forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrpc.html
1075 lines (867 loc) · 75.4 KB
/
rpc.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Distributed RPC Framework — PyTorch master documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/rpc.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="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="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="torch.random" href="random.html" />
<link rel="prev" title="Quantization" href="quantization.html" />
<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">
</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>
<div class="ecosystem-dropdown">
<a id="dropdownMenuButton" data-toggle="ecosystem-dropdown">
Ecosystem
</a>
<div class="ecosystem-dropdown-menu">
<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>
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem">
<span class=dropdown-title>Tools & Libraries</span>
<p>Explore the ecosystem of tools and libraries</p>
</a>
</div>
</div>
</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">
<a href="https://pytorch.org/docs/stable/index.html">Docs</a>
</li>
<li>
<div class="resources-dropdown">
<a id="resourcesDropdownButton" data-toggle="resources-dropdown">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/resources"">
<span class=dropdown-title>Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class=dropdown-title>About</span>
<p>Learn about PyTorch’s features and capabilities</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='http://pytorch.org/docs/versions.html'>1.4.0 ▼</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/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/distributed_autograd.html">Distributed Autograd Design</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/large_scale_deployments.html">Features for large-scale deployments</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/rref.html">Remote Reference Protocol</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>
<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">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/cppdocs/">C++ API</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="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="distributed.html">torch.distributed</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="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="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="quantization.html">Quantization</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">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="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="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">torchvision Reference</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchvision/index.html">torchvision</a></li>
</ul>
<p class="caption"><span class="caption-text">torchaudio Reference</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio">torchaudio</a></li>
</ul>
<p class="caption"><span class="caption-text">torchtext Reference</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text">torchtext</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>Distributed RPC Framework</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/rpc.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="distributed-rpc-framework">
<span id="id1"></span><h1>Distributed RPC Framework<a class="headerlink" href="#distributed-rpc-framework" title="Permalink to this headline">¶</a></h1>
<p>The distributed RPC framework provides mechanisms for multi-machine model
training through a set of primitives to allow for remote communication, and a
higher-level API to automatically differentiate models split across several
machines.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>The RPC API is experimental and subject to change.</p>
</div>
<div class="section" id="rpc-and-rref-framework">
<h2>RPC and RRef Framework<a class="headerlink" href="#rpc-and-rref-framework" title="Permalink to this headline">¶</a></h2>
<p>Before using RPC and distributed autograd primitives, initialization must take
place. To initialize the RPC framework we need to use
<a class="reference internal" href="#torch.distributed.rpc.init_rpc" title="torch.distributed.rpc.init_rpc"><code class="xref py py-meth docutils literal notranslate"><span class="pre">init_rpc()</span></code></a> which would initialize the RPC
framework, RRef framework and distributed autograd. By default, this will also
initialize the <cite>ProcessGroup</cite> (<a class="reference internal" href="distributed.html#torch.distributed.init_process_group" title="torch.distributed.init_process_group"><code class="xref py py-meth docutils literal notranslate"><span class="pre">init_process_group()</span></code></a>)
backend for RPC communication. The <cite>ProcessGroup</cite> backend internally uses gloo
for communication.</p>
<span class="target" id="module-torch.distributed.rpc"></span><dl class="function">
<dt id="torch.distributed.rpc.init_rpc">
<code class="sig-prename descclassname">torch.distributed.rpc.</code><code class="sig-name descname">init_rpc</code><span class="sig-paren">(</span><em class="sig-param">name</em>, <em class="sig-param">backend=BackendType.PROCESS_GROUP</em>, <em class="sig-param">rank=-1</em>, <em class="sig-param">world_size=None</em>, <em class="sig-param">rpc_backend_options=None</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/rpc.html#init_rpc"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.distributed.rpc.init_rpc" title="Permalink to this definition">¶</a></dt>
<dd><p>Initializes RPC primitives such as the local RPC agent
and distributed autograd.</p>
<p>Initializes the local RPC agent which immediately makes the current
process ready to send and receive RPCs. This method also properly
initializes a default process group backend that uses gloo for
collective communication.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>backend</strong> (<em>Enum</em>) – type of RPC backend implementation.
Currently, process group backend is the only
available backend implementation. (default:
<code class="docutils literal notranslate"><span class="pre">RpcBackend.PROCESS_GROUP</span></code>).</p></li>
<li><p><strong>name</strong> (<em>str</em>) – a globally unique name of this node. (e.g.,
<code class="docutils literal notranslate"><span class="pre">Trainer3</span></code>, <code class="docutils literal notranslate"><span class="pre">ParameterServer2</span></code>, <code class="docutils literal notranslate"><span class="pre">Master</span></code>,
<code class="docutils literal notranslate"><span class="pre">Worker1</span></code>) Name can only contain number, alphabet,
underscore, and/or dash, and must be shorter than
128 characters.</p></li>
<li><p><strong>rank</strong> (<em>python:int</em>) – a globally unique id/rank of this node.</p></li>
<li><p><strong>world_size</strong> (<em>python:int</em>) – The number of workers in the group.</p></li>
<li><p><strong>rpc_backend_options</strong> (<em>RpcBackendOptions</em>) – The options passed to RpcAgent
consturctor.</p></li>
</ul>
</dd>
</dl>
</dd></dl>
</div>
<div class="section" id="rref">
<span id="id2"></span><h2>RRef<a class="headerlink" href="#rref" title="Permalink to this headline">¶</a></h2>
<p>An <cite>RRef</cite> (Remote REFerence) is a reference to a value of some type <cite>T</cite>
(e.g. <cite>Tensor</cite>) on a remote worker. This handle keeps the referenced remote
value alive on the owner, but there is no implication that the value will be
transferred to the local worker in the future. RRefs can be used in
multi-machine training by holding references to <a class="reference external" href="https://pytorch.org/docs/stable/nn.html#torch.nn.Module">nn.Modules</a> that exist on
other workers, and calling the appropriate functions to retrieve or modify their
parameters during training. See <a class="reference internal" href="notes/rref.html#remote-reference-protocol"><span class="std std-ref">Remote Reference Protocol</span></a> for more
details.</p>
<dl class="class">
<dt id="torch.distributed.rpc.RRef">
<em class="property">class </em><code class="sig-prename descclassname">torch.distributed.rpc.</code><code class="sig-name descname">RRef</code><a class="headerlink" href="#torch.distributed.rpc.RRef" title="Permalink to this definition">¶</a></dt>
<dd><p>A class encapsulating a reference to a value of some type on a remote worker.
This handle will keep the referenced remote value alive on the worker.</p>
<dl class="method">
<dt id="torch.distributed.rpc.RRef.is_owner">
<code class="sig-name descname">is_owner</code><span class="sig-paren">(</span><em class="sig-param">self: torch.distributed.rpc.RRef</em><span class="sig-paren">)</span> → bool<a class="headerlink" href="#torch.distributed.rpc.RRef.is_owner" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns whether or not the current node is the owner of this <code class="docutils literal notranslate"><span class="pre">RRef</span></code>.</p>
</dd></dl>
<dl class="method">
<dt id="torch.distributed.rpc.RRef.local_value">
<code class="sig-name descname">local_value</code><span class="sig-paren">(</span><em class="sig-param">self: torch.distributed.rpc.RRef</em><span class="sig-paren">)</span> → object<a class="headerlink" href="#torch.distributed.rpc.RRef.local_value" title="Permalink to this definition">¶</a></dt>
<dd><p>If the current node is the owner, returns a reference to the local value.
Otherwise, throws an exception.</p>
</dd></dl>
<dl class="method">
<dt id="torch.distributed.rpc.RRef.owner">
<code class="sig-name descname">owner</code><span class="sig-paren">(</span><em class="sig-param">self: torch.distributed.rpc.RRef</em><span class="sig-paren">)</span> → torch.distributed.rpc.WorkerInfo<a class="headerlink" href="#torch.distributed.rpc.RRef.owner" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns worker information of the node that owns this <code class="docutils literal notranslate"><span class="pre">RRef</span></code>.</p>
</dd></dl>
<dl class="method">
<dt id="torch.distributed.rpc.RRef.to_here">
<code class="sig-name descname">to_here</code><span class="sig-paren">(</span><em class="sig-param">self: torch.distributed.rpc.RRef</em><span class="sig-paren">)</span> → object<a class="headerlink" href="#torch.distributed.rpc.RRef.to_here" title="Permalink to this definition">¶</a></dt>
<dd><p>Blocking call that copies the value of the RRef from the owner to the local node
and returns it. If the current node is the owner, returns a reference to the
local value.</p>
</dd></dl>
</dd></dl>
</div>
<div class="section" id="rpc-and-rref-primitives">
<h2>RPC and RRef primitives<a class="headerlink" href="#rpc-and-rref-primitives" title="Permalink to this headline">¶</a></h2>
<p>This library provides primitives allowing users to create and modify references
(RRefs) to remote data as well as remotely execute functions.</p>
<span class="target" id="module-torch.distributed.rpc"></span><dl class="function">
<dt id="torch.distributed.rpc.rpc_sync">
<code class="sig-prename descclassname">torch.distributed.rpc.</code><code class="sig-name descname">rpc_sync</code><span class="sig-paren">(</span><em class="sig-param">to</em>, <em class="sig-param">func</em>, <em class="sig-param">args=None</em>, <em class="sig-param">kwargs=None</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/rpc/api.html#rpc_sync"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.distributed.rpc.rpc_sync" title="Permalink to this definition">¶</a></dt>
<dd><p>Make a blocking RPC call to run function <code class="docutils literal notranslate"><span class="pre">func</span></code> on worker <code class="docutils literal notranslate"><span class="pre">to</span></code>. RPC
messages are sent and received in parallel to execution of Python code. This
method is thread-safe.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>to</strong> (<em>str</em><em> or </em><em>WorkerInfo</em>) – id or name of the destination worker.</p></li>
<li><p><strong>func</strong> (<em>callable</em>) – any callable function. builtin functions (like
<a class="reference internal" href="torch.html#torch.add" title="torch.add"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.add()</span></code></a>) can be sent over RPC more efficiently.</p></li>
<li><p><strong>args</strong> (<em>tuple</em>) – the argument tuple for the <code class="docutils literal notranslate"><span class="pre">func</span></code> invocation.</p></li>
<li><p><strong>kwargs</strong> (<em>dict</em>) – is a dictionary of keyword arguments for the <code class="docutils literal notranslate"><span class="pre">func</span></code>
invocation.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>Returns the result of running <code class="docutils literal notranslate"><span class="pre">func</span></code> on <code class="docutils literal notranslate"><span class="pre">args</span></code> and <code class="docutils literal notranslate"><span class="pre">kwargs</span></code>.</p>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">On</span> <span class="n">worker</span> <span class="mi">0</span><span class="p">:</span>
<span class="o">>>></span> <span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="s2">"worker0"</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">2</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">ret</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">rpc_sync</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">2</span><span class="p">),</span> <span class="mi">3</span><span class="p">))</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
<span class="n">On</span> <span class="n">worker</span> <span class="mi">1</span><span class="p">:</span>
<span class="o">>>></span> <span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">rank</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">world_size</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
</pre></div>
</div>
</dd></dl>
<dl class="function">
<dt id="torch.distributed.rpc.rpc_async">
<code class="sig-prename descclassname">torch.distributed.rpc.</code><code class="sig-name descname">rpc_async</code><span class="sig-paren">(</span><em class="sig-param">to</em>, <em class="sig-param">func</em>, <em class="sig-param">args=None</em>, <em class="sig-param">kwargs=None</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/rpc/api.html#rpc_async"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.distributed.rpc.rpc_async" title="Permalink to this definition">¶</a></dt>
<dd><p>Make a non-blocking RPC call to run function <code class="docutils literal notranslate"><span class="pre">func</span></code> on worker <code class="docutils literal notranslate"><span class="pre">to</span></code>. RPC
messages are sent and received in parallel to execution of Python code. This
method is thread-safe. This method will immediately return a
<code class="docutils literal notranslate"><span class="pre">torch.distributed.FutureMessage</span></code> that can be awaited on.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>to</strong> (<em>str</em><em> or </em><em>WorkerInfo</em>) – id or name of the destination worker.</p></li>
<li><p><strong>func</strong> (<em>callable</em>) – any callable function. builtin functions (like
<a class="reference internal" href="torch.html#torch.add" title="torch.add"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.add()</span></code></a>) can be sent over RPC more efficiently.</p></li>
<li><p><strong>args</strong> (<em>tuple</em>) – the argument tuple for the <code class="docutils literal notranslate"><span class="pre">func</span></code> invocation.</p></li>
<li><p><strong>kwargs</strong> (<em>dict</em>) – is a dictionary of keyword arguments for the <code class="docutils literal notranslate"><span class="pre">func</span></code>
invocation.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>Returns a <code class="docutils literal notranslate"><span class="pre">torch.distributed.FutureMessage</span></code> object that can be waited
on. When completed, the return value of <code class="docutils literal notranslate"><span class="pre">func</span></code> on <code class="docutils literal notranslate"><span class="pre">args</span></code> and
<code class="docutils literal notranslate"><span class="pre">kwargs</span></code> can be retrieved from the <code class="docutils literal notranslate"><span class="pre">FutureMessage</span></code> object.</p>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">On</span> <span class="n">worker</span> <span class="mi">0</span><span class="p">:</span>
<span class="o">>>></span> <span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="s2">"worker0"</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">2</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">fut1</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">rpc_async</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">2</span><span class="p">),</span> <span class="mi">3</span><span class="p">))</span>
<span class="o">>>></span> <span class="n">fut2</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">rpc_async</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="nb">min</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">))</span>
<span class="o">>>></span> <span class="n">result</span> <span class="o">=</span> <span class="n">fut1</span><span class="o">.</span><span class="n">wait</span><span class="p">()</span> <span class="o">+</span> <span class="n">fut2</span><span class="o">.</span><span class="n">wait</span><span class="p">()</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
<span class="n">On</span> <span class="n">worker</span> <span class="mi">1</span><span class="p">:</span>
<span class="o">>>></span> <span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">rank</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">world_size</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
</pre></div>
</div>
</dd></dl>
<dl class="function">
<dt id="torch.distributed.rpc.remote">
<code class="sig-prename descclassname">torch.distributed.rpc.</code><code class="sig-name descname">remote</code><span class="sig-paren">(</span><em class="sig-param">to</em>, <em class="sig-param">func</em>, <em class="sig-param">args=None</em>, <em class="sig-param">kwargs=None</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/rpc/api.html#remote"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.distributed.rpc.remote" title="Permalink to this definition">¶</a></dt>
<dd><p>Make a remote call to run <code class="docutils literal notranslate"><span class="pre">func</span></code> on worker <code class="docutils literal notranslate"><span class="pre">to</span></code> and return an
<a class="reference internal" href="#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 result value immediately.
Worker <code class="docutils literal notranslate"><span class="pre">to</span></code> will be the owner of the returned
<a class="reference internal" href="#torch.distributed.rpc.RRef" title="torch.distributed.rpc.RRef"><code class="xref py py-class docutils literal notranslate"><span class="pre">RRef</span></code></a>, and the worker calling <code class="docutils literal notranslate"><span class="pre">remote</span></code> is
a user. The owner manages the global reference count of its
<a class="reference internal" href="#torch.distributed.rpc.RRef" title="torch.distributed.rpc.RRef"><code class="xref py py-class docutils literal notranslate"><span class="pre">RRef</span></code></a>, and the owner
<a class="reference internal" href="#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> is only destructed when globally there
are no living references to it.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>to</strong> (<em>str</em><em> or </em><em>WorkerInfo</em>) – id or name of the destination worker.</p></li>
<li><p><strong>func</strong> (<em>callable</em>) – builtin functions (like <a class="reference internal" href="torch.html#torch.add" title="torch.add"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.add()</span></code></a>).</p></li>
<li><p><strong>args</strong> (<em>tuple</em>) – the argument tuple for the <code class="docutils literal notranslate"><span class="pre">func</span></code> invocation.</p></li>
<li><p><strong>kwargs</strong> (<em>dict</em>) – is a dictionary of keyword arguments for the <code class="docutils literal notranslate"><span class="pre">func</span></code>
invocation.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>A user <a class="reference internal" href="#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> instance to the result
value. Use the blocking API <a class="reference internal" href="#torch.distributed.rpc.RRef.to_here" title="torch.distributed.rpc.RRef.to_here"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.distributed.rpc.RRef.to_here()</span></code></a>
to retrieve the result value locally.</p>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">On</span> <span class="n">worker</span> <span class="mi">0</span><span class="p">:</span>
<span class="o">>>></span> <span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="s2">"worker0"</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">2</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">rref1</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">remote</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">2</span><span class="p">),</span> <span class="mi">3</span><span class="p">))</span>
<span class="o">>>></span> <span class="n">rref2</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">remote</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">2</span><span class="p">),</span> <span class="mi">1</span><span class="p">))</span>
<span class="o">>>></span> <span class="n">x</span> <span class="o">=</span> <span class="n">rref1</span><span class="o">.</span><span class="n">to_here</span><span class="p">()</span> <span class="o">+</span> <span class="n">rref2</span><span class="o">.</span><span class="n">to_here</span><span class="p">()</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
<span class="n">On</span> <span class="n">worker</span> <span class="mi">1</span><span class="p">:</span>
<span class="o">>>></span> <span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">rank</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">world_size</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
</pre></div>
</div>
</dd></dl>
<dl class="function">
<dt id="torch.distributed.rpc.get_worker_info">
<code class="sig-prename descclassname">torch.distributed.rpc.</code><code class="sig-name descname">get_worker_info</code><span class="sig-paren">(</span><em class="sig-param">worker_name=None</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/rpc/api.html#get_worker_info"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.distributed.rpc.get_worker_info" title="Permalink to this definition">¶</a></dt>
<dd><p>Get <code class="xref py py-class docutils literal notranslate"><span class="pre">WorkerInfo</span></code> of a given worker name.
Use this <code class="xref py py-class docutils literal notranslate"><span class="pre">WorkerInfo</span></code> to avoid passing an
expensive string on every invocation.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>worker_name</strong> (<em>str</em>) – the string name of a worker. If <code class="docutils literal notranslate"><span class="pre">None</span></code>, return the
the id of the current worker. (default <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p><code class="xref py py-class docutils literal notranslate"><span class="pre">WorkerInfo</span></code> instance for the given
<code class="docutils literal notranslate"><span class="pre">worker_name</span></code> or <code class="xref py py-class docutils literal notranslate"><span class="pre">WorkerInfo</span></code> of the
current worker if <code class="docutils literal notranslate"><span class="pre">worker_name</span></code> is <code class="docutils literal notranslate"><span class="pre">None</span></code>.</p>
</dd>
</dl>
</dd></dl>
<dl class="function">
<dt id="torch.distributed.rpc.shutdown">
<code class="sig-prename descclassname">torch.distributed.rpc.</code><code class="sig-name descname">shutdown</code><span class="sig-paren">(</span><em class="sig-param">graceful=True</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/rpc/api.html#shutdown"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.distributed.rpc.shutdown" title="Permalink to this definition">¶</a></dt>
<dd><p>Perform a shutdown of the RPC agent, and then destroy the RPC agent. This
stops the local agent from accepting outstanding requests, and shuts
down the RPC framework by terminating all RPC threads. If graceful=True,
then this will block until all local and remote RPC processes reach this
method and wait for all outstanding work to complete. Otherwise, if
graceful=False, then this is a local shutdown, and it does not wait for
other RPC processes to reach this method.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>graceful</strong> (<em>bool</em>) – Whether to do a graceful shutdown or not. If True,
this will block until all local and remote RPC
processes have reached this method and wait for all
outstanding work to complete.</p>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">On</span> <span class="n">worker</span> <span class="mi">0</span><span class="p">:</span>
<span class="o">>>></span> <span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="s2">"worker0"</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">2</span><span class="p">)</span>
<span class="o">>>></span> <span class="c1"># do some work</span>
<span class="o">>>></span> <span class="n">result</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">rpc_sync</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">1</span><span class="p">),</span> <span class="mi">1</span><span class="p">))</span>
<span class="o">>>></span> <span class="c1"># ready to shutdown</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
<span class="n">On</span> <span class="n">worker</span> <span class="mi">1</span><span class="p">:</span>
<span class="o">>>></span> <span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">rank</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">world_size</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
<span class="o">>>></span> <span class="c1"># wait for worker 0 to finish work, and then shutdown.</span>
<span class="o">>>></span> <span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
</pre></div>
</div>
</dd></dl>
</div>
<div class="section" id="distributed-autograd-framework">
<h2>Distributed Autograd Framework<a class="headerlink" href="#distributed-autograd-framework" title="Permalink to this headline">¶</a></h2>
<p>This module provides an RPC-based distributed autograd framework that can be
used for applications such as model parallel training. In short, applications
may send and receive gradient recording tensors over RPC. In the forward pass,
we record when gradient recording tensors are sent over RPC and during the
backward pass we use this information to perform a distributed backward pass
using RPC. For more details see <a class="reference internal" href="notes/distributed_autograd.html#distributed-autograd-design"><span class="std std-ref">Distributed Autograd Design</span></a>.</p>
<span class="target" id="module-torch.distributed.autograd"></span><dl class="class">
<dt id="torch.distributed.autograd.context">
<em class="property">class </em><code class="sig-prename descclassname">torch.distributed.autograd.</code><code class="sig-name descname">context</code><a class="reference internal" href="_modules/torch/distributed/autograd.html#context"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.distributed.autograd.context" title="Permalink to this definition">¶</a></dt>
<dd><p>Context object to wrap forward and backward passes when using
distributed autograd. The <code class="docutils literal notranslate"><span class="pre">context_id</span></code> generated in the <code class="docutils literal notranslate"><span class="pre">with</span></code>
statement is required to uniquely identify a distributed backward pass
on all workers. Each worker stores metadata associated with this
<code class="docutils literal notranslate"><span class="pre">context_id</span></code>, which is required to correctly execute a distributed
autograd pass.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="o">>></span> <span class="kn">import</span> <span class="nn">torch.distributed.autograd</span> <span class="k">as</span> <span class="nn">dist_autograd</span>
<span class="o">>></span> <span class="k">with</span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">context</span><span class="p">()</span> <span class="k">as</span> <span class="n">context_id</span><span class="p">:</span>
<span class="o">>></span> <span class="n">t1</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">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">),</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="o">>></span> <span class="n">t2</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">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">),</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="o">>></span> <span class="n">loss</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">rpc_sync</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">t1</span><span class="p">,</span> <span class="n">t2</span><span class="p">))</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span>
<span class="o">>></span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">backward</span><span class="p">([</span><span class="n">loss</span><span class="p">])</span>
</pre></div>
</div>
</dd></dl>
<dl class="function">
<dt id="torch.distributed.autograd.backward">
<code class="sig-prename descclassname">torch.distributed.autograd.</code><code class="sig-name descname">backward</code><span class="sig-paren">(</span><em class="sig-param">roots: List[Tensor]</em><span class="sig-paren">)</span> → None<a class="headerlink" href="#torch.distributed.autograd.backward" title="Permalink to this definition">¶</a></dt>
<dd><p>Kicks off the distributed backward pass using the provided roots. This
currently implements the <a class="reference internal" href="notes/distributed_autograd.html#fast-mode-algorithm"><span class="std std-ref">FAST mode algorithm</span></a> which
assumes all RPC messages sent in the same distributed autograd context
across workers would be part of the autograd graph during the backward pass.</p>
<p>We use the provided roots to discover the autograd graph and compute
appropriate dependencies. This method blocks until the entire
autograd computation is done.</p>
<p>We accumulate the gradients in the appropriate
<a class="reference internal" href="#torch.distributed.autograd.context" title="torch.distributed.autograd.context"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.distributed.autograd.context</span></code></a> on each of the nodes. The autograd
context used is the current autograd context of this node when
<a class="reference internal" href="#torch.distributed.autograd.backward" title="torch.distributed.autograd.backward"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.distributed.autograd.backward()</span></code></a> is called. If there is no valid
autograd context, we throw an error. You can retrieve the accumulated
gradients using the <a class="reference internal" href="#torch.distributed.autograd.get_gradients" title="torch.distributed.autograd.get_gradients"><code class="xref py py-meth docutils literal notranslate"><span class="pre">get_gradients()</span></code></a> API.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>roots</strong> (<em>list</em>) – Tensors which represent the roots of the autograd
computation. All the tensors should be scalars.</p>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="o">>></span> <span class="kn">import</span> <span class="nn">torch.distributed.autograd</span> <span class="k">as</span> <span class="nn">dist_autograd</span>
<span class="o">>></span> <span class="k">with</span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">context</span><span class="p">()</span> <span class="k">as</span> <span class="n">context_id</span><span class="p">:</span>
<span class="o">>></span> <span class="n">pred</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">forward</span><span class="p">()</span>
<span class="o">>></span> <span class="n">loss</span> <span class="o">=</span> <span class="n">loss_func</span><span class="p">(</span><span class="n">pred</span><span class="p">,</span> <span class="n">loss</span><span class="p">)</span>
<span class="o">>></span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">backward</span><span class="p">(</span><span class="n">loss</span><span class="p">)</span>
</pre></div>
</div>
</dd></dl>
<dl class="function">
<dt id="torch.distributed.autograd.get_gradients">
<code class="sig-prename descclassname">torch.distributed.autograd.</code><code class="sig-name descname">get_gradients</code><span class="sig-paren">(</span><em class="sig-param">context_id: int</em><span class="sig-paren">)</span> → Dict[Tensor, Tensor]<a class="headerlink" href="#torch.distributed.autograd.get_gradients" title="Permalink to this definition">¶</a></dt>
<dd><p>Retrieves a map from Tensor to the appropriate gradient for that Tensor
accumulated in the provided <code class="docutils literal notranslate"><span class="pre">context_id</span></code> as part of the distributed autograd
backward pass.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>context_id</strong> (<em>python:int</em>) – The autograd context id for which we should retrieve the
gradients.</p>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>A map where the key is the Tensor and the value is the associated gradient for that Tensor.</p>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="o">>></span> <span class="kn">import</span> <span class="nn">torch.distributed.autograd</span> <span class="k">as</span> <span class="nn">dist_autograd</span>
<span class="o">>></span> <span class="k">with</span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">context</span><span class="p">()</span> <span class="k">as</span> <span class="n">context_id</span><span class="p">:</span>
<span class="o">>></span> <span class="n">t1</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">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">),</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="o">>></span> <span class="n">t2</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">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">),</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="o">>></span> <span class="n">loss</span> <span class="o">=</span> <span class="n">t1</span> <span class="o">+</span> <span class="n">t2</span>
<span class="o">>></span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">backward</span><span class="p">([</span><span class="n">loss</span><span class="o">.</span><span class="n">sum</span><span class="p">()])</span>
<span class="o">>></span> <span class="n">grads</span> <span class="o">=</span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">get_gradients</span><span class="p">(</span><span class="n">context_id</span><span class="p">)</span>
<span class="o">>></span> <span class="nb">print</span> <span class="p">(</span><span class="n">grads</span><span class="p">[</span><span class="n">t1</span><span class="p">])</span>
<span class="o">>></span> <span class="nb">print</span> <span class="p">(</span><span class="n">grads</span><span class="p">[</span><span class="n">t2</span><span class="p">])</span>
</pre></div>
</div>
</dd></dl>
</div>
<div class="section" id="module-torch.distributed.optim">
<span id="distributed-optimizer"></span><h2>Distributed Optimizer<a class="headerlink" href="#module-torch.distributed.optim" title="Permalink to this headline">¶</a></h2>
<p><a class="reference internal" href="#module-torch.distributed.optim" title="torch.distributed.optim"><code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.distributed.optim</span></code></a> exposes DistributedOptimizer, which takes a list
of remote parameters (<a class="reference internal" href="#torch.distributed.rpc.RRef" title="torch.distributed.rpc.RRef"><code class="xref py py-class docutils literal notranslate"><span class="pre">RRef</span></code></a>) and runs the
optimizer locally on the workers where the parameters live. The distributed
optimizer can use any of the local optimizer <a class="reference internal" href="optim.html#optimizer-algorithms"><span class="std std-ref">Algorithms</span></a> to
apply the gradients on each worker.</p>
<dl class="class">
<dt id="torch.distributed.optim.DistributedOptimizer">
<em class="property">class </em><code class="sig-prename descclassname">torch.distributed.optim.</code><code class="sig-name descname">DistributedOptimizer</code><span class="sig-paren">(</span><em class="sig-param">optimizer_class</em>, <em class="sig-param">params_rref</em>, <em class="sig-param">*args</em>, <em class="sig-param">**kwargs</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/optimizer.html#DistributedOptimizer"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.distributed.optim.DistributedOptimizer" title="Permalink to this definition">¶</a></dt>
<dd><p>DistributedOptimizer takes remote references to parameters scattered
across workers and applies the given optimizer locally for each parameter.</p>
<p>This class uses <a class="reference internal" href="#torch.distributed.autograd.get_gradients" title="torch.distributed.autograd.get_gradients"><code class="xref py py-meth docutils literal notranslate"><span class="pre">get_gradients()</span></code></a> in order
to retrieve the gradients for specific parameters.</p>
<p>Concurrent calls to
<a class="reference internal" href="#torch.distributed.optim.DistributedOptimizer.step" title="torch.distributed.optim.DistributedOptimizer.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code></a>,
either from the same or different clients, will
be serialized on each worker – as each worker’s optimizer can only work
on one set of gradients at a time. However, there is no guarantee that
the full forward-backward-optimizer sequence will execute for one client
at a time. This means that the gradients being applied may not correspond
to the latest forward pass executed on a given worker. Also, there is no
guaranteed ordering across workers.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>optimizer_class</strong> (<a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.Optimizer"><em>optim.Optimizer</em></a>) – the class of optimizer to
instantiate on each worker.</p></li>
<li><p><strong>params_rref</strong> (<em>list</em><em>[</em><a class="reference internal" href="#torch.distributed.rpc.RRef" title="torch.distributed.rpc.RRef"><em>RRef</em></a><em>]</em>) – list of RRefs to local or remote parameters
to optimize.</p></li>
<li><p><strong>args</strong> – arguments to pass to the optimizer constructor on each worker.</p></li>
<li><p><strong>kwargs</strong> – arguments to pass to the optimizer constructor on each worker.</p></li>
</ul>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="o">>></span> <span class="kn">import</span> <span class="nn">torch.distributed.autograd</span> <span class="k">as</span> <span class="nn">dist_autograd</span>
<span class="o">>></span> <span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="o">>></span> <span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">optim</span>
<span class="o">>></span> <span class="kn">from</span> <span class="nn">torch.distributed.optim</span> <span class="kn">import</span> <span class="n">DistributedOptimizer</span>
<span class="o">>></span>
<span class="o">>></span> <span class="k">with</span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">context</span><span class="p">()</span> <span class="k">as</span> <span class="n">context_id</span><span class="p">:</span>
<span class="o">>></span> <span class="c1"># Forward pass.</span>
<span class="o">>></span> <span class="n">rref1</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">remote</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">2</span><span class="p">),</span> <span class="mi">3</span><span class="p">))</span>
<span class="o">>></span> <span class="n">rref2</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">remote</span><span class="p">(</span><span class="s2">"worker1"</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">2</span><span class="p">),</span> <span class="mi">1</span><span class="p">))</span>
<span class="o">>></span> <span class="n">loss</span> <span class="o">=</span> <span class="n">rref1</span><span class="o">.</span><span class="n">to_here</span><span class="p">()</span> <span class="o">+</span> <span class="n">rref2</span><span class="o">.</span><span class="n">to_here</span><span class="p">()</span>
<span class="o">>></span>
<span class="o">>></span> <span class="c1"># Backward pass.</span>
<span class="o">>></span> <span class="n">dist_autograd</span><span class="o">.</span><span class="n">backward</span><span class="p">([</span><span class="n">loss</span><span class="o">.</span><span class="n">sum</span><span class="p">()])</span>
<span class="o">>></span>
<span class="o">>></span> <span class="c1"># Optimizer.</span>
<span class="o">>></span> <span class="n">dist_optim</span> <span class="o">=</span> <span class="n">DistributedOptimizer</span><span class="p">(</span>
<span class="o">>></span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">,</span>
<span class="o">>></span> <span class="p">[</span><span class="n">rref1</span><span class="p">,</span> <span class="n">rref2</span><span class="p">],</span>
<span class="o">>></span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.05</span><span class="p">,</span>
<span class="o">>></span> <span class="p">)</span>
<span class="o">>></span> <span class="n">dist_optim</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<dl class="method">
<dt id="torch.distributed.optim.DistributedOptimizer.step">
<code class="sig-name descname">step</code><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/optim/optimizer.html#DistributedOptimizer.step"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.distributed.optim.DistributedOptimizer.step" title="Permalink to this definition">¶</a></dt>
<dd><p>Performs a single optimization step.</p>
<p>This will call <a class="reference internal" href="optim.html#torch.optim.Optimizer.step" title="torch.optim.Optimizer.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.optim.Optimizer.step()</span></code></a> on each worker
containing parameters to be optimized, and will block until all workers
return. The current distributed autograd
<a class="reference internal" href="#torch.distributed.autograd.context" title="torch.distributed.autograd.context"><code class="xref py py-class docutils literal notranslate"><span class="pre">context</span></code></a> will be used globally.</p>
</dd></dl>
</dd></dl>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="random.html" class="btn btn-neutral float-right" title="torch.random" accesskey="n" rel="next">Next <img src="_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="quantization.html" class="btn btn-neutral" title="Quantization" 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="#">Distributed RPC Framework</a><ul>
<li><a class="reference internal" href="#rpc-and-rref-framework">RPC and RRef Framework</a></li>
<li><a class="reference internal" href="#rref">RRef</a></li>
<li><a class="reference internal" href="#rpc-and-rref-primitives">RPC and RRef primitives</a></li>
<li><a class="reference internal" href="#distributed-autograd-framework">Distributed Autograd Framework</a></li>
<li><a class="reference internal" href="#module-torch.distributed.optim">Distributed Optimizer</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/language_data.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 type="text/javascript" src="_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<script>
(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
})(window,document,'script','https://www.google-analytics.com/analytics.js','ga');
ga('create', 'UA-90545585-1', 'auto');
ga('send', 'pageview');
</script>
<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>
<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://pytorch.org/resources">Resources</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/support">Support</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.slack.com" target="_blank">Slack</a></li>
<li><a href="https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md" target="_blank">Contributing</a></li>
</ul>
</div>
<div class="footer-links-col follow-us-col">
<ul>
<li class="list-title">Follow Us</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>
</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">