-
Notifications
You must be signed in to change notification settings - Fork 258
/
Copy pathrpc_param_server_tutorial.html
1106 lines (914 loc) ยท 111 KB
/
rpc_param_server_tutorial.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="ko" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="ko" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta property="og:title" content="Implementing a Parameter Server Using Distributed RPC Framework" />
<meta property="og:type" content="article" />
<meta property="og:url" content="https://tutorials.pytorch.kr/intermediate/rpc_param_server_tutorial.html" />
<meta property="og:site_name" content="PyTorch Tutorials KR" />
<meta property="og:description" content="Author: Rohan Varma Prerequisites: PyTorch Distributed Overview, RPC API documents. This tutorial walks through a simple example of implementing a parameter server using PyTorchโs Distributed RPC framework. The parameter server framework is a paradigm in which a set of servers store parameters, s..." />
<meta property="og:image" content="https://tutorials.pytorch.kr/_static/logos/logo-kr-sm-dark.png" />
<meta property="og:image:alt" content="PyTorch Tutorials KR" />
<meta name="description" content="Author: Rohan Varma Prerequisites: PyTorch Distributed Overview, RPC API documents. This tutorial walks through a simple example of implementing a parameter server using PyTorchโs Distributed RPC framework. The parameter server framework is a paradigm in which a set of servers store parameters, s..." />
<meta property="og:ignore_canonical" content="true" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Implementing a Parameter Server Using Distributed RPC Framework — ํ์ดํ ์น ํ๊ตญ์ด ํํ ๋ฆฌ์ผ (PyTorch tutorials in Korean)</title>
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="../_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.10/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-binder.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-dataframe.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-rendered-html.css" type="text/css" />
<link rel="stylesheet" href="../_static/sphinx-design.5ea377869091fd0449014c60fc090103.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.10.0-beta/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/custom.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/custom2.css" type="text/css" />
<link rel="index" title="์์ธ" href="../genindex.html" />
<link rel="search" title="๊ฒ์" href="../search.html" />
<link rel="next" title="Distributed Pipeline Parallelism Using RPC" href="dist_pipeline_parallel_tutorial.html" />
<link rel="prev" title="Getting Started with Distributed RPC Framework" href="rpc_tutorial.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/katex@0.10.0/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.kr/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.kr/get-started">์์ํ๊ธฐ</a>
</li>
<li class="active">
<a href="https://tutorials.pytorch.kr/">ํํ ๋ฆฌ์ผ</a>
</li>
<li>
<a href="https://pytorch.kr/hub">ํ๋ธ</a>
</li>
<li>
<a href="https://discuss.pytorch.kr/">์ปค๋ฎค๋ํฐ</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">
2.3.1+cu121
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="../search.html" method="get">
<input type="text" name="q" placeholder="Search Tutorials" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption" role="heading"><span class="caption-text">ํ์ดํ ์น(PyTorch) ๋ ์ํผ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../recipes/recipes_index.html">๋ชจ๋ ๋ ์ํผ ๋ณด๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../prototype/prototype_index.html">๋ชจ๋ ํ๋กํ ํ์
๋ ์ํผ ๋ณด๊ธฐ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ์ดํ ์น(PyTorch) ์์ํ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/intro.html">ํ์ดํ ์น(PyTorch) ๊ธฐ๋ณธ ์ตํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/quickstart_tutorial.html">๋น ๋ฅธ ์์(Quickstart)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/tensorqs_tutorial.html">ํ
์(Tensor)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/data_tutorial.html">Dataset๊ณผ DataLoader</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/transforms_tutorial.html">๋ณํ(Transform)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/buildmodel_tutorial.html">์ ๊ฒฝ๋ง ๋ชจ๋ธ ๊ตฌ์ฑํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/autogradqs_tutorial.html"><code class="docutils literal notranslate"><span class="pre">torch.autograd</span></code>๋ฅผ ์ฌ์ฉํ ์๋ ๋ฏธ๋ถ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/optimization_tutorial.html">๋ชจ๋ธ ๋งค๊ฐ๋ณ์ ์ต์ ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/saveloadrun_tutorial.html">๋ชจ๋ธ ์ ์ฅํ๊ณ ๋ถ๋ฌ์ค๊ธฐ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Introduction to PyTorch on YouTube</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt.html">PyTorch ์๊ฐ - YouTube ์๋ฆฌ์ฆ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/introyt1_tutorial.html">PyTorch ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensors_deeper_tutorial.html">Pytorch Tensor ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/autogradyt_tutorial.html">The Fundamentals of Autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/modelsyt_tutorial.html">Building Models with PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensorboardyt_tutorial.html">PyTorch TensorBoard Support</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/trainingyt.html">Training with PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/captumyt.html">Model Understanding with Captum</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ์ดํ ์น(PyTorch) ๋ฐฐ์ฐ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/deep_learning_60min_blitz.html">PyTorch๋ก ๋ฅ๋ฌ๋ํ๊ธฐ: 60๋ถ๋ง์ ๋์ฅ๋ด๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/pytorch_with_examples.html">์์ ๋ก ๋ฐฐ์ฐ๋ ํ์ดํ ์น(PyTorch)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/nn_tutorial.html"><cite>torch.nn</cite> ์ด <em>์ค์ ๋ก</em> ๋ฌด์์ธ๊ฐ์?</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard_tutorial.html">TensorBoard๋ก ๋ชจ๋ธ, ๋ฐ์ดํฐ, ํ์ต ์๊ฐํํ๊ธฐ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">์ด๋ฏธ์ง/๋น๋์ค</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchvision_tutorial.html">TorchVision Object Detection Finetuning Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/transfer_learning_tutorial.html">์ปดํจํฐ ๋น์ (Vision)์ ์ํ ์ ์ดํ์ต(Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/fgsm_tutorial.html">์ ๋์ ์์ ์์ฑ(Adversarial Example Generation)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/dcgan_faces_tutorial.html">DCGAN ํํ ๋ฆฌ์ผ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">๋ฐฐํฌ๋ฅผ ์ํด ๋น์ ํธ๋์คํฌ๋จธ(Vision Transformer) ๋ชจ๋ธ ์ต์ ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="tiatoolbox_tutorial.html">Whole Slide Image Classification Using PyTorch and TIAToolbox</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">์ค๋์ค</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_io_tutorial.html">Audio I/O</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_resampling_tutorial.html">Audio Resampling</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_data_augmentation_tutorial.html">Audio Data Augmentation</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_feature_extractions_tutorial.html">Audio Feature Extractions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_feature_augmentation_tutorial.html">Audio Feature Augmentation</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_datasets_tutorial.html">Audio Datasets</a></li>
<li class="toctree-l1"><a class="reference internal" href="speech_recognition_pipeline_tutorial.html">Speech Recognition with Wav2Vec2</a></li>
<li class="toctree-l1"><a class="reference internal" href="text_to_speech_with_torchaudio.html">Text-to-speech with Tacotron2</a></li>
<li class="toctree-l1"><a class="reference internal" href="forced_alignment_with_torchaudio_tutorial.html">wav2vec2์ ์ด์ฉํ ๊ฐ์ ์ ๋ ฌ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ
์คํธ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/bettertransformer_tutorial.html">Fast Transformer Inference with Better Transformer</a></li>
<li class="toctree-l1"><a class="reference internal" href="char_rnn_classification_tutorial.html">๊ธฐ์ด๋ถํฐ ์์ํ๋ NLP: ๋ฌธ์-๋จ์ RNN์ผ๋ก ์ด๋ฆ ๋ถ๋ฅํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="char_rnn_generation_tutorial.html">๊ธฐ์ด๋ถํฐ ์์ํ๋ NLP: ๋ฌธ์-๋จ์ RNN์ผ๋ก ์ด๋ฆ ์์ฑํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="seq2seq_translation_tutorial.html">๊ธฐ์ด๋ถํฐ ์์ํ๋ NLP: Sequence to Sequence ๋คํธ์ํฌ์ Attention์ ์ด์ฉํ ๋ฒ์ญ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/text_sentiment_ngrams_tutorial.html">torchtext ๋ผ์ด๋ธ๋ฌ๋ฆฌ๋ก ํ
์คํธ ๋ถ๋ฅํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/translation_transformer.html"><code class="docutils literal notranslate"><span class="pre">nn.Transformer</span></code> ์ torchtext๋ก ์ธ์ด ๋ฒ์ญํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/torchtext_custom_dataset_tutorial.html">Preprocess custom text dataset using Torchtext</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">๋ฐฑ์๋</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/onnx/intro_onnx.html">Introduction to ONNX</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">๊ฐํํ์ต</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="reinforcement_q_learning.html">๊ฐํ ํ์ต (DQN) ํํ ๋ฆฌ์ผ</a></li>
<li class="toctree-l1"><a class="reference internal" href="reinforcement_ppo.html">Reinforcement Learning (PPO) with TorchRL Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="mario_rl_tutorial.html">๋ง๋ฆฌ์ค ๊ฒ์ RL ์์ด์ ํธ๋ก ํ์ตํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/pendulum.html">Pendulum: Writing your environment and transforms with TorchRL</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch ๋ชจ๋ธ์ ํ๋ก๋์
ํ๊ฒฝ์ ๋ฐฐํฌํ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/onnx/intro_onnx.html">Introduction to ONNX</a></li>
<li class="toctree-l1"><a class="reference internal" href="flask_rest_api_tutorial.html">Flask๋ฅผ ์ฌ์ฉํ์ฌ Python์์ PyTorch๋ฅผ REST API๋ก ๋ฐฐํฌํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/Intro_to_TorchScript_tutorial.html">TorchScript ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_export.html">C++์์ TorchScript ๋ชจ๋ธ ๋ก๋ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/super_resolution_with_onnxruntime.html">(์ ํ) PyTorch ๋ชจ๋ธ์ ONNX์ผ๋ก ๋ณํํ๊ณ ONNX ๋ฐํ์์์ ์คํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="realtime_rpi.html">Raspberry Pi 4 ์์ ์ค์๊ฐ ์ถ๋ก (Inference) (30fps!)</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch ํ๋กํ์ผ๋ง</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch ๋ชจ๋ ํ๋กํ์ผ๋งํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hta_intro_tutorial.html">Introduction to Holistic Trace Analysis</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hta_trace_diff_tutorial.html">Trace Diff using Holistic Trace Analysis</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Code Transforms with FX</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="fx_conv_bn_fuser.html">(๋ฒ ํ) FX์์ ํฉ์ฑ๊ณฑ/๋ฐฐ์น ์ ๊ทํ(Convolution/Batch Norm) ๊ฒฐํฉ๊ธฐ(Fuser) ๋ง๋ค๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx_profiling_tutorial.html">(beta) Building a Simple CPU Performance Profiler with FX</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ๋ก ํธ์๋ API</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="memory_format_tutorial.html">(๋ฒ ํ) PyTorch๋ฅผ ์ฌ์ฉํ Channels Last ๋ฉ๋ชจ๋ฆฌ ํ์</a></li>
<li class="toctree-l1"><a class="reference internal" href="forward_ad_usage.html">Forward-mode Automatic Differentiation (Beta)</a></li>
<li class="toctree-l1"><a class="reference internal" href="jacobians_hessians.html">Jacobians, Hessians, hvp, vhp, and more: composing function transforms</a></li>
<li class="toctree-l1"><a class="reference internal" href="ensembling.html">๋ชจ๋ธ ์์๋ธ</a></li>
<li class="toctree-l1"><a class="reference internal" href="per_sample_grads.html">Per-sample-gradients</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_frontend.html">PyTorch C++ ํ๋ก ํธ์๋ ์ฌ์ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch-script-parallelism.html">TorchScript์ ๋์ ๋ณ๋ ฌ ์ฒ๋ฆฌ(Dynamic Parallelism)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_autograd.html">C++ ํ๋ก ํธ์๋์ ์๋ ๋ฏธ๋ถ (autograd)</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch ํ์ฅํ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="custom_function_double_backward_tutorial.html">Double Backward with Custom Functions</a></li>
<li class="toctree-l1"><a class="reference internal" href="custom_function_conv_bn_tutorial.html">Fusing Convolution and Batch Norm using Custom Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_extension.html">Custom C++ and CUDA Extensions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch_script_custom_ops.html">Extending TorchScript with Custom C++ Operators</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch_script_custom_classes.html">์ปค์คํ
C++ ํด๋์ค๋ก TorchScript ํ์ฅํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/dispatcher.html">Registering a Dispatched Operator in C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/extend_dispatcher.html">Extending dispatcher for a new backend in C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/privateuseone.html">Facilitating New Backend Integration by PrivateUse1</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">๋ชจ๋ธ ์ต์ ํ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch ๋ชจ๋ ํ๋กํ์ผ๋งํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard_profiler_tutorial.html">ํ
์๋ณด๋๋ฅผ ์ด์ฉํ ํ์ดํ ์น ํ๋กํ์ผ๋ฌ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hyperparameter_tuning_tutorial.html">Ray Tune์ ์ฌ์ฉํ ํ์ดํผํ๋ผ๋ฏธํฐ ํ๋</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">๋ฐฐํฌ๋ฅผ ์ํด ๋น์ ํธ๋์คํฌ๋จธ(Vision Transformer) ๋ชจ๋ธ ์ต์ ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="parametrizations.html">Parametrizations Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="pruning_tutorial.html">๊ฐ์ง์น๊ธฐ ๊ธฐ๋ฒ(Pruning) ํํ ๋ฆฌ์ผ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/dynamic_quantization_tutorial.html">(๋ฒ ํ) LSTM ๊ธฐ๋ฐ ๋จ์ด ๋จ์ ์ธ์ด ๋ชจ๋ธ์ ๋์ ์์ํ</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamic_quantization_bert_tutorial.html">(๋ฒ ํ) BERT ๋ชจ๋ธ ๋์ ์์ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantized_transfer_learning_tutorial.html">(๋ฒ ํ) ์ปดํจํฐ ๋น์ ํํ ๋ฆฌ์ผ์ ์ํ ์์ํ๋ ์ ์ดํ์ต(Quantized Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/static_quantization_tutorial.html">(๋ฒ ํ) PyTorch์์ Eager Mode๋ฅผ ์ด์ฉํ ์ ์ ์์ํ</a></li>
<li class="toctree-l1"><a class="reference internal" href="torchserve_with_ipex.html">Grokking PyTorch Intel CPU performance from first principles</a></li>
<li class="toctree-l1"><a class="reference internal" href="torchserve_with_ipex_2.html">Grokking PyTorch Intel CPU performance from first principles (Part 2)</a></li>
<li class="toctree-l1"><a class="reference internal" href="nvfuser_intro_tutorial.html">Getting Started - Accelerate Your Scripts with nvFuser</a></li>
<li class="toctree-l1"><a class="reference internal" href="ax_multiobjective_nas_tutorial.html">Multi-Objective NAS with Ax</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_compile_tutorial.html">Introduction to <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code></a></li>
<li class="toctree-l1"><a class="reference internal" href="inductor_debug_cpu.html">Inductor CPU backend debugging and profiling</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html">(Beta) Scaled Dot Product Attention (SDPA)๋ก ๊ณ ์ฑ๋ฅ ํธ๋์คํฌ๋จธ(Transformers) ๊ตฌํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#torch-compile-sdpa"><code class="docutils literal notranslate"><span class="pre">torch.compile</span></code> ๊ณผ ํจ๊ป SDPA ์ฌ์ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#sdpa-atteition-bias">SDPA๋ฅผ <code class="docutils literal notranslate"><span class="pre">atteition.bias</span></code> ํ์ ํด๋์ค์ ์ฌ์ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#id8">๊ฒฐ๋ก </a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/knowledge_distillation_tutorial.html">Knowledge Distillation Tutorial</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">๋ณ๋ ฌ ๋ฐ ๋ถ์ฐ ํ์ต</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../distributed/home.html">Distributed and Parallel Training Tutorials</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/dist_overview.html">PyTorch Distributed Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/ddp_series_intro.html">Distributed Data Parallel in PyTorch - Video Tutorials</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_parallel_tutorial.html">๋จ์ผ ๋จธ์ ์ ์ฌ์ฉํ ๋ชจ๋ธ ๋ณ๋ ฌํ ๋ชจ๋ฒ ์ฌ๋ก</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_tutorial.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ์์ํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="dist_tuto.html">PyTorch๋ก ๋ถ์ฐ ์ดํ๋ฆฌ์ผ์ด์
๊ฐ๋ฐํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="FSDP_tutorial.html">Getting Started with Fully Sharded Data Parallel(FSDP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="FSDP_adavnced_tutorial.html">Advanced Model Training with Fully Sharded Data Parallel (FSDP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="TP_tutorial.html">Large Scale Transformer model training with Tensor Parallel (TP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="process_group_cpp_extension_tutorial.html">Cpp ํ์ฅ์ ์ฌ์ฉํ ํ๋ก์ธ์ค ๊ทธ๋ฃน ๋ฐฑ์๋ ์ฌ์ฉ์ ์ ์</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_tutorial.html">Getting Started with Distributed RPC Framework</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Implementing a Parameter Server Using Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="dist_pipeline_parallel_tutorial.html">Distributed Pipeline Parallelism Using RPC</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_async_execution.html">Implementing Batch RPC Processing Using Asynchronous Executions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/rpc_ddp_tutorial.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ(DDP)๊ณผ ๋ถ์ฐ RPC ํ๋ ์์ํฌ ๊ฒฐํฉ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/ddp_pipeline.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ์ ๋ณ๋ ฌ ์ฒ๋ฆฌ ํ์ดํ๋ผ์ธ์ ์ฌ์ฉํ ํธ๋์คํฌ๋จธ ๋ชจ๋ธ ํ์ต</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/generic_join.html">Distributed Training with Uneven Inputs Using the Join Context Manager</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Edge with ExecuTorch</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/tutorials/export-to-executorch-tutorial.html">Exporting to ExecuTorch Tutorial</a></li>
<li class="toctree-l1"><a class="reference external" href=" https://pytorch.org/executorch/stable/running-a-model-cpp-tutorial.html">Running an ExecuTorch Model in C++ Tutorial</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/tutorials/sdk-integration-tutorial.html">Using the ExecuTorch SDK to Profile a Model</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/demo-apps-ios.html">Building an ExecuTorch iOS Demo App</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/demo-apps-android.html">Building an ExecuTorch Android Demo App</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/examples-end-to-end-to-lower-model-to-delegate.html">Lowering a Model as a Delegate</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">์ถ์ฒ ์์คํ
</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchrec_tutorial.html">TorchRec ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/sharding.html">Exploring TorchRec sharding</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Multimodality</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/flava_finetuning_tutorial.html">TorchMultimodal ํํ ๋ฆฌ์ผ: FLAVA ๋ฏธ์ธ์กฐ์ </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">
Tutorials
</a> >
</li>
<li>Implementing a Parameter Server Using Distributed RPC Framework</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/intermediate/rpc_param_server_tutorial.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="pytorch-call-to-action-links">
<div id="tutorial-type">intermediate/rpc_param_server_tutorial</div>
<div id="google-colab-link">
<img class="call-to-action-img" src="../_static/images/pytorch-colab.svg"/>
<div class="call-to-action-desktop-view">Run in Google Colab</div>
<div class="call-to-action-mobile-view">Colab</div>
</div>
<div id="download-notebook-link">
<img class="call-to-action-notebook-img" src="../_static/images/pytorch-download.svg"/>
<div class="call-to-action-desktop-view">Download Notebook</div>
<div class="call-to-action-mobile-view">Notebook</div>
</div>
<div id="github-view-link">
<img class="call-to-action-img" src="../_static/images/pytorch-github.svg"/>
<div class="call-to-action-desktop-view">View on GitHub</div>
<div class="call-to-action-mobile-view">GitHub</div>
</div>
</div>
<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="implementing-a-parameter-server-using-distributed-rpc-framework">
<h1>Implementing a Parameter Server Using Distributed RPC Framework<a class="headerlink" href="#implementing-a-parameter-server-using-distributed-rpc-framework" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h1>
<p><strong>Author</strong>: <a class="reference external" href="https://github.com/rohan-varma">Rohan Varma</a></p>
<div class="admonition note">
<p class="admonition-title">์ฐธ๊ณ </p>
<p><a class="reference internal" href="../_images/pencil-16.png"><img alt="edit" src="../_images/pencil-16.png" style="width: 16px; height: 16px;" /></a> View and edit this tutorial in <a class="reference external" href="https://github.com/pytorch/tutorials/blob/main/intermediate_source/rpc_param_server_tutorial.rst">github</a>.</p>
</div>
<p>Prerequisites:</p>
<ul class="simple">
<li><p><a class="reference external" href="../beginner/dist_overview.html">PyTorch Distributed Overview</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/docs/master/rpc.html">RPC API documents</a></p></li>
</ul>
<p>This tutorial walks through a simple example of implementing a parameter server using PyTorchโs <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html">Distributed RPC framework</a>. The parameter server framework is a paradigm in which a set of servers store parameters, such as large embedding tables, and several trainers query the parameter servers in order to retrieve the most up to date parameters. These trainers can run a training loop locally and occasionally synchronize with the parameter server to get the latest parameters. For more reading on the parameter server approach, check out <a class="reference external" href="https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf">this paper</a>.</p>
<p>Using the Distributed RPC Framework, weโll build an example where multiple trainers use RPC to communicate with the same parameter server and use <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html#torch.distributed.rpc.RRef">RRef</a> to access states on the remote parameter server instance. Each trainer will launch its dedicated backward pass in a distributed fashion through stitching of the autograd graph across multiple nodes using distributed autograd.</p>
<p><strong>Note</strong>: This tutorial covers the use of the Distributed RPC Framework, which is useful for splitting a model onto multiple machines, or for implementing a parameter-server training strategy where network trainers fetch parameters hosted on a different machine. If instead you are looking for replicating your model across many GPUs, please see the <a class="reference external" href="https://tutorials.pytorch.kr/intermediate/ddp_tutorial.html">Distributed Data Parallel tutorial</a>. There is also another <a class="reference external" href="https://tutorials.pytorch.kr/intermediate/rpc_tutorial.html">RPC tutorial</a> that covers reinforcement learning and RNN use cases.</p>
<p>Letโs start with the familiar: importing our required modules and defining a simple ConvNet that will train on the MNIST dataset. The below network is largely adopted from the network defined in the <a class="reference external" href="https://github.com/pytorch/examples/tree/master/mnist">pytorch/examples repo</a>.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">argparse</span>
<span class="kn">import</span> <span class="nn">os</span>
<span class="kn">import</span> <span class="nn">time</span>
<span class="kn">from</span> <span class="nn">threading</span> <span class="kn">import</span> <span class="n">Lock</span>
<span class="kn">import</span> <span class="nn">torch</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="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="kn">import</span> <span class="nn">torch.multiprocessing</span> <span class="k">as</span> <span class="nn">mp</span>
<span class="kn">import</span> <span class="nn">torch.nn</span> <span class="k">as</span> <span class="nn">nn</span>
<span class="kn">import</span> <span class="nn">torch.nn.functional</span> <span class="k">as</span> <span class="nn">F</span>
<span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">optim</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="kn">from</span> <span class="nn">torchvision</span> <span class="kn">import</span> <span class="n">datasets</span><span class="p">,</span> <span class="n">transforms</span>
<span class="c1"># --------- MNIST Network to train, from pytorch/examples -----</span>
<span class="k">class</span> <span class="nc">Net</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">num_gpus</span><span class="o">=</span><span class="mi">0</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">Net</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Using </span><span class="si">{</span><span class="n">num_gpus</span><span class="si">}</span><span class="s2"> GPUs to train"</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">num_gpus</span> <span class="o">=</span> <span class="n">num_gpus</span>
<span class="n">device</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">device</span><span class="p">(</span>
<span class="s2">"cuda:0"</span> <span class="k">if</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">is_available</span><span class="p">()</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_gpus</span> <span class="o">></span> <span class="mi">0</span> <span class="k">else</span> <span class="s2">"cpu"</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Putting first 2 convs on </span><span class="si">{</span><span class="nb">str</span><span class="p">(</span><span class="n">device</span><span class="p">)</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="c1"># Put conv layers on the first cuda device, or CPU if no cuda device</span>
<span class="bp">self</span><span class="o">.</span><span class="n">conv1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">32</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">conv2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</span><span class="p">(</span><span class="mi">32</span><span class="p">,</span> <span class="mi">64</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span>
<span class="c1"># Put rest of the network on the 2nd cuda device, if there is one</span>
<span class="k">if</span> <span class="s2">"cuda"</span> <span class="ow">in</span> <span class="nb">str</span><span class="p">(</span><span class="n">device</span><span class="p">)</span> <span class="ow">and</span> <span class="n">num_gpus</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
<span class="n">device</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">device</span><span class="p">(</span><span class="s2">"cuda:1"</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Putting rest of layers on </span><span class="si">{</span><span class="nb">str</span><span class="p">(</span><span class="n">device</span><span class="p">)</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">dropout1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Dropout2d</span><span class="p">(</span><span class="mf">0.25</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">dropout2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Dropout2d</span><span class="p">(</span><span class="mf">0.5</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">fc1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">9216</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">fc2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">128</span><span class="p">,</span> <span class="mi">10</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv1</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv2</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">max_pool2d</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">dropout1</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">flatten</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="c1"># Move tensor to next device if necessary</span>
<span class="n">next_device</span> <span class="o">=</span> <span class="nb">next</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">fc1</span><span class="o">.</span><span class="n">parameters</span><span class="p">())</span><span class="o">.</span><span class="n">device</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">x</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">next_device</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">fc1</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">dropout2</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">fc2</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">log_softmax</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="k">return</span> <span class="n">output</span>
</pre></div>
</div>
<p>Next, letโs define some helper functions that will be useful for the rest of our script. The following uses <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html#torch.distributed.rpc.rpc_sync">rpc_sync</a> and <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html#torch.distributed.rpc.RRef">RRef</a> in order to define a function that invokes a given method on an object living on a remote node. Below, our handle to the remote object is given by the <code class="docutils literal notranslate"><span class="pre">rref</span></code> argument, and we run it on its owning node: <code class="docutils literal notranslate"><span class="pre">rref.owner()</span></code>. On the caller node, we run this command synchronously through the use of <code class="docutils literal notranslate"><span class="pre">rpc_sync</span></code>, meaning that we will block until a response is received.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># --------- Helper Methods --------------------</span>
<span class="c1"># On the local node, call a method with first arg as the value held by the</span>
<span class="c1"># RRef. Other args are passed in as arguments to the function called.</span>
<span class="c1"># Useful for calling instance methods. method could be any matching function, including</span>
<span class="c1"># class methods.</span>
<span class="k">def</span> <span class="nf">call_method</span><span class="p">(</span><span class="n">method</span><span class="p">,</span> <span class="n">rref</span><span class="p">,</span> <span class="o">*</span><span class="n">args</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
<span class="k">return</span> <span class="n">method</span><span class="p">(</span><span class="n">rref</span><span class="o">.</span><span class="n">local_value</span><span class="p">(),</span> <span class="o">*</span><span class="n">args</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">)</span>
<span class="c1"># Given an RRef, return the result of calling the passed in method on the value</span>
<span class="c1"># held by the RRef. This call is done on the remote node that owns</span>
<span class="c1"># the RRef and passes along the given argument.</span>
<span class="c1"># Example: If the value held by the RRef is of type Foo, then</span>
<span class="c1"># remote_method(Foo.bar, rref, arg1, arg2) is equivalent to calling</span>
<span class="c1"># <foo_instance>.bar(arg1, arg2) on the remote node and getting the result</span>
<span class="c1"># back.</span>
<span class="k">def</span> <span class="nf">remote_method</span><span class="p">(</span><span class="n">method</span><span class="p">,</span> <span class="n">rref</span><span class="p">,</span> <span class="o">*</span><span class="n">args</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
<span class="n">args</span> <span class="o">=</span> <span class="p">[</span><span class="n">method</span><span class="p">,</span> <span class="n">rref</span><span class="p">]</span> <span class="o">+</span> <span class="nb">list</span><span class="p">(</span><span class="n">args</span><span class="p">)</span>
<span class="k">return</span> <span class="n">rpc</span><span class="o">.</span><span class="n">rpc_sync</span><span class="p">(</span><span class="n">rref</span><span class="o">.</span><span class="n">owner</span><span class="p">(),</span> <span class="n">call_method</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="n">args</span><span class="p">,</span> <span class="n">kwargs</span><span class="o">=</span><span class="n">kwargs</span><span class="p">)</span>
</pre></div>
</div>
<p>Now, weโre ready to define our parameter server. We will subclass <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code> and save a handle to our network defined above. Weโll also save an input device which will be the device our input is transferred to before invoking the model.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># --------- Parameter Server --------------------</span>
<span class="k">class</span> <span class="nc">ParameterServer</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">num_gpus</span><span class="o">=</span><span class="mi">0</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">Net</span><span class="p">(</span><span class="n">num_gpus</span><span class="o">=</span><span class="n">num_gpus</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="n">model</span>
<span class="bp">self</span><span class="o">.</span><span class="n">input_device</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">device</span><span class="p">(</span>
<span class="s2">"cuda:0"</span> <span class="k">if</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">is_available</span><span class="p">()</span> <span class="ow">and</span> <span class="n">num_gpus</span> <span class="o">></span> <span class="mi">0</span> <span class="k">else</span> <span class="s2">"cpu"</span><span class="p">)</span>
</pre></div>
</div>
<p>Next, weโll define our forward pass. Note that regardless of the device of the model output, we move the output to CPU, as the Distributed RPC Framework currently only supports sending CPU tensors over RPC. We have intentionally disabled sending CUDA tensors over RPC due to the potential for different devices (CPU/GPU) on on the caller/callee, but may support this in future releases.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">ParameterServer</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="o">...</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">inp</span><span class="p">):</span>
<span class="n">inp</span> <span class="o">=</span> <span class="n">inp</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">input_device</span><span class="p">)</span>
<span class="n">out</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">(</span><span class="n">inp</span><span class="p">)</span>
<span class="c1"># This output is forwarded over RPC, which as of 1.5.0 only accepts CPU tensors.</span>
<span class="c1"># Tensors must be moved in and out of GPU memory due to this.</span>
<span class="n">out</span> <span class="o">=</span> <span class="n">out</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="s2">"cpu"</span><span class="p">)</span>
<span class="k">return</span> <span class="n">out</span>
</pre></div>
</div>
<p>Next, weโll define a few miscellaneous functions useful for training and verification purposes. The first, <code class="docutils literal notranslate"><span class="pre">get_dist_gradients</span></code>, will take in a Distributed Autograd context ID and call into the <code class="docutils literal notranslate"><span class="pre">dist_autograd.get_gradients</span></code> API in order to retrieve gradients computed by distributed autograd. More information can be found in the <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html#distributed-autograd-framework">distributed autograd documentation</a>. Note that we also iterate through the resulting dictionary and convert each tensor to a CPU tensor, as the framework currently only supports sending tensors over RPC. Next, <code class="docutils literal notranslate"><span class="pre">get_param_rrefs</span></code> will iterate through our model parameters and wrap them as a (local) <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html#torch.distributed.rpc.RRef">RRef</a>. This method will be invoked over RPC by trainer nodes and will return a list of the parameters to be optimized. This is required as input to the <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html#module-torch.distributed.optim">Distributed Optimizer</a>, which requires all parameters it must optimize as a list of <code class="docutils literal notranslate"><span class="pre">RRef</span></code>s.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Use dist autograd to retrieve gradients accumulated for this model.</span>
<span class="c1"># Primarily used for verification.</span>
<span class="k">def</span> <span class="nf">get_dist_gradients</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">cid</span><span class="p">):</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">cid</span><span class="p">)</span>
<span class="c1"># This output is forwarded over RPC, which as of 1.5.0 only accepts CPU tensors.</span>
<span class="c1"># Tensors must be moved in and out of GPU memory due to this.</span>
<span class="n">cpu_grads</span> <span class="o">=</span> <span class="p">{}</span>
<span class="k">for</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span> <span class="ow">in</span> <span class="n">grads</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="n">k_cpu</span><span class="p">,</span> <span class="n">v_cpu</span> <span class="o">=</span> <span class="n">k</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="s2">"cpu"</span><span class="p">),</span> <span class="n">v</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="s2">"cpu"</span><span class="p">)</span>
<span class="n">cpu_grads</span><span class="p">[</span><span class="n">k_cpu</span><span class="p">]</span> <span class="o">=</span> <span class="n">v_cpu</span>
<span class="k">return</span> <span class="n">cpu_grads</span>
<span class="c1"># Wrap local parameters in a RRef. Needed for building the</span>
<span class="c1"># DistributedOptimizer which optimizes paramters remotely.</span>
<span class="k">def</span> <span class="nf">get_param_rrefs</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">param_rrefs</span> <span class="o">=</span> <span class="p">[</span><span class="n">rpc</span><span class="o">.</span><span class="n">RRef</span><span class="p">(</span><span class="n">param</span><span class="p">)</span> <span class="k">for</span> <span class="n">param</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">()]</span>
<span class="k">return</span> <span class="n">param_rrefs</span>
</pre></div>
</div>
<p>Finally, weโll create methods to initialize our parameter server. Note that there will only be one instance of a parameter server across all processes, and all trainers will talk to the same parameter server and update the same stored model. As seen in <code class="docutils literal notranslate"><span class="pre">run_parameter_server</span></code>, the server itself does not take any independent actions; it waits for requests from trainers (which are yet to be defined) and responds to them by running the requested function.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># The global parameter server instance.</span>
<span class="n">param_server</span> <span class="o">=</span> <span class="kc">None</span>
<span class="c1"># A lock to ensure we only have one parameter server.</span>
<span class="n">global_lock</span> <span class="o">=</span> <span class="n">Lock</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">get_parameter_server</span><span class="p">(</span><span class="n">num_gpus</span><span class="o">=</span><span class="mi">0</span><span class="p">):</span>
<span class="w"> </span><span class="sd">"""</span>
<span class="sd"> Returns a singleton parameter server to all trainer processes</span>
<span class="sd"> """</span>
<span class="k">global</span> <span class="n">param_server</span>
<span class="c1"># Ensure that we get only one handle to the ParameterServer.</span>
<span class="k">with</span> <span class="n">global_lock</span><span class="p">:</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">param_server</span><span class="p">:</span>
<span class="c1"># construct it once</span>
<span class="n">param_server</span> <span class="o">=</span> <span class="n">ParameterServer</span><span class="p">(</span><span class="n">num_gpus</span><span class="o">=</span><span class="n">num_gpus</span><span class="p">)</span>
<span class="k">return</span> <span class="n">param_server</span>
<span class="k">def</span> <span class="nf">run_parameter_server</span><span class="p">(</span><span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="p">):</span>
<span class="c1"># The parameter server just acts as a host for the model and responds to</span>
<span class="c1"># requests from trainers.</span>
<span class="c1"># rpc.shutdown() will wait for all workers to complete by default, which</span>
<span class="c1"># in this case means that the parameter server will wait for all trainers</span>
<span class="c1"># to complete, and then exit.</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"PS master initializing RPC"</span><span class="p">)</span>
<span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s2">"parameter_server"</span><span class="p">,</span> <span class="n">rank</span><span class="o">=</span><span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="o">=</span><span class="n">world_size</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"RPC initialized! Running parameter server..."</span><span class="p">)</span>
<span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"RPC shutdown on parameter server."</span><span class="p">)</span>
</pre></div>
</div>
<p>Note that above, <code class="docutils literal notranslate"><span class="pre">rpc.shutdown()</span></code> will not immediately shut down the Parameter Server. Instead, it will wait for all workers (trainers in this case) to also call into <code class="docutils literal notranslate"><span class="pre">rpc.shutdown()</span></code>. This gives us the guarantee that the parameter server will not go offline before all trainers (yet to be define) have completed their training process.</p>
<p>Next, weโll define our <code class="docutils literal notranslate"><span class="pre">TrainerNet</span></code> class. This will also be a subclass of <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code>, and our <code class="docutils literal notranslate"><span class="pre">__init__</span></code> method will use the <code class="docutils literal notranslate"><span class="pre">rpc.remote</span></code> API to obtain an RRef, or Remote Reference, to our parameter server. Note that here we are not copying the parameter server to our local process, instead, we can think of <code class="docutils literal notranslate"><span class="pre">self.param_server_rref</span></code> as a distributed shared pointer to the parameter server that lives on a separate process.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># --------- Trainers --------------------</span>
<span class="c1"># nn.Module corresponding to the network trained by this trainer. The</span>
<span class="c1"># forward() method simply invokes the network on the given parameter</span>
<span class="c1"># server.</span>
<span class="k">class</span> <span class="nc">TrainerNet</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">num_gpus</span><span class="o">=</span><span class="mi">0</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">num_gpus</span> <span class="o">=</span> <span class="n">num_gpus</span>
<span class="bp">self</span><span class="o">.</span><span class="n">param_server_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="s2">"parameter_server"</span><span class="p">,</span> <span class="n">get_parameter_server</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">num_gpus</span><span class="p">,))</span>
</pre></div>
</div>
<p>Next, weโll define a method called <code class="docutils literal notranslate"><span class="pre">get_global_param_rrefs</span></code>. To motivate the need for this method, it is worth it to read through the documentation on <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html#module-torch.distributed.optim">DistributedOptimizer</a>, specifically the API signature. The optimizer must be passed a list of <code class="docutils literal notranslate"><span class="pre">RRef</span></code>s corresponding to the remote parameters to be optimized, so here we obtain the necessary <code class="docutils literal notranslate"><span class="pre">RRef</span></code>s. Since the only remote worker that a given <code class="docutils literal notranslate"><span class="pre">TrainerNet</span></code> interacts with is the <code class="docutils literal notranslate"><span class="pre">ParameterServer</span></code>, we simply invoke a <code class="docutils literal notranslate"><span class="pre">remote_method</span></code> on the <code class="docutils literal notranslate"><span class="pre">ParameterServer</span></code>. We use the <code class="docutils literal notranslate"><span class="pre">get_param_rrefs</span></code> method which we defined in the <code class="docutils literal notranslate"><span class="pre">ParameterServer</span></code> class. This method will return a list of <code class="docutils literal notranslate"><span class="pre">RRef</span></code>s to the parameters that need to be optimized. Note that in this case our <code class="docutils literal notranslate"><span class="pre">TrainerNet</span></code> does not define its own paramaters; if it did, we would need to wrap each parameter in an <code class="docutils literal notranslate"><span class="pre">RRef</span></code> as well and include it into our input to <code class="docutils literal notranslate"><span class="pre">DistributedOptimizer</span></code>.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">TrainerNet</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="o">...</span>
<span class="k">def</span> <span class="nf">get_global_param_rrefs</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">remote_params</span> <span class="o">=</span> <span class="n">remote_method</span><span class="p">(</span>
<span class="n">ParameterServer</span><span class="o">.</span><span class="n">get_param_rrefs</span><span class="p">,</span>
<span class="bp">self</span><span class="o">.</span><span class="n">param_server_rref</span><span class="p">)</span>
<span class="k">return</span> <span class="n">remote_params</span>
</pre></div>
</div>
<p>Now, weโre ready to define our <code class="docutils literal notranslate"><span class="pre">forward</span></code> method, which will invoke (synchronous) RPC to run the forward pass of the network defined on the <code class="docutils literal notranslate"><span class="pre">ParameterServer</span></code>. Note that we pass in <code class="docutils literal notranslate"><span class="pre">self.param_server_rref</span></code>, which is a remote handle to our <code class="docutils literal notranslate"><span class="pre">ParameterServer</span></code>, to our RPC call. This call will send an RPC to the node on which our <code class="docutils literal notranslate"><span class="pre">ParameterServer</span></code> is running, invoke the <code class="docutils literal notranslate"><span class="pre">forward</span></code> pass, and return the <code class="docutils literal notranslate"><span class="pre">Tensor</span></code> corresponding to the modelโs output.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">TrainerNet</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="o">...</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">model_output</span> <span class="o">=</span> <span class="n">remote_method</span><span class="p">(</span>
<span class="n">ParameterServer</span><span class="o">.</span><span class="n">forward</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">param_server_rref</span><span class="p">,</span> <span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model_output</span>
</pre></div>
</div>
<p>With our trainer fully defined, itโs now time to write our neural network training loop that will create our network and optimizer, run some inputs through the network and compute the loss. The training loop looks a lot like that of a local training program, with some modifications due to the nature of our network being distributed across machines.</p>
<p>Below, we initialize our <code class="docutils literal notranslate"><span class="pre">TrainerNet</span></code> and build a <code class="docutils literal notranslate"><span class="pre">DistributedOptimizer</span></code>. Note that as mentioned above, we must pass in all of the global (across all nodes participating in distributed training) parameters that we want to be optimized. In addition, we pass in the local optimizer to be used, in this case, SGD. Note that we can configure the underlying optimizer algorithm in the same way as creating a local optimizer - all arguments for <code class="docutils literal notranslate"><span class="pre">optimizer.SGD</span></code> will be forwarded properly. As an example, we pass in a custom learning rate that will be used as the learning rate for all local optimizers.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">run_training_loop</span><span class="p">(</span><span class="n">rank</span><span class="p">,</span> <span class="n">num_gpus</span><span class="p">,</span> <span class="n">train_loader</span><span class="p">,</span> <span class="n">test_loader</span><span class="p">):</span>
<span class="c1"># Runs the typical nueral network forward + backward + optimizer step, but</span>
<span class="c1"># in a distributed fashion.</span>
<span class="n">net</span> <span class="o">=</span> <span class="n">TrainerNet</span><span class="p">(</span><span class="n">num_gpus</span><span class="o">=</span><span class="n">num_gpus</span><span class="p">)</span>
<span class="c1"># Build DistributedOptimizer.</span>
<span class="n">param_rrefs</span> <span class="o">=</span> <span class="n">net</span><span class="o">.</span><span class="n">get_global_param_rrefs</span><span class="p">()</span>
<span class="n">opt</span> <span class="o">=</span> <span class="n">DistributedOptimizer</span><span class="p">(</span><span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">,</span> <span class="n">param_rrefs</span><span class="p">,</span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.03</span><span class="p">)</span>
</pre></div>
</div>
<p>Next, we define our main training loop. We loop through iterables given by PyTorchโs <a class="reference external" href="https://pytorch.org/docs/stable/data.html">DataLoader</a>. Before writing our typical forward/backward/optimizer loop, we first wrap the logic within a <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html#torch.distributed.autograd.context">Distributed Autograd context</a>. Note that this is needed to record RPCs invoked in the modelโs forward pass, so that an appropriate graph can be constructed which includes all participating distributed workers in the backward pass. The distributed autograd context returns a <code class="docutils literal notranslate"><span class="pre">context_id</span></code> which serves as an identifier for accumulating and optimizing gradients corresponding to a particular iteration.</p>
<p>As opposed to calling the typical <code class="docutils literal notranslate"><span class="pre">loss.backward()</span></code> which would kick off the backward pass on this local worker, we call <code class="docutils literal notranslate"><span class="pre">dist_autograd.backward()</span></code> and pass in our context_id as well as <code class="docutils literal notranslate"><span class="pre">loss</span></code>, which is the root at which we want the backward pass to begin. In addition, we pass this <code class="docutils literal notranslate"><span class="pre">context_id</span></code> into our optimizer call, which is required to be able to look up the corresponding gradients computed by this particular backwards pass across all nodes.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">run_training_loop</span><span class="p">(</span><span class="n">rank</span><span class="p">,</span> <span class="n">num_gpus</span><span class="p">,</span> <span class="n">train_loader</span><span class="p">,</span> <span class="n">test_loader</span><span class="p">):</span>
<span class="o">...</span>
<span class="k">for</span> <span class="n">i</span><span class="p">,</span> <span class="p">(</span><span class="n">data</span><span class="p">,</span> <span class="n">target</span><span class="p">)</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">train_loader</span><span class="p">):</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">cid</span><span class="p">:</span>
<span class="n">model_output</span> <span class="o">=</span> <span class="n">net</span><span class="p">(</span><span class="n">data</span><span class="p">)</span>
<span class="n">target</span> <span class="o">=</span> <span class="n">target</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">model_output</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
<span class="n">loss</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">nll_loss</span><span class="p">(</span><span class="n">model_output</span><span class="p">,</span> <span class="n">target</span><span class="p">)</span>
<span class="k">if</span> <span class="n">i</span> <span class="o">%</span> <span class="mi">5</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Rank </span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s2"> training batch </span><span class="si">{</span><span class="n">i</span><span class="si">}</span><span class="s2"> loss </span><span class="si">{</span><span class="n">loss</span><span class="o">.</span><span class="n">item</span><span class="p">()</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="n">dist_autograd</span><span class="o">.</span><span class="n">backward</span><span class="p">(</span><span class="n">cid</span><span class="p">,</span> <span class="p">[</span><span class="n">loss</span><span class="p">])</span>
<span class="c1"># Ensure that dist autograd ran successfully and gradients were</span>
<span class="c1"># returned.</span>
<span class="k">assert</span> <span class="n">remote_method</span><span class="p">(</span>
<span class="n">ParameterServer</span><span class="o">.</span><span class="n">get_dist_gradients</span><span class="p">,</span>
<span class="n">net</span><span class="o">.</span><span class="n">param_server_rref</span><span class="p">,</span>
<span class="n">cid</span><span class="p">)</span> <span class="o">!=</span> <span class="p">{}</span>
<span class="n">opt</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">cid</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Training complete!"</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Getting accuracy...."</span><span class="p">)</span>
<span class="n">get_accuracy</span><span class="p">(</span><span class="n">test_loader</span><span class="p">,</span> <span class="n">net</span><span class="p">)</span>
</pre></div>
</div>
<p>The following simply computes the accuracy of our model after weโre done training, much like a traditional local model. However, note that the <code class="docutils literal notranslate"><span class="pre">net</span></code> we pass into this function above is an instance of <code class="docutils literal notranslate"><span class="pre">TrainerNet</span></code> and therefore the forward pass invokes RPC in a transparent fashion.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">get_accuracy</span><span class="p">(</span><span class="n">test_loader</span><span class="p">,</span> <span class="n">model</span><span class="p">):</span>
<span class="n">model</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="n">correct_sum</span> <span class="o">=</span> <span class="mi">0</span>
<span class="c1"># Use GPU to evaluate if possible</span>
<span class="n">device</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">device</span><span class="p">(</span><span class="s2">"cuda:0"</span> <span class="k">if</span> <span class="n">model</span><span class="o">.</span><span class="n">num_gpus</span> <span class="o">></span> <span class="mi">0</span>
<span class="ow">and</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">is_available</span><span class="p">()</span> <span class="k">else</span> <span class="s2">"cpu"</span><span class="p">)</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">():</span>
<span class="k">for</span> <span class="n">i</span><span class="p">,</span> <span class="p">(</span><span class="n">data</span><span class="p">,</span> <span class="n">target</span><span class="p">)</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">test_loader</span><span class="p">):</span>
<span class="n">out</span> <span class="o">=</span> <span class="n">model</span><span class="p">(</span><span class="n">data</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">)</span>
<span class="n">pred</span> <span class="o">=</span> <span class="n">out</span><span class="o">.</span><span class="n">argmax</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">keepdim</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="n">pred</span><span class="p">,</span> <span class="n">target</span> <span class="o">=</span> <span class="n">pred</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">),</span> <span class="n">target</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span>
<span class="n">correct</span> <span class="o">=</span> <span class="n">pred</span><span class="o">.</span><span class="n">eq</span><span class="p">(</span><span class="n">target</span><span class="o">.</span><span class="n">view_as</span><span class="p">(</span><span class="n">pred</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">item</span><span class="p">()</span>
<span class="n">correct_sum</span> <span class="o">+=</span> <span class="n">correct</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Accuracy </span><span class="si">{</span><span class="n">correct_sum</span><span class="w"> </span><span class="o">/</span><span class="w"> </span><span class="nb">len</span><span class="p">(</span><span class="n">test_loader</span><span class="o">.</span><span class="n">dataset</span><span class="p">)</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
</pre></div>
</div>
<p>Next, similar to how we defined <code class="docutils literal notranslate"><span class="pre">run_parameter_server</span></code> as the main loop for our <code class="docutils literal notranslate"><span class="pre">ParameterServer</span></code> that is responsible for initializing RPC, letโs define a similar loop for our trainers. The difference will be that our trainers must run the training loop we defined above:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Main loop for trainers.</span>
<span class="k">def</span> <span class="nf">run_worker</span><span class="p">(</span><span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="p">,</span> <span class="n">num_gpus</span><span class="p">,</span> <span class="n">train_loader</span><span class="p">,</span> <span class="n">test_loader</span><span class="p">):</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Worker rank </span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s2"> initializing RPC"</span><span class="p">)</span>
<span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span>
<span class="n">name</span><span class="o">=</span><span class="sa">f</span><span class="s2">"trainer_</span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s2">"</span><span class="p">,</span>
<span class="n">rank</span><span class="o">=</span><span class="n">rank</span><span class="p">,</span>
<span class="n">world_size</span><span class="o">=</span><span class="n">world_size</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Worker </span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s2"> done initializing RPC"</span><span class="p">)</span>
<span class="n">run_training_loop</span><span class="p">(</span><span class="n">rank</span><span class="p">,</span> <span class="n">num_gpus</span><span class="p">,</span> <span class="n">train_loader</span><span class="p">,</span> <span class="n">test_loader</span><span class="p">)</span>
<span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
</pre></div>
</div>
<p>Note that similar to <code class="docutils literal notranslate"><span class="pre">run_parameter_server</span></code>, <code class="docutils literal notranslate"><span class="pre">rpc.shutdown()</span></code> will by default wait for all workers, both trainers and ParameterServers, to call into <code class="docutils literal notranslate"><span class="pre">rpc.shutdown()</span></code> before this node exits. This ensures that nodes are terminated gracefully and no node goes offline while another is expecting it to be online.</p>
<p>Weโve now completed our trainer and parameter server specific code, and all thatโs left is to add code to launch trainers and parameter servers. First, we must take in various arguments that apply to our parameter server and trainers. <code class="docutils literal notranslate"><span class="pre">world_size</span></code> corresponds to the total number of nodes that will participate in training, and is the sum of all trainers and the parameter server. We also must pass in a unique <code class="docutils literal notranslate"><span class="pre">rank</span></code> for each individual process, from 0 (where we will run our single parameter server) to <code class="docutils literal notranslate"><span class="pre">world_size</span> <span class="pre">-</span> <span class="pre">1</span></code>. <code class="docutils literal notranslate"><span class="pre">master_addr</span></code> and <code class="docutils literal notranslate"><span class="pre">master_port</span></code> are arguments that can be used to identify where the rank 0 process is running, and will be used by individual nodes to discover each other. To test this example out locally, simply pass in <code class="docutils literal notranslate"><span class="pre">localhost</span></code> and the same <code class="docutils literal notranslate"><span class="pre">master_port</span></code> to all instances spawned. Note that for demonstration purposes, this example supports only between 0-2 GPUs, although the pattern can be extended to make use of additional GPUs.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">if</span> <span class="vm">__name__</span> <span class="o">==</span> <span class="s1">'__main__'</span><span class="p">:</span>
<span class="n">parser</span> <span class="o">=</span> <span class="n">argparse</span><span class="o">.</span><span class="n">ArgumentParser</span><span class="p">(</span>
<span class="n">description</span><span class="o">=</span><span class="s2">"Parameter-Server RPC based training"</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span>
<span class="s2">"--world_size"</span><span class="p">,</span>
<span class="nb">type</span><span class="o">=</span><span class="nb">int</span><span class="p">,</span>
<span class="n">default</span><span class="o">=</span><span class="mi">4</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s2">"""Total number of participating processes. Should be the sum of</span>
<span class="s2"> master node and all training nodes."""</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span>
<span class="s2">"--rank"</span><span class="p">,</span>
<span class="nb">type</span><span class="o">=</span><span class="nb">int</span><span class="p">,</span>
<span class="n">default</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s2">"Global rank of this process. Pass in 0 for master."</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span>
<span class="s2">"--num_gpus"</span><span class="p">,</span>
<span class="nb">type</span><span class="o">=</span><span class="nb">int</span><span class="p">,</span>
<span class="n">default</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s2">"""Number of GPUs to use for training, Currently supports between 0</span>
<span class="s2"> and 2 GPUs. Note that this argument will be passed to the parameter servers."""</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span>
<span class="s2">"--master_addr"</span><span class="p">,</span>
<span class="nb">type</span><span class="o">=</span><span class="nb">str</span><span class="p">,</span>
<span class="n">default</span><span class="o">=</span><span class="s2">"localhost"</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s2">"""Address of master, will default to localhost if not provided.</span>
<span class="s2"> Master must be able to accept network traffic on the address + port."""</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span>
<span class="s2">"--master_port"</span><span class="p">,</span>
<span class="nb">type</span><span class="o">=</span><span class="nb">str</span><span class="p">,</span>
<span class="n">default</span><span class="o">=</span><span class="s2">"29500"</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s2">"""Port that master is listening on, will default to 29500 if not</span>
<span class="s2"> provided. Master must be able to accept network traffic on the host and port."""</span><span class="p">)</span>
<span class="n">args</span> <span class="o">=</span> <span class="n">parser</span><span class="o">.</span><span class="n">parse_args</span><span class="p">()</span>
<span class="k">assert</span> <span class="n">args</span><span class="o">.</span><span class="n">rank</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">,</span> <span class="s2">"must provide rank argument."</span>
<span class="k">assert</span> <span class="n">args</span><span class="o">.</span><span class="n">num_gpus</span> <span class="o"><=</span> <span class="mi">3</span><span class="p">,</span> <span class="sa">f</span><span class="s2">"Only 0-2 GPUs currently supported (got </span><span class="si">{</span><span class="n">args</span><span class="o">.</span><span class="n">num_gpus</span><span class="si">}</span><span class="s2">)."</span>
<span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'MASTER_ADDR'</span><span class="p">]</span> <span class="o">=</span> <span class="n">args</span><span class="o">.</span><span class="n">master_addr</span>
<span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s2">"MASTER_PORT"</span><span class="p">]</span> <span class="o">=</span> <span class="n">args</span><span class="o">.</span><span class="n">master_port</span>
</pre></div>
</div>
<p>Now, weโll create a process corresponding to either a parameter server or trainer depending on our command line arguments. Weโll create a <code class="docutils literal notranslate"><span class="pre">ParameterServer</span></code> if our passed in rank is 0, and a <code class="docutils literal notranslate"><span class="pre">TrainerNet</span></code> otherwise. Note that weโre using <code class="docutils literal notranslate"><span class="pre">torch.multiprocessing</span></code> to launch a subprocess corresponding to the function that we want to execute, and waiting on this processโs completion from the main thread with <code class="docutils literal notranslate"><span class="pre">p.join()</span></code>. In the case of initializing our trainers, we also use PyTorchโs <a class="reference external" href="https://pytorch.org/docs/stable/data.html">dataloaders</a> in order to specify train and test data loaders on the MNIST dataset.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">processes</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">world_size</span> <span class="o">=</span> <span class="n">args</span><span class="o">.</span><span class="n">world_size</span>
<span class="k">if</span> <span class="n">args</span><span class="o">.</span><span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="n">p</span> <span class="o">=</span> <span class="n">mp</span><span class="o">.</span><span class="n">Process</span><span class="p">(</span><span class="n">target</span><span class="o">=</span><span class="n">run_parameter_server</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">world_size</span><span class="p">))</span>
<span class="n">p</span><span class="o">.</span><span class="n">start</span><span class="p">()</span>
<span class="n">processes</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">p</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="c1"># Get data to train on</span>
<span class="n">train_loader</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span>
<span class="n">datasets</span><span class="o">.</span><span class="n">MNIST</span><span class="p">(</span><span class="s1">'../data'</span><span class="p">,</span> <span class="n">train</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">download</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="n">transform</span><span class="o">=</span><span class="n">transforms</span><span class="o">.</span><span class="n">Compose</span><span class="p">([</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">ToTensor</span><span class="p">(),</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">Normalize</span><span class="p">((</span><span class="mf">0.1307</span><span class="p">,),</span> <span class="p">(</span><span class="mf">0.3081</span><span class="p">,))</span>
<span class="p">])),</span>
<span class="n">batch_size</span><span class="o">=</span><span class="mi">32</span><span class="p">,</span> <span class="n">shuffle</span><span class="o">=</span><span class="kc">True</span><span class="p">,)</span>
<span class="n">test_loader</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span>
<span class="n">datasets</span><span class="o">.</span><span class="n">MNIST</span><span class="p">(</span>
<span class="s1">'../data'</span><span class="p">,</span>
<span class="n">train</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
<span class="n">transform</span><span class="o">=</span><span class="n">transforms</span><span class="o">.</span><span class="n">Compose</span><span class="p">([</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">ToTensor</span><span class="p">(),</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">Normalize</span><span class="p">((</span><span class="mf">0.1307</span><span class="p">,),</span> <span class="p">(</span><span class="mf">0.3081</span><span class="p">,))</span>
<span class="p">])),</span>
<span class="n">batch_size</span><span class="o">=</span><span class="mi">32</span><span class="p">,</span>
<span class="n">shuffle</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="p">)</span>
<span class="c1"># start training worker on this node</span>
<span class="n">p</span> <span class="o">=</span> <span class="n">mp</span><span class="o">.</span><span class="n">Process</span><span class="p">(</span>
<span class="n">target</span><span class="o">=</span><span class="n">run_worker</span><span class="p">,</span>
<span class="n">args</span><span class="o">=</span><span class="p">(</span>
<span class="n">args</span><span class="o">.</span><span class="n">rank</span><span class="p">,</span>
<span class="n">world_size</span><span class="p">,</span> <span class="n">args</span><span class="o">.</span><span class="n">num_gpus</span><span class="p">,</span>
<span class="n">train_loader</span><span class="p">,</span>
<span class="n">test_loader</span><span class="p">))</span>
<span class="n">p</span><span class="o">.</span><span class="n">start</span><span class="p">()</span>
<span class="n">processes</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">p</span><span class="p">)</span>
<span class="k">for</span> <span class="n">p</span> <span class="ow">in</span> <span class="n">processes</span><span class="p">:</span>
<span class="n">p</span><span class="o">.</span><span class="n">join</span><span class="p">()</span>
</pre></div>
</div>
<p>To run the example locally, run the following command worker for the server and each worker you wish to spawn, in separate terminal windows: <code class="docutils literal notranslate"><span class="pre">python</span> <span class="pre">rpc_parameter_server.py</span> <span class="pre">--world_size=WORLD_SIZE</span> <span class="pre">--rank=RANK</span></code>. For example, for a master node with world size of 2, the command would be <code class="docutils literal notranslate"><span class="pre">python</span> <span class="pre">rpc_parameter_server.py</span> <span class="pre">--world_size=2</span> <span class="pre">--rank=0</span></code>. The trainer can then be launched with the command <code class="docutils literal notranslate"><span class="pre">python</span> <span class="pre">rpc_parameter_server.py</span> <span class="pre">--world_size=2</span> <span class="pre">--rank=1</span></code> in a separate window, and this will begin training with one server and a single trainer. Note that this tutorial assumes that training occurs using between 0 and 2 GPUs, and this argument can be configured by passing <code class="docutils literal notranslate"><span class="pre">--num_gpus=N</span></code> into the training script.</p>
<p>You can pass in the command line arguments <code class="docutils literal notranslate"><span class="pre">--master_addr=ADDRESS</span></code> and <code class="docutils literal notranslate"><span class="pre">--master_port=PORT</span></code> to indicate the address and port that the master worker is listening on, for example, to test functionality where trainers and master nodes run on different machines.</p>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="dist_pipeline_parallel_tutorial.html" class="btn btn-neutral float-right" title="Distributed Pipeline Parallelism Using RPC" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="rpc_tutorial.html" class="btn btn-neutral" title="Getting Started with Distributed RPC Framework" accesskey="p" rel="prev"><img src="../_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr class="community-hr hr-top" />
<div class="community-container">
<div class="community-prompt">๋ ๊ถ๊ธํ์๊ฑฐ๋ ๊ฐ์ ํ ๋ด์ฉ์ด ์์ผ์ ๊ฐ์? ์ปค๋ฎค๋ํฐ์ ์ฐธ์ฌํด๋ณด์ธ์!</div>
<div class="community-link"><a href="https://discuss.pytorch.kr/" aria-label="PyTorchKoreaCommunity">ํ๊ตญ์ด ์ปค๋ฎค๋ํฐ ๋ฐ๋ก๊ฐ๊ธฐ</a></div>
</div>
<hr class="community-hr hr-bottom"/>
<hr class="rating-hr hr-top" />
<div class="rating-container">
<div class="rating-prompt">์ด ํํ ๋ฆฌ์ผ์ด ์ด๋ ์
จ๋์? ํ๊ฐํด์ฃผ์๋ฉด ์ดํ ๊ฐ์ ์ ์ฐธ๊ณ ํ๊ฒ ์ต๋๋ค! :)</div>
<div class="stars-outer">
<i class="far fa-star" title="1 Star" data-behavior="tutorial-rating" data-count="1"></i>
<i class="far fa-star" title="2 Stars" data-behavior="tutorial-rating" data-count="2"></i>
<i class="far fa-star" title="3 Stars" data-behavior="tutorial-rating" data-count="3"></i>
<i class="far fa-star" title="4 Stars" data-behavior="tutorial-rating" data-count="4"></i>
<i class="far fa-star" title="5 Stars" data-behavior="tutorial-rating" data-count="5"></i>
</div>
</div>
<hr class="rating-hr hr-bottom"/>
<div role="contentinfo">
<p>
© Copyright 2018-2024, PyTorch & ํ์ดํ ์น ํ๊ตญ ์ฌ์ฉ์ ๋ชจ์(PyTorch Korea User Group).
</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>
<script>
if((window.location.href.indexOf("/prototype/")!= -1) && (window.location.href.indexOf("/prototype/prototype_index")< 1))
{
var div = '<div class="admonition note"><p class="admonition-title">Note</p><p><i class="fa fa-flask" aria-hidden="true"> </i> ์ด ํํ ๋ฆฌ์ผ์ ํ๋กํ ํ์
(prototype) ๊ธฐ๋ฅ๋ค์ ๋ํด์ ์ค๋ช
ํ๊ณ ์์ต๋๋ค. ํ๋กํ ํ์
๊ธฐ๋ฅ์ ์ผ๋ฐ์ ์ผ๋ก ํผ๋๋ฐฑ ๋ฐ ํ
์คํธ์ฉ์ผ๋ก, ๋ฐํ์ ํ๋๊ทธ ์์ด๋ PyPI๋ Conda๋ก ๋ฐฐํฌ๋๋ ๋ฐ์ด๋๋ฆฌ์์๋ ์ฌ์ฉํ ์ ์์ต๋๋ค.</p></div>'
document.getElementById("pytorch-article").insertAdjacentHTML('afterBegin', div)
}
</script>
</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="#">Implementing a Parameter Server Using Distributed RPC Framework</a></li>
</ul>
</div>
</div>
</div>
</section>
</div>
<script type="text/javascript" id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script data-url_root="../" id="documentation_options" src="../_static/documentation_options.js"></script>
<script src="../_static/jquery.js"></script>
<script src="../_static/underscore.js"></script>
<script src="../_static/_sphinx_javascript_frameworks_compat.js"></script>
<script src="../_static/doctools.js"></script>
<script src="../_static/clipboard.min.js"></script>
<script src="../_static/copybutton.js"></script>
<script src="../_static/translations.js"></script>
<script src="../_static/katex.min.js"></script>
<script src="../_static/auto-render.min.js"></script>
<script src="../_static/katex_autorenderer.js"></script>
<script src="../_static/design-tabs.js"></script>
<script type="text/javascript" src="../_static/js/vendor/popper.min.js"></script>
<script type="text/javascript" src="../_static/js/vendor/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/list.js/1.5.0/list.min.js"></script>
<script type="text/javascript" src="../_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<script>
// Helper function to make it easier to call dataLayer.push()
function gtag(){window.dataLayer.push(arguments);}
//add microsoft link
if(window.location.href.indexOf("/beginner/basics/")!= -1)
{
var url="https://docs.microsoft.com/learn/paths/pytorch-fundamentals/?wt.mc_id=aiml-7486-cxa";
switch(window.location.pathname.split("/").pop().replace('.html',''))
{
case"quickstart_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/9-quickstart?WT.mc_id=aiml-7486-cxa";
break;
case"tensorqs_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/2-tensors?WT.mc_id=aiml-7486-cxa";
break;
case"data_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/3-data?WT.mc_id=aiml-7486-cxa";
break;
case"transforms_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/4-transforms?WT.mc_id=aiml-7486-cxa";
break;
case"buildmodel_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/5-model?WT.mc_id=aiml-7486-cxa";
break;
case"autogradqs_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/6-autograd?WT.mc_id=aiml-7486-cxa";
break;
case"optimization_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/7-optimization?WT.mc_id=aiml-7486-cxa";
break;
case"saveloadrun_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/8-inference?WT.mc_id=aiml-7486-cxa";
}
$(".pytorch-call-to-action-links").children().first().before("<a href="+url+' data-behavior="call-to-action-event" data-response="Run in Microsoft Learn" target="_blank"><div id="microsoft-learn-link" style="padding-bottom: 0.625rem;border-bottom: 1px solid #f3f4f7;padding-right: 2.5rem;display: -webkit-box; display: -ms-flexbox; display: flex; -webkit-box-align: center;-ms-flex-align: center;align-items: center;"><img class="call-to-action-img" src="../../_static/images/microsoft-logo.svg"/><div class="call-to-action-desktop-view">Run in Microsoft Learn</div><div class="call-to-action-mobile-view">Learn</div></div></a>')
}
</script>
<script async src="https://www.googletagmanager.com/gtag/js?id=G-LZRD6GXDLF"></script>
<script data-cfasync="false">
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'G-LZRD6GXDLF'); // GA4
gtag('config', 'UA-71919972-3'); // UA
</script>
<script data-cfasync="false">
$("[data-behavior='call-to-action-event']").on('click', function(){
ga('send', {
hitType: 'event',
eventCategory: $(this).attr("data-response"),
eventAction: 'click',
eventLabel: window.location.href
});
gtag('event', 'click', {
'event_category': $(this).attr("data-response"),
'event_label': $("h1").first().text(),
'tutorial_link': window.location.href
});
});
$("[data-behavior='tutorial-rating']").on('click', function(){
gtag('event', 'click', {
'event_category': 'Tutorial Rating',
'event_label': $("h1").first().text(),
'value': $(this).attr("data-count"),
'customEvent:Rating': $(this).attr("data-count") // send to GA custom dimension customEvent:Rating.
});
});
if (location.pathname == "/") {
$(".rating-container").hide();
$(".hr-bottom").hide();
}
</script>
<script type="text/javascript">
var collapsedSections = ['ํ์ดํ ์น(PyTorch) ๋ ์ํผ', 'ํ์ดํ ์น(PyTorch) ๋ฐฐ์ฐ๊ธฐ', 'Introduction to PyTorch on YouTube', '์ด๋ฏธ์ง/๋น๋์ค', '์ค๋์ค', 'ํ
์คํธ', '๋ฐฑ์๋', '๊ฐํํ์ต', 'PyTorch ๋ชจ๋ธ์ ํ๋ก๋์
ํ๊ฒฝ์ ๋ฐฐํฌํ๊ธฐ', 'PyTorch ํ๋กํ์ผ๋ง', 'Code Transforms with FX', 'ํ๋ก ํธ์๋ API', 'PyTorch ํ์ฅํ๊ธฐ', '๋ชจ๋ธ ์ต์ ํ', '๋ณ๋ ฌ ๋ฐ ๋ถ์ฐ ํ์ต', 'Edge with ExecuTorch', '์ถ์ฒ ์์คํ
', 'Multimodality'];
</script>
<!-- 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">