forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrref.html
909 lines (714 loc) · 54.4 KB
/
rref.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
<!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>Remote Reference Protocol — PyTorch master documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/notes/rref.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="Serialization semantics" href="serialization.html" />
<link rel="prev" title="Reproducibility" href="randomness.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 class="current">
<li class="toctree-l1"><a class="reference internal" href="autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed_autograd.html">Distributed Autograd Design</a></li>
<li class="toctree-l1"><a class="reference internal" href="extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="randomness.html">Reproducibility</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Remote Reference Protocol</a></li>
<li class="toctree-l1"><a class="reference internal" href="serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="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>
<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"><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="../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>Remote Reference Protocol</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/notes/rref.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="remote-reference-protocol">
<span id="id1"></span><h1>Remote Reference Protocol<a class="headerlink" href="#remote-reference-protocol" title="Permalink to this headline">¶</a></h1>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>The <a class="reference internal" href="../rpc.html#rref"><span class="std std-ref">RRef</span></a> API is experimental and subject to change.</p>
</div>
<p>This note describes the design details of Remote Reference protocol and walks
through message flows in different scenarios. Make sure you’re familiar with the
<a class="reference internal" href="../rpc.html#distributed-rpc-framework"><span class="std std-ref">Distributed RPC Framework</span></a> before proceeding.</p>
<div class="section" id="background">
<h2>Background<a class="headerlink" href="#background" title="Permalink to this headline">¶</a></h2>
<p>RRef stands for Remote REFerence. It is a reference of an object which is
located on the local or a remote worker, and transparently handles reference
counting under the hood. Conceptually, it can be considered as a distributed
shared pointer. Applications can create an RRef by calling
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.remote" title="torch.distributed.rpc.remote"><code class="xref py py-meth docutils literal notranslate"><span class="pre">remote()</span></code></a>. Each RRef is owned by the callee worker
of the <a class="reference internal" href="../rpc.html#torch.distributed.rpc.remote" title="torch.distributed.rpc.remote"><code class="xref py py-meth docutils literal notranslate"><span class="pre">remote()</span></code></a> call (i.e., owner) and can be used
by multiple users. The owner stores the real data and keeps track of the global
reference count. Every RRef can be uniquely identified by a global <code class="docutils literal notranslate"><span class="pre">RRefId</span></code>,
which is assigned at the time of creation on the caller of the
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.remote" title="torch.distributed.rpc.remote"><code class="xref py py-meth docutils literal notranslate"><span class="pre">remote()</span></code></a> call.</p>
<p>On the owner worker, there is only one <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> instance, which contains
the real data, while on user workers, there can be as many <code class="docutils literal notranslate"><span class="pre">UserRRefs</span></code> as
necessary, and <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> does not hold the data. All usage on the owner will
retrieve the unique <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> instance using the globally unique <code class="docutils literal notranslate"><span class="pre">RRefId</span></code>.
A <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> will be created when it is used as an argument or return value in
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.rpc_sync" title="torch.distributed.rpc.rpc_sync"><code class="xref py py-meth docutils literal notranslate"><span class="pre">rpc_sync()</span></code></a>,
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.rpc_async" title="torch.distributed.rpc.rpc_async"><code class="xref py py-meth docutils literal notranslate"><span class="pre">rpc_async()</span></code></a> or
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.remote" title="torch.distributed.rpc.remote"><code class="xref py py-meth docutils literal notranslate"><span class="pre">remote()</span></code></a> invocation, and the owner will be notified
according to update the reference count. An <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> and its data will be
deleted when there is no <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> instances globally and there are no
reference to the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> on the owner as well.</p>
</div>
<div class="section" id="assumptions">
<h2>Assumptions<a class="headerlink" href="#assumptions" title="Permalink to this headline">¶</a></h2>
<p>RRef protocol is designed with the following assumptions.</p>
<ul class="simple">
<li><p><strong>Transient Network Failures</strong>: The RRef design aims to handle transient
network failures by retrying messages. Node crashes or permanent network
partition is beyond the scope. When those incidents occur, the application
may take down all workers, revert to the previous checkpoint, and resume
training.</p></li>
<li><p><strong>Non-idempotent UDFs</strong>: We assume the user functions (UDF) provided to
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.rpc_sync" title="torch.distributed.rpc.rpc_sync"><code class="xref py py-meth docutils literal notranslate"><span class="pre">rpc_sync()</span></code></a>,
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.rpc_async" title="torch.distributed.rpc.rpc_async"><code class="xref py py-meth docutils literal notranslate"><span class="pre">rpc_async()</span></code></a> or
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.remote" title="torch.distributed.rpc.remote"><code class="xref py py-meth docutils literal notranslate"><span class="pre">remote()</span></code></a> are not idempotent and therefore
cannot be retried. However, internal RRef control messages will be made
idempotent and retryable.</p></li>
<li><p><strong>Out of Order Message Delivery</strong>: We do not assume message delivery order
between any pair of nodes, because both sender and receiver are using multiple
threads. There is no guarantee on which message will be processed first.</p></li>
</ul>
</div>
<div class="section" id="rref-lifetime">
<h2>RRef Lifetime<a class="headerlink" href="#rref-lifetime" title="Permalink to this headline">¶</a></h2>
<p>The goal of the protocol is to delete an <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> at an appropriate time.
The right time to delete an <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> is when there are no living
<code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> instances and user code is not holding references to the
<code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> either. The tricky part is to determine if there are any living
<code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> instances.</p>
<div class="section" id="design-reasoning">
<h3>Design Reasoning<a class="headerlink" href="#design-reasoning" title="Permalink to this headline">¶</a></h3>
<p>A user can get a <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> in three situations:</p>
<ol class="arabic simple">
<li><p>Receiving a <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> from the owner.</p></li>
<li><p>Receiving a <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> from another user.</p></li>
<li><p>Creating a new <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> owned by another worker.</p></li>
</ol>
<p>Case 1 is the simplest where the owner passes its RRef to a user, where the
owner calls <a class="reference internal" href="../rpc.html#torch.distributed.rpc.rpc_sync" title="torch.distributed.rpc.rpc_sync"><code class="xref py py-meth docutils literal notranslate"><span class="pre">rpc_sync()</span></code></a>,
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.rpc_async" title="torch.distributed.rpc.rpc_async"><code class="xref py py-meth docutils literal notranslate"><span class="pre">rpc_async()</span></code></a>, or
<a class="reference internal" href="../rpc.html#torch.distributed.rpc.remote" title="torch.distributed.rpc.remote"><code class="xref py py-meth docutils literal notranslate"><span class="pre">remote()</span></code></a> and uses its RRef as an argument. In this
case a new <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> will be created on the user. As the owner is the caller,
it can easily update its local reference count on the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code>.</p>
<p>The only requirement is that any
<code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> must notify the owner upon destruction. Hence, we need the first
guarantee:</p>
<p><strong>G1. The owner will be notified when any ``UserRRef`` is deleted.</strong></p>
<p>As messages might come delayed or out-of-order, we need one more guarantee to
make sure the delete message is not processed too soon. If A sends a message to
B that involves an RRef, we call the RRef on A the parent RRef and the RRef on B
the child RRef.</p>
<p><strong>G2. Parent RRef will NOT be deleted until the child RRef is confirmed by the
owner.</strong></p>
<p>In cases 2 and 3, it is possible that the owner has only partial or no knowledge
at all about the RRef fork graph. For example, an RRef could be
constructed on a user, and before the owner receives any RPC call, the
creator user might have already shared the RRef with other users, and those
users could further share the RRef. One invariant is that the fork graph of
any RRef is always a tree, because forking an RRef always
creates a new <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> instance on the callee (except if the callee is the
owner), and hence every RRef has a single parent.</p>
<p>The owner’s view on any <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> in the tree has three stages:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="mi">1</span><span class="p">)</span> <span class="n">unknown</span> <span class="o">-></span> <span class="mi">2</span><span class="p">)</span> <span class="n">known</span> <span class="o">-></span> <span class="mi">3</span><span class="p">)</span> <span class="n">deleted</span><span class="o">.</span>
</pre></div>
</div>
<p>The owner’s view of the entire tree keeps changing. The owner deletes its
<code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> instance when it thinks there are no living <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code>
instances, i.e.,
when <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> is deleted, all <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> instances could be either indeed
deleted or unknown. The dangerous case is when some forks are unknown and others
are deleted.</p>
<p><strong>G2</strong> trivially guarantees that no parent <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> can be deleted before
the owner knows all of its children <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> instances. However, it is
possible that the child <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> may be deleted before the owner knows its
parent <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code>.</p>
<p>Consider the following example, where the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> forks to A, then A forks
to Y, and Y forks to Z.:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">OwnerRRef</span> <span class="o">-></span> <span class="n">A</span> <span class="o">-></span> <span class="n">Y</span> <span class="o">-></span> <span class="n">Z</span>
</pre></div>
</div>
<p>If all of Z’s messages, including the delete message, are processed by the
owner before all messages from Y, the owner will learn Z’s deletion before
knowing Y. Nevertheless, this does not cause any problem. Because, at least
one of Y’s ancestors will be alive (in this case, A) and it will
prevent the owner from deleting the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code>. More specifically, if the
owner does not know Y, A cannot be deleted due to <strong>G2</strong>, and the owner knows A
as the owner is A’s parent.</p>
<p>Things get a little trickier if the RRef is created on a user:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">OwnerRRef</span>
<span class="o">^</span>
<span class="o">|</span>
<span class="n">A</span> <span class="o">-></span> <span class="n">Y</span> <span class="o">-></span> <span class="n">Z</span>
</pre></div>
</div>
<p>If Z calls <a class="reference internal" href="../rpc.html#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">to_here()</span></code></a> on the <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code>, the
owner at least knows A when Z is deleted, because otherwise,
<a class="reference internal" href="../rpc.html#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">to_here()</span></code></a> wouldn’t finish. If Z does not call
<a class="reference internal" href="../rpc.html#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">to_here()</span></code></a>, it is possible that the owner
receives all messages from Z before any message from A and Y. In this case, as
the real data of the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> has not been created yet, there is nothing to
be deleted either. It is the same as Z does not exist at all. Hence, it’s still
OK.</p>
</div>
<div class="section" id="implementation">
<h3>Implementation<a class="headerlink" href="#implementation" title="Permalink to this headline">¶</a></h3>
<p><strong>G1</strong> is implemented by sending out a delete message in <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code>
destructor. To provide <strong>G2</strong>, the parent <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> is put into a context
whenever it is forked, indexed by the new <code class="docutils literal notranslate"><span class="pre">ForkId</span></code>. The parent <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> is
only removed from the context when it receives an acknowledgement message (ACK)
from the child, and the child will only send out the ACK when it is confirmed by
the owner.</p>
</div>
</div>
<div class="section" id="protocol-scenarios">
<h2>Protocol Scenarios<a class="headerlink" href="#protocol-scenarios" title="Permalink to this headline">¶</a></h2>
<p>Let’s now discuss how the above designs translate to the protocol in four
scenarios.</p>
<div class="section" id="user-share-rref-with-owner-as-return-value">
<h3>User Share RRef with Owner as Return Value<a class="headerlink" href="#user-share-rref-with-owner-as-return-value" title="Permalink to this headline">¶</a></h3>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</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="c1"># on worker A</span>
<span class="n">rref</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="s1">'B'</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="c1"># say the rref has RRefId 100 and ForkId 1</span>
<span class="n">rref</span><span class="o">.</span><span class="n">to_here</span><span class="p">()</span>
</pre></div>
</div>
<p>In this case, the <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> is created on the user worker A, then it is
passed to the owner worker B together with the remote message, and then B
creates the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code>. The method <a class="reference internal" href="../rpc.html#torch.distributed.rpc.remote" title="torch.distributed.rpc.remote"><code class="xref py py-meth docutils literal notranslate"><span class="pre">remote()</span></code></a>
returns immediately, meaning that the <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> can be forked/used before
the owner knows about it.</p>
<p>On the owner, when receiving the <a class="reference internal" href="../rpc.html#torch.distributed.rpc.remote" title="torch.distributed.rpc.remote"><code class="xref py py-meth docutils literal notranslate"><span class="pre">remote()</span></code></a> call, it
will create the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code>, and returns an ACK to acknowledge <code class="docutils literal notranslate"><span class="pre">{100,</span> <span class="pre">1}</span></code>
(<code class="docutils literal notranslate"><span class="pre">RRefId</span></code>, <code class="docutils literal notranslate"><span class="pre">ForkId</span></code>). Only after receiving this ACK, can A delete its
<code class="docutils literal notranslate"><span class="pre">UserRRef</span></code>. This involves both <strong>G1</strong> and <strong>G2</strong>. <strong>G1</strong> is obvious. For
<strong>G2</strong>, the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> is a child of the <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code>, and the <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code>
is not deleted until it receives the ACK from the owner.</p>
<a class="reference internal image-reference" href="https://user-images.githubusercontent.com/16999635/69164772-98181300-0abe-11ea-93a7-9ad9f757cd94.png"><img alt="user_to_owner_ret.png" src="https://user-images.githubusercontent.com/16999635/69164772-98181300-0abe-11ea-93a7-9ad9f757cd94.png" style="width: 500px;" /></a>
<p>The diagram above shows the message flow, where solid arrow contains user
function and dashed arrow are builtin messages. Note that the first two messages
from A to B (<a class="reference internal" href="../rpc.html#torch.distributed.rpc.remote" title="torch.distributed.rpc.remote"><code class="xref py py-meth docutils literal notranslate"><span class="pre">remote()</span></code></a> and
<a class="reference internal" href="../rpc.html#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">to_here()</span></code></a>) may
arrive at B in any order, but the final delete message will only be sent out
when:</p>
<ul class="simple">
<li><p>B acknowledges <code class="docutils literal notranslate"><span class="pre">UserRRef</span> <span class="pre">{100,</span> <span class="pre">1}</span></code> (G2), and</p></li>
<li><p>Python GC agrees to delete the local <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> instance. This occurs when
the RRef is no longer in scope and is eligible for garbage collection.</p></li>
</ul>
</div>
<div class="section" id="user-share-rref-with-owner-as-argument">
<h3>User Share RRef with Owner as Argument<a class="headerlink" href="#user-share-rref-with-owner-as-argument" title="Permalink to this headline">¶</a></h3>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</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="c1"># on worker A and worker B</span>
<span class="k">def</span> <span class="nf">func</span><span class="p">(</span><span class="n">rref</span><span class="p">):</span>
<span class="k">pass</span>
<span class="c1"># on worker A</span>
<span class="n">rref</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="s1">'B'</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="c1"># say the rref has RRefId 100 and ForkId 1</span>
<span class="n">rpc</span><span class="o">.</span><span class="n">rpc_async</span><span class="p">(</span><span class="s1">'B'</span><span class="p">,</span> <span class="n">func</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">rref</span><span class="p">,</span> <span class="p">))</span>
</pre></div>
</div>
<p>In this case, after creating the <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> on A, A uses it as an argument in
a followup RPC call to B. A will keep <code class="docutils literal notranslate"><span class="pre">UserRRef</span> <span class="pre">{100,</span> <span class="pre">1}</span></code> alive until it
receives the acknowledge from B (<strong>G2</strong>, not the return value of the RPC call).
This is necessary because A should not send out the delete message until all
previous messages are received, otherwise, the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> could be
deleted before usage as we do not guarantee message delivery order. This is done
by creating a child <code class="docutils literal notranslate"><span class="pre">ForkId</span></code> of RRef, holding them in a map until receives the
owner confirms the child <code class="docutils literal notranslate"><span class="pre">ForkId</span></code>. The figure below shows the message flow.</p>
<a class="reference internal image-reference" href="https://user-images.githubusercontent.com/16999635/69164845-b67e0e80-0abe-11ea-93fa-d24674e75a2b.png"><img alt="user_to_owner_arg.png" src="https://user-images.githubusercontent.com/16999635/69164845-b67e0e80-0abe-11ea-93fa-d24674e75a2b.png" style="width: 500px;" /></a>
<p>Note that the <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> could be deleted on B before func finishes or even
starts. However this is OK, as at the time B sends out ACK for the child
<code class="docutils literal notranslate"><span class="pre">ForkId</span></code>, it already acquired the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> instance, which would prevent
it been deleted too soon.</p>
</div>
<div class="section" id="owner-share-rref-with-user">
<h3>Owner Share RRef with User<a class="headerlink" href="#owner-share-rref-with-user" title="Permalink to this headline">¶</a></h3>
<p>Owner to user is the simplest case, where the owner can update reference
counting locally, and does not need any additional control message to notify
others. Regarding <strong>G2</strong>, it is same as the parent receives the ACK from the
owner immediately, as the parent is the owner.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">RRef</span><span class="o">,</span> <span class="nn">rpc</span>
<span class="c1"># on worker B and worker C</span>
<span class="k">def</span> <span class="nf">func</span><span class="p">(</span><span class="n">rref</span><span class="p">):</span>
<span class="k">pass</span>
<span class="c1"># on worker B, creating a local RRef</span>
<span class="n">rref</span> <span class="o">=</span> <span class="n">RRef</span><span class="p">(</span><span class="s2">"data"</span><span class="p">)</span>
<span class="c1"># say the rref has RRefId 100</span>
<span class="n">dist</span><span class="o">.</span><span class="n">rpc_async</span><span class="p">(</span><span class="s1">'C'</span><span class="p">,</span> <span class="n">func</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">rref</span><span class="p">,</span> <span class="p">))</span>
</pre></div>
</div>
<a class="reference internal image-reference" href="https://user-images.githubusercontent.com/16999635/69164921-c990de80-0abe-11ea-9250-d32ad00cf4ae.png"><img alt="owner_to_user.png" src="https://user-images.githubusercontent.com/16999635/69164921-c990de80-0abe-11ea-9250-d32ad00cf4ae.png" style="width: 500px;" /></a>
<p>The figure above shows the message flow. Note that when the <code class="docutils literal notranslate"><span class="pre">OwnerRRef</span></code> exits
scope after the rpc_async call, it will not be deleted, because internally
there is a map to hold it alive if there is any known forks, in which case is
<code class="docutils literal notranslate"><span class="pre">UserRRef</span> <span class="pre">{100,</span> <span class="pre">1}</span></code>. (<strong>G2</strong>)</p>
</div>
<div class="section" id="user-share-rref-with-user">
<h3>User Share RRef with User<a class="headerlink" href="#user-share-rref-with-user" title="Permalink to this headline">¶</a></h3>
<p>This is the most complicated case where caller user (parent <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code>),
callee user (child <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code>), and the owner all need to get involved.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</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="c1"># on worker A and worker C</span>
<span class="k">def</span> <span class="nf">func</span><span class="p">(</span><span class="n">rref</span><span class="p">):</span>
<span class="k">pass</span>
<span class="c1"># on worker A</span>
<span class="n">rref</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="s1">'B'</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="c1"># say the rref has RRefId 100 and ForkId 1</span>
<span class="n">rpc</span><span class="o">.</span><span class="n">rpc_async</span><span class="p">(</span><span class="s1">'C'</span><span class="p">,</span> <span class="n">func</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">rref</span><span class="p">,</span> <span class="p">))</span>
</pre></div>
</div>
<a class="reference internal image-reference" href="https://user-images.githubusercontent.com/16999635/69164971-d6adcd80-0abe-11ea-971d-6b7af131f0fd.png"><img alt="user_to_user.png" src="https://user-images.githubusercontent.com/16999635/69164971-d6adcd80-0abe-11ea-971d-6b7af131f0fd.png" style="width: 500px;" /></a>
<p>When C receives the child <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> from A, it sends out a fork request to
the owner B. Later, when the B confirms the <code class="docutils literal notranslate"><span class="pre">UserRRef</span></code> on C, C will perform
two actions in parallel: 1) send out the child ACK to A ,and 2) run the user
provided function. During this time, the parent (A) will hold its
<code class="docutils literal notranslate"><span class="pre">UserRRef</span> <span class="pre">{100,</span> <span class="pre">1}</span></code> alive to achieve <strong>G2</strong>.</p>
</div>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="serialization.html" class="btn btn-neutral float-right" title="Serialization semantics" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="randomness.html" class="btn btn-neutral" title="Reproducibility" 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="#">Remote Reference Protocol</a><ul>
<li><a class="reference internal" href="#background">Background</a></li>
<li><a class="reference internal" href="#assumptions">Assumptions</a></li>
<li><a class="reference internal" href="#rref-lifetime">RRef Lifetime</a><ul>
<li><a class="reference internal" href="#design-reasoning">Design Reasoning</a></li>
<li><a class="reference internal" href="#implementation">Implementation</a></li>
</ul>
</li>
<li><a class="reference internal" href="#protocol-scenarios">Protocol Scenarios</a><ul>
<li><a class="reference internal" href="#user-share-rref-with-owner-as-return-value">User Share RRef with Owner as Return Value</a></li>
<li><a class="reference internal" href="#user-share-rref-with-owner-as-argument">User Share RRef with Owner as Argument</a></li>
<li><a class="reference internal" href="#owner-share-rref-with-user">Owner Share RRef with User</a></li>
<li><a class="reference internal" href="#user-share-rref-with-user">User Share RRef with User</a></li>
</ul>
</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">
<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/features">Features</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="active">
<a href="https://pytorch.org/docs/stable/index.html">Docs</a>
</li>
<li>
<a href="https://pytorch.org/resources">Resources</a>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">Github</a>
</li>
</ul>
</div>
</div>
</div>
<!-- End Mobile Menu -->
<script type="text/javascript" src="../_static/js/vendor/anchor.min.js"></script>
<script type="text/javascript">
$(document).ready(function() {
mobileMenu.bind();
mobileTOC.bind();
pytorchAnchors.bind();
sideMenus.bind();
scrollToAnchor.bind();
highlightNavigation.bind();
mainMenuDropdown.bind();
// Add class to links that have code blocks, since we cannot create links in code blocks
$("article.pytorch-article a span.pre").each(function(e) {
$(this).closest("a").addClass("has-code");
});
})
</script>
</body>
</html>