never'</span></code> disables checkpointing completely, <code class="docutils literal notranslate"><span class="pre">'except_last'</span></code>
enables checkpointing for all micro-batches except the last one
and <code class="docutils literal notranslate"><span class="pre">'always'</span></code> enables checkpointing for all micro-batches.</p></li>
<li><p><strong>deferred_batch_norm</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.10)"><em>bool</em></a>) – whether to use deferred <code class="docutils literal notranslate"><span class="pre">BatchNorm</span></code> moving statistics (default:
<a class="reference external" href="https://docs.python.org/3/library/constants.html#False" title="(in Python v3.10)"><code class="xref py py-data docutils literal notranslate"><span class="pre">False</span></code></a>). If set to <a class="reference external" href="https://docs.python.org/3/library/constants.html#True" title="(in Python v3.10)"><code class="xref py py-data docutils literal notranslate"><span class="pre">True</span></code></a>, we track statistics across
multiple micro-batches to update the running statistics per
mini-batch.</p></li>
</ul>
</dd>
<dt class="field-even">Raises</dt>
<dd class="field-even"><ul class="simple">
<li><p><a class="reference external" href="https://docs.python.org/3/library/exceptions.html#TypeError" title="(in Python v3.10)"><strong>TypeError</strong></a> – the module is not a <a class="reference internal" href="generated/torch.nn.Sequential.html#torch.nn.Sequential" title="torch.nn.Sequential"><code class="xref py py-class docutils literal notranslate"><span class="pre">nn.Sequential</span></code></a>.</p></li>
<li><p><a class="reference external" href="https://docs.python.org/3/library/exceptions.html#ValueError" title="(in Python v3.10)"><strong>ValueError</strong></a> – invalid arguments</p></li>
</ul>
</dd>
</dl>
<dl>
<dt>Example::</dt><dd><p>Pipeline of two FC layers across GPUs 0 and 1.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="c1"># Need to initialize RPC framework first.</span>
<span class="gp">>>> </span><span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'MASTER_ADDR'</span><span class="p">]</span> <span class="o">=</span> <span class="s1">'localhost'</span>
<span class="gp">>>> </span><span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'MASTER_PORT'</span><span class="p">]</span> <span class="o">=</span> <span class="s1">'29500'</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">distributed</span><span class="o">.</span><span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="s1">'worker'</span><span class="p">,</span> <span class="n">rank</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span> <span class="n">world_size</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># Build pipe.</span>
<span class="gp">>>> </span><span class="n">fc1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">16</span><span class="p">,</span> <span class="mi">8</span><span class="p">)</span><span class="o">.</span><span class="n">cuda</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">fc2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">8</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span><span class="o">.</span><span class="n">cuda</span><span class="p">(</span><span class="mi">1</span><span class="p">)</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="n">fc1</span><span class="p">,</span> <span class="n">fc2</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">model</span> <span class="o">=</span> <span class="n">Pipe</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">chunks</span><span class="o">=</span><span class="mi">8</span><span class="p">)</span>
<span class="gp">>>> </span><span class="nb">input</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">16</span><span class="p">,</span> <span class="mi">16</span><span class="p">)</span><span class="o">.</span><span class="n">cuda</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">output_rref</span> <span class="o">=</span> <span class="n">model</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>You can wrap a <a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a> model 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> only when the
checkpoint parameter of <a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a> is <code class="docutils literal notranslate"><span class="pre">'never'</span></code>.</p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p><a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a> only supports intra-node pipelining currently, but
will be expanded to support inter-node pipelining in the future.
The forward function returns an <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>
to allow for inter-node pipelining in the future, where the output
might be on a remote host. For intra-node pipelinining you can use
<a class="reference internal" href="rpc.html#torch.distributed.rpc.RRef.local_value" title="torch.distributed.rpc.RRef.local_value"><code class="xref py py-meth docutils literal notranslate"><span class="pre">local_value()</span></code></a> to retrieve the
output locally.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p><a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a> is experimental and subject to change.</p>
</div>
<dl class="py method">
<dt id="torch.distributed.pipeline.sync.Pipe.forward">
<code class="sig-name descname"><span class="pre">forward</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">inputs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/pipeline/sync/pipe.html#Pipe.forward"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.pipeline.sync.Pipe.forward" title="Permalink to this definition">¶</a></dt>
<dd><p>Processes a single input mini-batch through the pipe and returns an
<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> pointing to the output.
<a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a> is a fairly transparent module wrapper. It doesn’t
modify the input and output signature of the underlying module. But
there’s type restriction. Input and output have to contain at least one
tensor. This restriction is applied at partition boundaries too.</p>
<p>The sequence of inputs are fed into the first stage of the pipeline as
<code class="docutils literal notranslate"><span class="pre">*inputs</span></code>. As a result the positional args for this function should
match the positional args for the first stage of the pipeline. The same
condition applies for output of one stage of the pipeline which is the
input for the next stage.</p>
<p>The input tensor is split into multiple micro-batches based on the
<code class="docutils literal notranslate"><span class="pre">chunks</span></code> parameter used to initialize <a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a>. The batch size
is assumed to be the first dimension of the tensor and if the batch
size is less than <code class="docutils literal notranslate"><span class="pre">chunks</span></code>, the number of micro-batches is equal to
the batch size.</p>
<p>Only tensors are split into multiple micro-batches, non-Tensor inputs
are just replicated as-is in each micro-batch. For non-Tensor outputs
in the last stage of the pipeline, they are aggregated as a <code class="docutils literal notranslate"><span class="pre">List</span></code>
and returned the user. For example, if you have 2 micro-batches
returning the integer 5, the user would receive the consolidated
output of <cite>[5, 5]</cite></p>
<p>All the input tensors need to be on the same device as the first
partition of the pipeline.</p>
<p>If a tensor is wrapped with the <code class="xref py py-class docutils literal notranslate"><span class="pre">NoChunk</span></code> wrapper, the tensor
is not split across micro-batches and is replicated as-is similar to
non-tensors.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>inputs</strong> – input mini-batch</p>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p><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> to the output of the mini-batch</p>
</dd>
<dt class="field-odd">Raises</dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/exceptions.html#TypeError" title="(in Python v3.10)"><strong>TypeError</strong></a> – input doesn’t contain at least one tensor</p>
</dd>
</dl>
</dd></dl>
</dd></dl>
<div class="section" id="skip-connections">
<h3>Skip connections<a class="headerlink" href="#skip-connections" title="Permalink to this headline">¶</a></h3>
<p>Certain models like ResNeXt are not completely sequential and have skip
connections between layers. Naively implementing as part of pipeling
parallelism would imply that we need to copy outputs for certain layers through
multiple GPUs till we eventually reach the GPU where the layer for the skip
connection resides. To avoid this copy overhead, we provide APIs below to stash
and pop Tensors in different layers of the model.</p>
<dl class="py function">
<dt id="torch.distributed.pipeline.sync.skip.skippable.skippable">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.pipeline.sync.skip.skippable.</span></code><code class="sig-name descname"><span class="pre">skippable</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">stash</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">()</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">pop</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">()</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/pipeline/sync/skip/skippable.html#skippable"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.pipeline.sync.skip.skippable.skippable" title="Permalink to this definition">¶</a></dt>
<dd><p>The decorator to define a <a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><code class="xref py py-class docutils literal notranslate"><span class="pre">nn.Module</span></code></a> with skip
connections. Decorated modules are called “skippable”. This functionality
works perfectly fine even when the module is not wrapped by
<a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a>.</p>
<p>Each skip tensor is managed by its name. Before manipulating skip tensors,
a skippable module must statically declare the names for skip tensors by
<cite>stash</cite> and/or <cite>pop</cite> parameters. Skip tensors with pre-declared name can be
stashed by <code class="docutils literal notranslate"><span class="pre">yield</span> <span class="pre">stash(name,</span> <span class="pre">tensor)</span></code> or popped by <code class="docutils literal notranslate"><span class="pre">tensor</span> <span class="pre">=</span> <span class="pre">yield</span>
<span class="pre">pop(name)</span></code>.</p>
<p>Here is an example with three layers. A skip tensor named “1to3” is stashed
and popped at the first and last layer, respectively:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="nd">@skippable</span><span class="p">(</span><span class="n">stash</span><span class="o">=</span><span class="p">[</span><span class="s1">'1to3'</span><span class="p">])</span>
<span class="k">class</span> <span class="nc">Layer1</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="k">yield</span> <span class="n">stash</span><span class="p">(</span><span class="s1">'1to3'</span><span class="p">,</span> <span class="nb">input</span><span class="p">)</span>
<span class="k">return</span> <span class="n">f1</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="k">class</span> <span class="nc">Layer2</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="k">return</span> <span class="n">f2</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="nd">@skippable</span><span class="p">(</span><span class="n">pop</span><span class="o">=</span><span class="p">[</span><span class="s1">'1to3'</span><span class="p">])</span>
<span class="k">class</span> <span class="nc">Layer3</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="n">skip_1to3</span> <span class="o">=</span> <span class="k">yield</span> <span class="n">pop</span><span class="p">(</span><span class="s1">'1to3'</span><span class="p">)</span>
<span class="k">return</span> <span class="n">f3</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span> <span class="o">+</span> <span class="n">skip_1to3</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="n">Layer1</span><span class="p">(),</span> <span class="n">Layer2</span><span class="p">(),</span> <span class="n">Layer3</span><span class="p">())</span>
</pre></div>
</div>
<p>One skippable module can stash or pop multiple skip tensors:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="nd">@skippable</span><span class="p">(</span><span class="n">stash</span><span class="o">=</span><span class="p">[</span><span class="s1">'alice'</span><span class="p">,</span> <span class="s1">'bob'</span><span class="p">],</span> <span class="n">pop</span><span class="o">=</span><span class="p">[</span><span class="s1">'carol'</span><span class="p">])</span>
<span class="k">class</span> <span class="nc">StashStashPop</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="k">yield</span> <span class="n">stash</span><span class="p">(</span><span class="s1">'alice'</span><span class="p">,</span> <span class="n">f_alice</span><span class="p">(</span><span class="nb">input</span><span class="p">))</span>
<span class="k">yield</span> <span class="n">stash</span><span class="p">(</span><span class="s1">'bob'</span><span class="p">,</span> <span class="n">f_bob</span><span class="p">(</span><span class="nb">input</span><span class="p">))</span>
<span class="n">carol</span> <span class="o">=</span> <span class="k">yield</span> <span class="n">pop</span><span class="p">(</span><span class="s1">'carol'</span><span class="p">)</span>
<span class="k">return</span> <span class="nb">input</span> <span class="o">+</span> <span class="n">carol</span>
</pre></div>
</div>
<p>Every skip tensor must be associated with exactly one pair of <cite>stash</cite> and
<cite>pop</cite>. <a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a> checks this
restriction automatically when wrapping a module. You can also check the
restriction by <a class="reference internal" href="#torch.distributed.pipeline.sync.skip.skippable.verify_skippables" title="torch.distributed.pipeline.sync.skip.skippable.verify_skippables"><code class="xref py py-func docutils literal notranslate"><span class="pre">verify_skippables()</span></code></a>
without <a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a>.</p>
</dd></dl>
<dl class="py class">
<dt id="torch.distributed.pipeline.sync.skip.skippable.stash">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.pipeline.sync.skip.skippable.</span></code><code class="sig-name descname"><span class="pre">stash</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">name</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">tensor</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/pipeline/sync/skip/skippable.html#stash"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.pipeline.sync.skip.skippable.stash" title="Permalink to this definition">¶</a></dt>
<dd><p>The command to stash a skip tensor.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="k">yield</span> <span class="n">stash</span><span class="p">(</span><span class="s1">'name'</span><span class="p">,</span> <span class="nb">input</span><span class="p">)</span>
<span class="k">return</span> <span class="n">f</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
</pre></div>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>name</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.10)"><em>str</em></a>) – name of skip tensor</p></li>
<li><p><strong>input</strong> (<a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>torch.Tensor</em></a><em> or </em><a class="reference external" href="https://docs.python.org/3/library/constants.html#None" title="(in Python v3.10)"><em>None</em></a>) – tensor to pass to the skip connection</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="py class">
<dt id="torch.distributed.pipeline.sync.skip.skippable.pop">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.distributed.pipeline.sync.skip.skippable.</span></code><code class="sig-name descname"><span class="pre">pop</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">name</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/pipeline/sync/skip/skippable.html#pop"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.pipeline.sync.skip.skippable.pop" title="Permalink to this definition">¶</a></dt>
<dd><p>The command to pop a skip tensor.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="n">skip</span> <span class="o">=</span> <span class="k">yield</span> <span class="n">pop</span><span class="p">(</span><span class="s1">'name'</span><span class="p">)</span>
<span class="k">return</span> <span class="n">f</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span> <span class="o">+</span> <span class="n">skip</span>
</pre></div>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>name</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.10)"><em>str</em></a>) – name of skip tensor</p>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>the skip tensor previously stashed by another layer under the same name</p>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt id="torch.distributed.pipeline.sync.skip.skippable.verify_skippables">
<code class="sig-prename descclassname"><span class="pre">torch.distributed.pipeline.sync.skip.skippable.</span></code><code class="sig-name descname"><span class="pre">verify_skippables</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">module</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/pipeline/sync/skip/skippable.html#verify_skippables"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.pipeline.sync.skip.skippable.verify_skippables" title="Permalink to this definition">¶</a></dt>
<dd><p>Verifies if the underlying skippable modules satisfy integrity.</p>
<p>Every skip tensor must have only one pair of <cite>stash</cite> and <cite>pop</cite>. If there
are one or more unmatched pairs, it will raise <a class="reference external" href="https://docs.python.org/3/library/exceptions.html#TypeError" title="(in Python v3.10)"><code class="xref py py-exc docutils literal notranslate"><span class="pre">TypeError</span></code></a> with the
detailed messages.</p>
<p>Here are a few failure cases. <a class="reference internal" href="#torch.distributed.pipeline.sync.skip.skippable.verify_skippables" title="torch.distributed.pipeline.sync.skip.skippable.verify_skippables"><code class="xref py py-func docutils literal notranslate"><span class="pre">verify_skippables()</span></code></a> will report failure
for these cases:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Layer1 stashes "1to3".</span>
<span class="c1"># Layer3 pops "1to3".</span>
<span class="n">nn</span><span class="o">.</span><span class="n">Sequential</span><span class="p">(</span><span class="n">Layer1</span><span class="p">(),</span> <span class="n">Layer2</span><span class="p">())</span>
<span class="c1"># └──── ?</span>
<span class="n">nn</span><span class="o">.</span><span class="n">Sequential</span><span class="p">(</span><span class="n">Layer2</span><span class="p">(),</span> <span class="n">Layer3</span><span class="p">())</span>
<span class="c1"># ? ────┘</span>
<span class="n">nn</span><span class="o">.</span><span class="n">Sequential</span><span class="p">(</span><span class="n">Layer1</span><span class="p">(),</span> <span class="n">Layer2</span><span class="p">(),</span> <span class="n">Layer3</span><span class="p">(),</span> <span class="n">Layer3</span><span class="p">())</span>
<span class="c1"># └───────────────────┘ ^^^^^^</span>
<span class="n">nn</span><span class="o">.</span><span class="n">Sequential</span><span class="p">(</span><span class="n">Layer1</span><span class="p">(),</span> <span class="n">Layer1</span><span class="p">(),</span> <span class="n">Layer2</span><span class="p">(),</span> <span class="n">Layer3</span><span class="p">())</span>
<span class="c1"># ^^^^^^ └───────────────────┘</span>
</pre></div>
</div>
<p>To use the same name for multiple skip tensors, they must be isolated by
different namespaces. See <code class="xref py py-meth docutils literal notranslate"><span class="pre">isolate()</span></code>.</p>
<dl class="field-list simple">
<dt class="field-odd">Raises</dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/exceptions.html#TypeError" title="(in Python v3.10)"><strong>TypeError</strong></a> – one or more pairs of <cite>stash</cite> and <cite>pop</cite> are not matched.</p>
</dd>
</dl>
</dd></dl>
</div>
</div>
<div class="section" id="tutorials">
<h2>Tutorials<a class="headerlink" href="#tutorials" title="Permalink to this headline">¶</a></h2>
<p>The following tutorials give a good overview of how to use the
<a class="reference internal" href="#torch.distributed.pipeline.sync.Pipe" title="torch.distributed.pipeline.sync.Pipe"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipe</span></code></a> API to train your models with the
rest of the components that PyTorch provides:</p>
<ul class="simple">
<li><p><a class="reference external" href="https://pytorch.org/tutorials/intermediate/pipeline_tutorial.html">Training Transformer models using Pipeline Parallelism</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/tutorials/advanced/ddp_pipeline.html">Training Transformer models using Distributed Data Parallel and Pipeline Parallelism</a></p></li>
</ul>
</div>
<div class="section" id="acknowledgements">
<h2>Acknowledgements<a class="headerlink" href="#acknowledgements" title="Permalink to this headline">¶</a></h2>
<p>The implementation for pipeline parallelism is based on <a class="reference external" href="https://github.com/facebookresearch/fairscale/tree/master/fairscale/nn/pipe">fairscale’s pipe implementation</a> and
<a class="reference external" href="https://github.com/kakaobrain/torchgpipe">torchgpipe</a>. We would like to
thank both teams for their contributions and guidance towards bringing pipeline
parallelism into PyTorch.</p>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="quantization.html" class="btn btn-neutral float-right" title="Quantization" accesskey="n" rel="next">Next <img src="_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="ddp_comm_hooks.html" class="btn btn-neutral" title="DDP Communication Hooks" 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="#">Pipeline Parallelism</a><ul>
<li><a class="reference internal" href="#model-parallelism-using-multiple-gpus">Model Parallelism using multiple GPUs</a></li>
<li><a class="reference internal" href="#pipelined-execution">Pipelined Execution</a></li>
<li><a class="reference internal" href="#pipe-apis-in-pytorch">Pipe APIs in PyTorch</a><ul>
<li><a class="reference internal" href="#skip-connections">Skip connections</a></li>
</ul>
</li>
<li><a class="reference internal" href="#tutorials">Tutorials</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">
<p class="gdpr-notice">To analyze traffic and optimize your experience, we serve cookies on this site. By clicking or navigating, you agree to allow our usage of cookies. As the current maintainers of this site, Facebook’s Cookies Policy applies. Learn more, including about available controls: <a href="https://www.facebook.com/policies/cookies/">Cookies Policy</a>.</p>
<img class="close-button" src="_static/images/pytorch-x.svg">
</div>
</div>
<!-- End Footer -->
<!-- Begin Mobile Menu -->
<div class="mobile-main-menu">
<div class="container-fluid">
<div class="container">
<div class="mobile-main-menu-header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<a class="main-menu-close-button" href="#" data-behavior="close-mobile-menu"></a>
</div>
</div>
</div>
<div class="mobile-main-menu-links-container">
<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/hub">PyTorch Hub</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="resources-mobile-menu-title" class="active">
Docs
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/docs/stable/index.html">PyTorch</a>
</li>
<li>
<a href="https://pytorch.org/audio/stable/index.html">torchaudio</a>
</li>
<li>
<a href="https://pytorch.org/text/stable/index.html">torchtext</a>
</li>
<li>
<a href="https://pytorch.org/vision/stable/index.html">torchvision</a>
</li>
<li>
<a href="https://pytorch.org/elastic/">TorchElastic</a>
</li>
<li>
<a href="https://pytorch.org/serve/">TorchServe</a>
</li>
<li>
<a href="https://pytorch.org/xla">PyTorch on XLA Devices</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
Resources
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/resources">Developer Resources</a>
</li>
<li>