-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathdistributed.tensor.parallel.html
1330 lines (1148 loc) · 101 KB
/
distributed.tensor.parallel.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Tensor Parallelism - torch.distributed.tensor.parallel — PyTorch 2.7 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/distributed.tensor.parallel.html"/>
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="_static/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/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="_static/sphinx-dropdown.css" type="text/css" />
<link rel="stylesheet" href="_static/panels-bootstrap.min.css" type="text/css" />
<link rel="stylesheet" href="_static/css/jit.css" type="text/css" />
<link rel="stylesheet" href="_static/css/custom.css" type="text/css" />
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="Distributed Optimizers" href="distributed.optim.html" />
<link rel="prev" title="torch.distributed.fsdp.fully_shard" href="distributed.fsdp.fully_shard.html" />
<!-- Google Tag Manager -->
<script>(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':
new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],
j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src=
'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);
})(window,document,'script','dataLayer','GTM-T8XT4PS');</script>
<!-- End Google Tag Manager -->
<script src="_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Learn
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/get-started">
<span class=dropdown-title>Get Started</span>
<p>Run PyTorch locally or get started quickly with one of the supported cloud platforms</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials">
<span class="dropdown-title">Tutorials</span>
<p>Whats new in PyTorch tutorials</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/basics/intro.html">
<span class="dropdown-title">Learn the Basics</span>
<p>Familiarize yourself with PyTorch concepts and modules</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/recipes/recipes_index.html">
<span class="dropdown-title">PyTorch Recipes</span>
<p>Bite-size, ready-to-deploy PyTorch code examples</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/introyt.html">
<span class="dropdown-title">Intro to PyTorch - YouTube Series</span>
<p>Master PyTorch basics with our engaging YouTube tutorial series</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Ecosystem
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem">
<span class="dropdown-title">Tools</span>
<p>Learn about the tools and frameworks in the PyTorch Ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class=dropdown-title>Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class=dropdown-title>Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class=dropdown-title>Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem/contributor-awards-2024">
<span class="dropdown-title">Contributor Awards - 2024</span>
<p>Award winners announced at this year's PyTorch Conference</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Edge
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/edge">
<span class="dropdown-title">About PyTorch Edge</span>
<p>Build innovative and privacy-aware AI experiences for edge devices</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch-overview">
<span class="dropdown-title">ExecuTorch</span>
<p>End-to-end solution for enabling on-device inference capabilities across mobile and edge devices</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch/stable/index.html">
<span class="dropdown-title">ExecuTorch Docs</span>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p>Explore the documentation for comprehensive guidance on how to use PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/pytorch-domains">
<span class="dropdown-title">PyTorch Domains</span>
<p>Read the PyTorch Domains documentation to learn more about domain-specific libraries</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Blogs & News
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/blog/">
<span class="dropdown-title">PyTorch Blog</span>
<p>Catch up on the latest technical news and happenings</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/community-blog">
<span class="dropdown-title">Community Blog</span>
<p>Stories from the PyTorch ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/videos">
<span class="dropdown-title">Videos</span>
<p>Learn about the latest PyTorch tutorials, new, and more </p>
<a class="nav-dropdown-item" href="https://pytorch.org/community-stories">
<span class="dropdown-title">Community Stories</span>
<p>Learn how our community solves real, everyday machine learning problems with PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/newsletter">
<span class="dropdown-title">Newsletter</span>
<p>Stay up-to-date with the latest updates</p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
About
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn more about the PyTorch Foundation</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/governing-board">
<span class="dropdown-title">Governing Board</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/credits">
<span class="dropdown-title">Cloud Credit Program</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tac">
<span class="dropdown-title">Technical Advisory Council</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/staff">
<span class="dropdown-title">Staff</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/contact-us">
<span class="dropdown-title">Contact Us</span>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div class="no-dropdown">
<a href="https://pytorch.org/join" data-cta="join">
Become a Member
</a>
</div>
</li>
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="github-icon">
</a>
</div>
</li>
<!--- TODO: This block adds the search icon to the nav bar. We will enable it later.
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="search-icon">
</a>
</div>
</li>
--->
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>2.7 ▼</a>
</div>
<div id="searchBox">
<div class="searchbox" id="googleSearchBox">
<script async src="https://cse.google.com/cse.js?cx=e65585f8c3ea1440e"></script>
<div class="gcse-search"></div>
</div>
<div id="sphinxSearchBox" style="display: none;">
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
</div>
<form id="searchForm">
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="google" checked>
Google Search
</label>
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="sphinx">
Classic Search
</label>
</form>
<script>
document.addEventListener('DOMContentLoaded', function() {
const searchForm = document.getElementById('searchForm');
const googleSearchBox = document.getElementById('googleSearchBox');
const sphinxSearchBox = document.getElementById('sphinxSearchBox');
// Function to toggle search box visibility
function toggleSearchBox(searchType) {
googleSearchBox.style.display = searchType === 'google' ? 'block' : 'none';
sphinxSearchBox.style.display = searchType === 'sphinx' ? 'block' : 'none';
}
// Determine the default search type
let defaultSearchType;
const currentUrl = window.location.href;
if (currentUrl.startsWith('https://pytorch.org/docs/stable')) {
// For the stable documentation, default to Google
defaultSearchType = localStorage.getItem('searchType') || 'google';
} else {
// For any other version, including docs-preview, default to Sphinx
defaultSearchType = 'sphinx';
}
// Set the default search type
document.querySelector(`input[name="searchType"][value="${defaultSearchType}"]`).checked = true;
toggleSearchBox(defaultSearchType);
// Event listener for changes in search type
searchForm.addEventListener('change', function(event) {
const selectedSearchType = event.target.value;
localStorage.setItem('searchType', selectedSearchType);
toggleSearchBox(selectedSearchType);
});
// Set placeholder text for Google search box
window.onload = function() {
var placeholderText = "Search Docs";
var googleSearchboxText = document.querySelector("#gsc-i-id1");
if (googleSearchboxText) {
googleSearchboxText.placeholder = placeholderText;
googleSearchboxText.style.fontFamily = 'FreightSans';
googleSearchboxText.style.fontSize = "1.2rem";
googleSearchboxText.style.color = '#262626';
}
};
});
</script>
</div>
<p class="caption" role="heading"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="community/build_ci_governance.html">PyTorch Governance | Build + CI</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/design.html">PyTorch Design Philosophy</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/governance.html">PyTorch Governance | Mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/persons_of_interest.html">PyTorch Governance | Maintainers</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Developer Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="notes/amp_examples.html">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/custom_operators.html">PyTorch Custom Operators Landing Page</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.func.html">Extending torch.func with autograd.Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/fsdp.html">FSDP Notes</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/get_start_xpu.html">Getting Started on Intel GPU</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/libtorch_stable_abi.html">LibTorch Stable ABI</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/mps.html">MPS backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/numerical_accuracy.html">Numerical accuracy</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
<li class="toctree-l1"><a class="reference internal" href="deploy.html">torch::deploy</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python API</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="amp.html">torch.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="library.html">torch.library</a></li>
<li class="toctree-l1"><a class="reference internal" href="accelerator.html">torch.accelerator</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpu.html">torch.cpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html">Understanding CUDA Memory Usage</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#generating-a-snapshot">Generating a Snapshot</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#using-the-visualizer">Using the visualizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#snapshot-api-reference">Snapshot API Reference</a></li>
<li class="toctree-l1"><a class="reference internal" href="mps.html">torch.mps</a></li>
<li class="toctree-l1"><a class="reference internal" href="xpu.html">torch.xpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="mtia.html">torch.mtia</a></li>
<li class="toctree-l1"><a class="reference internal" href="mtia.memory.html">torch.mtia.memory</a></li>
<li class="toctree-l1"><a class="reference internal" href="meta.html">Meta device</a></li>
<li class="toctree-l1"><a class="reference internal" href="backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="export.html">torch.export</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.tensor.html">torch.distributed.tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="fsdp.html">torch.distributed.fsdp</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.fsdp.fully_shard.html">torch.distributed.fsdp.fully_shard</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">torch.distributed.tensor.parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.pipelining.html">torch.distributed.pipelining</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.checkpoint.html">torch.distributed.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.compiler.html">torch.compiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="func.html">torch.func</a></li>
<li class="toctree-l1"><a class="reference internal" href="futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.experimental.html">torch.fx.experimental</a></li>
<li class="toctree-l1"><a class="reference internal" href="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="monitor.html">torch.monitor</a></li>
<li class="toctree-l1"><a class="reference internal" href="signal.html">torch.signal</a></li>
<li class="toctree-l1"><a class="reference internal" href="special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.attention.html">torch.nn.attention</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="masked.html">torch.masked</a></li>
<li class="toctree-l1"><a class="reference internal" href="nested.html">torch.nested</a></li>
<li class="toctree-l1"><a class="reference internal" href="size.html">torch.Size</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="utils.html">torch.utils</a></li>
<li class="toctree-l1"><a class="reference internal" href="benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="deterministic.html">torch.utils.deterministic</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit_utils.html">torch.utils.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="module_tracker.html">torch.utils.module_tracker</a></li>
<li class="toctree-l1"><a class="reference internal" href="type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="config_mod.html">torch.__config__</a></li>
<li class="toctree-l1"><a class="reference internal" href="future_mod.html">torch.__future__</a></li>
<li class="toctree-l1"><a class="reference internal" href="logging.html">torch._logging</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_environment_variables.html">Torch Environment Variables</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/data">TorchData</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/torchrec">TorchRec</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/ao">torchao</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="index.html">
Docs
</a> >
</li>
<li>Tensor Parallelism - torch.distributed.tensor.parallel</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/distributed.tensor.parallel.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">
<!-- Google Tag Manager (noscript) -->
<noscript><iframe src="https://www.googletagmanager.com/ns.html?id=GTM-T8XT4PS"
height="0" width="0" style="display:none;visibility:hidden"></iframe></noscript>
<!-- End Google Tag Manager (noscript) -->
<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="tensor-parallelism-torch-distributed-tensor-parallel">
<h1>Tensor Parallelism - torch.distributed.tensor.parallel<a class="headerlink" href="#tensor-parallelism-torch-distributed-tensor-parallel" title="Permalink to this heading">¶</a></h1>
<p>Tensor Parallelism(TP) is built on top of the PyTorch DistributedTensor
(<a class="reference external" href="https://github.com/pytorch/pytorch/blob/main/torch/distributed/tensor/README.md">DTensor</a>)
and provides different parallelism styles: Colwise, Rowwise, and Sequence Parallelism.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Tensor Parallelism APIs are experimental and subject to change.</p>
</div>
<p>The entrypoint to parallelize your <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code> using Tensor Parallelism is:</p>
<span class="target" id="module-torch.distributed.tensor.parallel"></span><dl class="py function">
<dt class="sig sig-object py" id="torch.distributed.tensor.parallel.parallelize_module">
<span class="sig-prename descclassname"><span class="pre">torch.distributed.tensor.parallel.</span></span><span class="sig-name descname"><span class="pre">parallelize_module</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">module</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">device_mesh</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">parallelize_plan</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">src_data_rank</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/tensor/parallel/api.html#parallelize_module"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/tensor/parallel/api.py#L16"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.tensor.parallel.parallelize_module" title="Permalink to this definition">¶</a></dt>
<dd><p>Apply Tensor Parallelism in PyTorch by parallelizing modules or sub-modules based on a user-specified plan.</p>
<p>We parallelize module or sub_modules based on a parallelize_plan. The parallelize_plan contains
<code class="xref py py-class docutils literal notranslate"><span class="pre">ParallelStyle</span></code>, which indicates how user wants the module or sub_module
to be parallelized.</p>
<p>User can also specify different parallel style per module fully qualified name (FQN).</p>
<p>Note that <code class="docutils literal notranslate"><span class="pre">parallelize_module</span></code> only accepts a 1-D <code class="xref py py-class docutils literal notranslate"><span class="pre">DeviceMesh</span></code>, if you have a 2-D or N-D <code class="xref py py-class docutils literal notranslate"><span class="pre">DeviceMesh</span></code>,
slice the DeviceMesh to a 1-D sub DeviceMesh first then pass to this API(i.e. <code class="docutils literal notranslate"><span class="pre">device_mesh["tp"]</span></code>)</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>module</strong> (<code class="xref py py-class docutils literal notranslate"><span class="pre">nn.Module</span></code>) – Module to be parallelized.</p></li>
<li><p><strong>device_mesh</strong> (<code class="xref py py-class docutils literal notranslate"><span class="pre">DeviceMesh</span></code>, optional) – Object which describes the mesh topology of devices for the DTensor.
If not specified, the call must be under a DeviceMesh context.</p></li>
<li><p><strong>parallelize_plan</strong> (Union[<code class="xref py py-class docutils literal notranslate"><span class="pre">ParallelStyle</span></code>, Dict[str, <code class="xref py py-class docutils literal notranslate"><span class="pre">ParallelStyle</span></code>]], optional) – The plan used to parallelize the module. It can be either a
<code class="xref py py-class docutils literal notranslate"><span class="pre">ParallelStyle</span></code> object which contains how we prepare
input/output for Tensor Parallelism or it can be a dict of module
FQN and its corresponding <code class="xref py py-class docutils literal notranslate"><span class="pre">ParallelStyle</span></code> object. If not
specified, the call will do nothing at the moment.</p></li>
</ul>
</dd>
<dt class="field-even">Keyword Arguments</dt>
<dd class="field-even"><p><strong>src_data_rank</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.13)"><em>int</em></a><em>, </em><em>optional</em>) – the rank of the source data for the logical/global tensor, it is used by
<code class="xref py py-meth docutils literal notranslate"><span class="pre">distribute_tensor()</span></code> to scatter/broadcast the shards/replicas to other ranks. By default,
we use <code class="docutils literal notranslate"><span class="pre">group_rank=0</span></code> on each DeviceMesh dimension as the source data to preserve the single-device
semantic. If passing <code class="docutils literal notranslate"><span class="pre">None</span></code> explicitly, <a class="reference internal" href="#torch.distributed.tensor.parallel.parallelize_module" title="torch.distributed.tensor.parallel.parallelize_module"><code class="xref py py-meth docutils literal notranslate"><span class="pre">parallelize_module()</span></code></a> simply uses its local data instead
of trying to preserve the single-device semantic via scatter/broadcast. Default: 0</p>
</dd>
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>A <code class="xref py py-class docutils literal notranslate"><span class="pre">nn.Module</span></code> object parallelized.</p>
</dd>
<dt class="field-even">Return type</dt>
<dd class="field-even"><p><p id="torch.nn.Module"/><a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.modules.module.Module"><em>Module</em></a></p>
</dd>
</dl>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.tensor.parallel</span> <span class="kn">import</span> <span class="n">parallelize_module</span><span class="p">,</span> <span class="n">ColwiseParallel</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.device_mesh</span> <span class="kn">import</span> <span class="n">init_device_mesh</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># Define the module.</span>
<span class="gp">>>> </span><span class="n">m</span> <span class="o">=</span> <span class="n">Model</span><span class="p">(</span><span class="o">...</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">tp_mesh</span> <span class="o">=</span> <span class="n">init_device_mesh</span><span class="p">(</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="p">(</span><span class="mi">8</span><span class="p">,))</span>
<span class="gp">>>> </span><span class="n">m</span> <span class="o">=</span> <span class="n">parallelize_module</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">tp_mesh</span><span class="p">,</span> <span class="p">{</span><span class="s2">"w1"</span><span class="p">:</span> <span class="n">ColwiseParallel</span><span class="p">(),</span> <span class="s2">"w2"</span><span class="p">:</span> <span class="n">RowwiseParallel</span><span class="p">()})</span>
<span class="go">>>></span>
</pre></div>
</div>
</dd>
</dl>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>For complex module architecture like Attention, MLP layers, we recommend composing
different ParallelStyles together (i.e. <code class="docutils literal notranslate"><span class="pre">ColwiseParallel</span></code> and <code class="docutils literal notranslate"><span class="pre">RowwiseParallel</span></code>) and pass
as a parallelize_plan, to achieves the desired sharding computation.</p>
</div>
</dd></dl>
<p>Tensor Parallelism supports the following parallel styles:</p>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.tensor.parallel.ColwiseParallel">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.tensor.parallel.</span></span><span class="sig-name descname"><span class="pre">ColwiseParallel</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">input_layouts</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">output_layouts</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">use_local_output</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/tensor/parallel/style.html#ColwiseParallel"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/tensor/parallel/style.py#L44"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.tensor.parallel.ColwiseParallel" title="Permalink to this definition">¶</a></dt>
<dd><p>Partition a compatible nn.Module in a column-wise fashion. Currently supports nn.Linear and nn.Embedding.
Users can compose it together with RowwiseParallel to achieve the sharding of more complicated modules.
(i.e. MLP, Attention)</p>
<dl class="field-list simple">
<dt class="field-odd">Keyword Arguments</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>input_layouts</strong> (<a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>, </em><em>optional</em>) – The DTensor layout of input tensor for the nn.Module, this is used to annotate the input tensor to
become a DTensor. If not specified, we assume the input tensor to be replicated.</p></li>
<li><p><strong>output_layouts</strong> (<a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>, </em><em>optional</em>) – The DTensor layout of the output for the nn.Module, this is used to ensure the output of the nn.Module
with the user desired layout. If not specified, the output tensor is sharded on the last dimension.</p></li>
<li><p><strong>use_local_output</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><em>bool</em></a><em>, </em><em>optional</em>) – Whether to use local <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a> instead of <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code> for the module output, default: True.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>A <code class="xref py py-class docutils literal notranslate"><span class="pre">ParallelStyle</span></code> object that represents Colwise sharding of the nn.Module.</p>
</dd>
</dl>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.tensor.parallel</span> <span class="kn">import</span> <span class="n">parallelize_module</span><span class="p">,</span> <span class="n">ColwiseParallel</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.device_mesh</span> <span class="kn">import</span> <span class="n">init_device_mesh</span>
<span class="gp">>>> </span><span class="o">...</span>
<span class="gp">>>> </span><span class="n">m</span> <span class="o">=</span> <span class="n">Model</span><span class="p">(</span><span class="o">...</span><span class="p">)</span> <span class="c1"># m is a nn.Module that contains a "w1" nn.Linear submodule</span>
<span class="gp">>>> </span><span class="n">tp_mesh</span> <span class="o">=</span> <span class="n">init_device_mesh</span><span class="p">(</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="p">(</span><span class="mi">8</span><span class="p">,))</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># By default, the input of the "w1" Linear will be converted to Replicated DTensor</span>
<span class="gp">>>> </span><span class="c1"># and the output of "w1" will return :class:`torch.Tensor` that shards on the last dim.</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="n">sharded_mod</span> <span class="o">=</span> <span class="n">parallelize_module</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">tp_mesh</span><span class="p">,</span> <span class="p">{</span><span class="s2">"w1"</span><span class="p">:</span> <span class="n">ColwiseParallel</span><span class="p">()})</span>
<span class="gp">>>> </span><span class="o">...</span>
</pre></div>
</div>
</dd>
</dl>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>By default <code class="docutils literal notranslate"><span class="pre">ColwiseParallel</span></code> output is sharded on the last dimension if the <code class="docutils literal notranslate"><span class="pre">output_layouts</span></code> not
specified, if there’re operators that require specific tensor shape (i.e. before the paired <code class="docutils literal notranslate"><span class="pre">RowwiseParallel</span></code>),
keep in mind that if the output is sharded the operator might need to be adjusted to the sharded size.</p>
</div>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.tensor.parallel.RowwiseParallel">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.tensor.parallel.</span></span><span class="sig-name descname"><span class="pre">RowwiseParallel</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">input_layouts</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">output_layouts</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">use_local_output</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/tensor/parallel/style.html#RowwiseParallel"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/tensor/parallel/style.py#L170"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.tensor.parallel.RowwiseParallel" title="Permalink to this definition">¶</a></dt>
<dd><p>Partition a compatible nn.Module in a row-wise fashion. Currently supports nn.Linear and nn.Embedding.
Users can compose it with ColwiseParallel to achieve the sharding of more complicated modules.
(i.e. MLP, Attention)</p>
<dl class="field-list simple">
<dt class="field-odd">Keyword Arguments</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>input_layouts</strong> (<a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>, </em><em>optional</em>) – The DTensor layout of input tensor for the nn.Module, this is used to annotate the input tensor to
become a DTensor. If not specified, we assume the input tensor to be sharded on the last dimension.</p></li>
<li><p><strong>output_layouts</strong> (<a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>, </em><em>optional</em>) – The DTensor layout of the output for the nn.Module, this is used to ensure the output of the nn.Module
with the user desired layout. If not specified, the output tensor is replicated.</p></li>
<li><p><strong>use_local_output</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><em>bool</em></a><em>, </em><em>optional</em>) – Whether to use local <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a> instead of <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code> for the module output, default: True.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>A <code class="xref py py-class docutils literal notranslate"><span class="pre">ParallelStyle</span></code> object that represents Rowwise sharding of the nn.Module.</p>
</dd>
</dl>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.tensor.parallel</span> <span class="kn">import</span> <span class="n">parallelize_module</span><span class="p">,</span> <span class="n">RowwiseParallel</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.device_mesh</span> <span class="kn">import</span> <span class="n">init_device_mesh</span>
<span class="gp">>>> </span><span class="o">...</span>
<span class="gp">>>> </span><span class="n">m</span> <span class="o">=</span> <span class="n">Model</span><span class="p">(</span><span class="o">...</span><span class="p">)</span> <span class="c1"># m is a nn.Module that contains a "w2" nn.Linear submodule</span>
<span class="gp">>>> </span><span class="n">tp_mesh</span> <span class="o">=</span> <span class="n">init_device_mesh</span><span class="p">(</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="p">(</span><span class="mi">8</span><span class="p">,))</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># By default, the input of the "w2" Linear will be converted to DTensor that shards on the last dim</span>
<span class="gp">>>> </span><span class="c1"># and the output of "w2" will return a replicated :class:`torch.Tensor`.</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="n">sharded_mod</span> <span class="o">=</span> <span class="n">parallelize_module</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">tp_mesh</span><span class="p">,</span> <span class="p">{</span><span class="s2">"w2"</span><span class="p">:</span> <span class="n">RowwiseParallel</span><span class="p">()}),</span>
<span class="gp">>>> </span><span class="o">...</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.tensor.parallel.SequenceParallel">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.tensor.parallel.</span></span><span class="sig-name descname"><span class="pre">SequenceParallel</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">sequence_dim</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">1</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">use_local_output</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/tensor/parallel/style.html#SequenceParallel"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/tensor/parallel/style.py#L307"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.tensor.parallel.SequenceParallel" title="Permalink to this definition">¶</a></dt>
<dd><p>SequenceParallel replicates a compatible <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code> parameters and runs the sharded computation with
input sharded on the sequence dimension. This currently supports <code class="docutils literal notranslate"><span class="pre">nn.LayerNorm</span></code>, <code class="docutils literal notranslate"><span class="pre">nn.Dropout</span></code>, and the
<a class="reference external" href="https://github.com/facebookresearch/llama/blob/main/llama/model.py#L34">RMSNorm python implementation</a></p>
<p>This style implements the operation that is described in the paper
<a class="reference external" href="https://arxiv.org/abs/2205.05198">Reducing Activation Recomputation in Large Transformer Models</a></p>
<p>If the input passed in to this <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code> is a <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a>, it assumes that the input is already sharded
on the sequence dimension and converts the input to a <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code> sharded on the sequence dimension. If the input
passed in to this <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code> is already a <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code> but is not sharded on the sequence dimension, it would
redistribute the input to be sharded on the sequence dimension.</p>
<p>The output of the <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code> will be sharded on the sequence dimension.</p>
<dl class="field-list simple">
<dt class="field-odd">Keyword Arguments</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>sequence_dim</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.13)"><em>int</em></a><em>, </em><em>optional</em>) – The sequence dimension of the input tensor for the <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code>, this is used to annotate the input tensor to
become a DTensor that is sharded on the sequence dimension, default: 1.</p></li>
<li><p><strong>use_local_output</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><em>bool</em></a><em>, </em><em>optional</em>) – Whether to use local <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a> instead of <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code> for the module output, default: False.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>A <code class="xref py py-class docutils literal notranslate"><span class="pre">ParallelStyle</span></code> object that represents Sequence Parallel of the <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code>.</p>
</dd>
</dl>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.tensor.parallel</span> <span class="kn">import</span> <span class="n">parallelize_module</span><span class="p">,</span> <span class="n">SequenceParallel</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.device_mesh</span> <span class="kn">import</span> <span class="n">init_device_mesh</span>
<span class="gp">>>> </span><span class="o">...</span>
<span class="gp">>>> </span><span class="n">m</span> <span class="o">=</span> <span class="n">Model</span><span class="p">(</span><span class="o">...</span><span class="p">)</span> <span class="c1"># m is a nn.Module that contains a "norm" nn.LayerNorm submodule</span>
<span class="gp">>>> </span><span class="n">tp_mesh</span> <span class="o">=</span> <span class="n">init_device_mesh</span><span class="p">(</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="p">(</span><span class="mi">8</span><span class="p">,))</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># By default, the input of the "norm" will be converted to DTensor that shards on the sequence dim</span>
<span class="gp">>>> </span><span class="c1"># and the output of "norm" will return a sharded on sequence dimension :class:`DTensor`.</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="n">sharded_mod</span> <span class="o">=</span> <span class="n">parallelize_module</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">tp_mesh</span><span class="p">,</span> <span class="p">{</span><span class="s2">"norm"</span><span class="p">:</span> <span class="n">SequenceParallel</span><span class="p">()}),</span>
<span class="gp">>>> </span><span class="o">...</span>
</pre></div>
</div>
</dd>
</dl>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>SequenceParallel style assumes ones initialization if there are weights in the nn.Module (i.e.
<code class="docutils literal notranslate"><span class="pre">nn.LayerNorm</span></code> or <code class="docutils literal notranslate"><span class="pre">RMSNorm</span></code>, and they by default have ones initialization). If you have custom
inits for the weights on those modules, you need to broadcast the weights before/after parallelizing
to ensure that they are replicated.</p>
</div>
</dd></dl>
<p>To simply configure the nn.Module’s inputs and outputs with DTensor layouts
and perform necessary layout redistributions, without distribute the module
parameters to DTensors, the following <code class="docutils literal notranslate"><span class="pre">ParallelStyle</span></code> s can be used in
the <code class="docutils literal notranslate"><span class="pre">parallelize_plan</span></code> when calling <code class="docutils literal notranslate"><span class="pre">parallelize_module</span></code>:</p>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.tensor.parallel.PrepareModuleInput">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.tensor.parallel.</span></span><span class="sig-name descname"><span class="pre">PrepareModuleInput</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">input_layouts</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">desired_input_layouts</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">input_kwarg_layouts</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">desired_input_kwarg_layouts</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">use_local_output</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/tensor/parallel/style.html#PrepareModuleInput"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/tensor/parallel/style.py#L402"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.tensor.parallel.PrepareModuleInput" title="Permalink to this definition">¶</a></dt>
<dd><p>Configure the nn.Module’s inputs to convert the input tensors of the nn.Module to DTensors at runtime according to
<code class="docutils literal notranslate"><span class="pre">input_layouts</span></code>, and perform layout redistribution according to the <code class="docutils literal notranslate"><span class="pre">desired_input_layouts</span></code>.</p>
<dl class="field-list simple">
<dt class="field-odd">Keyword Arguments</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>input_layouts</strong> (<em>Union</em><em>[</em><a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>, </em><em>Tuple</em><em>[</em><em>Optional</em><em>[</em><a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>]</em><em>]</em><em>]</em>) – The DTensor layouts of input tensors for the nn.Module, this is used to convert the input tensors to
DTensors. If some inputs are not torch.Tensor or no need to convert to DTensors, <code class="docutils literal notranslate"><span class="pre">None</span></code> need to be specified
as a placeholder. default: None.</p></li>
<li><p><strong>desired_input_layouts</strong> (<em>Union</em><em>[</em><a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>, </em><em>Tuple</em><em>[</em><em>Optional</em><em>[</em><a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>]</em><em>]</em><em>]</em>) – The desired DTensor layout of input tensors for the nn.Module, this is used to ensure the inputs of the nn.Module
have the desired DTensor layouts. This argument needs to have the same length with <code class="docutils literal notranslate"><span class="pre">input_layouts</span></code>. default: None.</p></li>
<li><p><strong>input_kwarg_layouts</strong> (<em>Dict</em><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.13)"><em>str</em></a><em>, </em><a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>]</em>) – The DTensor layouts of input kwargs for the nn.Module, this is used to convert the input kwarg tensors to DTensors.
default: None</p></li>
<li><p><strong>desired_input_kwarg_layouts</strong> – (Dict[str, Placement]):
The desired DTensor layout of input kwargs for the nn.Module, this is used to ensure the inputs of the nn.Module
have the desired DTensor layouts. default: None.</p></li>
<li><p><strong>use_local_output</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><em>bool</em></a><em>, </em><em>optional</em>) – Whether to use local <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a> instead of <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code> for the module inputs, default: False.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>A <code class="xref py py-class docutils literal notranslate"><span class="pre">ParallelStyle</span></code> object that prepares the sharding layouts of the nn.Module’s inputs.</p>
</dd>
</dl>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.tensor.parallel</span> <span class="kn">import</span> <span class="n">parallelize_module</span><span class="p">,</span> <span class="n">PrepareModuleInput</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.device_mesh</span> <span class="kn">import</span> <span class="n">init_device_mesh</span>
<span class="gp">>>> </span><span class="o">...</span>
<span class="gp">>>> </span><span class="n">block</span> <span class="o">=</span> <span class="n">TransformerBlock</span><span class="p">(</span><span class="o">...</span><span class="p">)</span> <span class="c1"># block is a nn.Module that contains an "attn" Attention submodule</span>
<span class="gp">>>> </span><span class="n">tp_mesh</span> <span class="o">=</span> <span class="n">init_device_mesh</span><span class="p">(</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="p">(</span><span class="mi">8</span><span class="p">,))</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># According to the style specified below, the first input of attn will be annotated to Sharded DTensor</span>
<span class="gp">>>> </span><span class="c1"># and then redistributed to Replicated DTensor.</span>
<span class="gp">>>> </span><span class="n">parallelize_module</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">block</span><span class="p">,</span> <span class="c1"># this can be a submodule or module</span>
<span class="gp">>>> </span> <span class="n">tp_mesh</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="n">parallelize_plan</span><span class="o">=</span><span class="p">{</span>
<span class="gp">>>> </span> <span class="s2">"attn"</span><span class="p">:</span> <span class="n">PrepareModuleInput</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">input_layouts</span><span class="o">=</span><span class="p">(</span><span class="n">Shard</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span> <span class="kc">None</span><span class="p">,</span> <span class="kc">None</span><span class="p">,</span> <span class="o">...</span><span class="p">),</span>
<span class="gp">>>> </span> <span class="n">desired_input_layouts</span><span class="o">=</span><span class="p">(</span><span class="n">Replicate</span><span class="p">(),</span> <span class="kc">None</span><span class="p">,</span> <span class="kc">None</span><span class="p">,</span> <span class="o">...</span><span class="p">)</span>
<span class="gp">>>> </span> <span class="p">),</span>
<span class="gp">>>> </span> <span class="p">}</span>
<span class="gp">>>> </span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.tensor.parallel.PrepareModuleOutput">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.tensor.parallel.</span></span><span class="sig-name descname"><span class="pre">PrepareModuleOutput</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">output_layouts</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">desired_output_layouts</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">use_local_output</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/tensor/parallel/style.html#PrepareModuleOutput"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/tensor/parallel/style.py#L561"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.tensor.parallel.PrepareModuleOutput" title="Permalink to this definition">¶</a></dt>
<dd><p>Configure the nn.Module’s outputs to convert the output tensors of the nn.Module to DTensors at runtime according to
<code class="docutils literal notranslate"><span class="pre">output_layouts</span></code>, and perform layout redistribution according to the <code class="docutils literal notranslate"><span class="pre">desired_output_layouts</span></code>.</p>
<dl class="field-list simple">
<dt class="field-odd">Keyword Arguments</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>output_layouts</strong> (<em>Union</em><em>[</em><a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>, </em><em>Tuple</em><em>[</em><a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>]</em><em>]</em>) – The DTensor layouts of output tensors for the nn.Module, this is used to convert the output tensors to
DTensors if they are <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a>. If some outputs are not torch.Tensor or no need to convert to DTensors,
<code class="docutils literal notranslate"><span class="pre">None</span></code> need to be specified as a placeholder.</p></li>
<li><p><strong>desired_output_layouts</strong> (<em>Union</em><em>[</em><a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>, </em><em>Tuple</em><em>[</em><a class="reference internal" href="distributed.tensor.html#torch.distributed.tensor.placement_types.Placement" title="torch.distributed.tensor.placement_types.Placement"><em>Placement</em></a><em>]</em><em>]</em>) – The desired DTensor layouts of output tensors for the nn.Module, this is used to ensure the outputs of the nn.Module
have the desired DTensor layouts.</p></li>
<li><p><strong>use_local_output</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><em>bool</em></a><em>, </em><em>optional</em>) – Whether to use local <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a> instead of <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code> for the module outputs, default: True.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>A ParallelStyle object that prepares the sharding layouts of the nn.Module’s outputs.</p>
</dd>
</dl>
<dl>
<dt>Example::</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.tensor.parallel</span> <span class="kn">import</span> <span class="n">parallelize_module</span><span class="p">,</span> <span class="n">PrepareModuleOutput</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.device_mesh</span> <span class="kn">import</span> <span class="n">init_device_mesh</span>
<span class="gp">>>> </span><span class="o">...</span>
<span class="gp">>>> </span><span class="n">block</span> <span class="o">=</span> <span class="n">TransformerBlock</span><span class="p">(</span><span class="o">...</span><span class="p">)</span> <span class="c1"># block is a nn.Module that contains an "attn" Attention submodule</span>
<span class="gp">>>> </span><span class="n">tp_mesh</span> <span class="o">=</span> <span class="n">init_device_mesh</span><span class="p">(</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="p">(</span><span class="mi">8</span><span class="p">,))</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># According to the style specified below, the output of the TransformerBlock will be converted to Replicated DTensor</span>
<span class="gp">>>> </span><span class="c1"># and then redistributed to Sharded DTensor.</span>
<span class="gp">>>> </span><span class="n">parallelize_module</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">block</span><span class="p">,</span> <span class="c1"># this can be a submodule or module</span>
<span class="gp">>>> </span> <span class="n">tp_mesh</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="n">parallelize_plan</span> <span class="o">=</span> <span class="n">PrepareModuleOutput</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">output_layouts</span><span class="o">=</span><span class="n">Replicate</span><span class="p">(),</span>
<span class="gp">>>> </span> <span class="n">desired_output_layouts</span><span class="o">=</span><span class="n">Shard</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
<span class="gp">>>> </span> <span class="p">)</span>
<span class="gp">>>> </span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
</dd></dl>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>when using the <code class="docutils literal notranslate"><span class="pre">Shard(dim)</span></code> as the input/output layouts for the above
<code class="docutils literal notranslate"><span class="pre">ParallelStyle</span></code> s, we assume the input/output activation tensors are evenly sharded on
the tensor dimension <code class="docutils literal notranslate"><span class="pre">dim</span></code> on the <code class="docutils literal notranslate"><span class="pre">DeviceMesh</span></code> that TP operates on. For instance,
since <code class="docutils literal notranslate"><span class="pre">RowwiseParallel</span></code> accepts input that is sharded on the last dimension, it assumes
the input tensor has already been evenly sharded on the last dimension. For the case of uneven
sharded activation tensors, one could pass in DTensor directly to the partitioned modules,
and use <code class="docutils literal notranslate"><span class="pre">use_local_output=False</span></code> to return DTensor after each <code class="docutils literal notranslate"><span class="pre">ParallelStyle</span></code>, where
DTensor could track the uneven sharding information.</p>
</div>
<p>For models like Transformer, we recommend users to use <code class="docutils literal notranslate"><span class="pre">ColwiseParallel</span></code>
and <code class="docutils literal notranslate"><span class="pre">RowwiseParallel</span></code> together in the parallelize_plan for achieve the desired
sharding for the entire model (i.e. Attention and MLP).</p>
<p>Parallelized cross-entropy loss computation (loss parallelism), is supported via the following context manager:</p>
<dl class="py function">
<dt class="sig sig-object py" id="torch.distributed.tensor.parallel.loss_parallel">
<span class="sig-prename descclassname"><span class="pre">torch.distributed.tensor.parallel.</span></span><span class="sig-name descname"><span class="pre">loss_parallel</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/tensor/parallel/loss.html#loss_parallel"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/tensor/parallel/loss.py#L29"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.tensor.parallel.loss_parallel" title="Permalink to this definition">¶</a></dt>
<dd><p>A context manager that enables loss parallelism, where efficient parallelized loss computation
can be performed when the input is sharded on the class dimension. Currently only the cross-entropy
loss is supported.</p>
<p>Within this context manager, one can use <a class="reference internal" href="generated/torch.nn.functional.cross_entropy.html#torch.nn.functional.cross_entropy" title="torch.nn.functional.cross_entropy"><code class="xref py py-func docutils literal notranslate"><span class="pre">cross_entropy()</span></code></a> or
<a class="reference internal" href="generated/torch.nn.CrossEntropyLoss.html#torch.nn.CrossEntropyLoss" title="torch.nn.CrossEntropyLoss"><code class="xref py py-class docutils literal notranslate"><span class="pre">CrossEntropyLoss</span></code></a> as usual, with the following assumptions on the input parameters.
The corresponding <code class="docutils literal notranslate"><span class="pre">backward()</span></code> call, if any, also needs to happen under this context manager.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>input</strong> (<code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code>) – Input logits. Assumed to be sharded on the class dimension.</p></li>
<li><p><strong>target</strong> (Union[<a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a>, <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code>]) – Must be ground truth class indices (class probabilities currently not supported).
Assumed to be replicated across the <code class="docutils literal notranslate"><span class="pre">DeviceMesh</span></code>.</p></li>
<li><p><strong>weight</strong> (Union[<a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a>, <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code>], optional) – If given, assumed to be replicated across the <code class="docutils literal notranslate"><span class="pre">DeviceMesh</span></code>.</p></li>
<li><p><strong>label_smoothing</strong> – Currently not supported.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>A replicated <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code>.</p>
</dd>
</dl>
<p class="rubric">Example</p>
<p>A sharded DTensor is manually created here to showcase the usage.
In practice, it is usually the output of a TP module.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.tensor.parallel</span> <span class="kn">import</span> <span class="n">loss_parallel</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.device_mesh</span> <span class="kn">import</span> <span class="n">init_device_mesh</span>
<span class="gp">>>> </span><span class="o">...</span>
<span class="gp">>>> </span><span class="n">device_mesh</span> <span class="o">=</span> <span class="n">init_device_mesh</span><span class="p">(</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="p">(</span><span class="mi">8</span><span class="p">,))</span>
<span class="gp">>>> </span><span class="nb">input</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">16</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">dist_input</span> <span class="o">=</span> <span class="n">distribute_tensor</span><span class="p">(</span><span class="nb">input</span><span class="p">,</span> <span class="n">device_mesh</span><span class="p">,</span> <span class="n">placements</span><span class="o">=</span><span class="p">[</span><span class="n">Shard</span><span class="p">(</span><span class="mi">1</span><span class="p">)])</span>
<span class="gp">>>> </span><span class="n">target</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randint</span><span class="p">(</span><span class="mi">16</span><span class="p">,</span> <span class="p">(</span><span class="mi">4</span><span class="p">,),</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="gp">>>> </span><span class="k">with</span> <span class="n">loss_parallel</span><span class="p">():</span>
<span class="gp">>>> </span> <span class="n">loss</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">cross_entropy</span><span class="p">(</span><span class="n">dist_input</span><span class="p">,</span> <span class="n">target</span><span class="p">,</span> <span class="n">reduction</span><span class="o">=</span><span class="s2">"mean"</span><span class="p">)</span>
<span class="gp">>>> </span> <span class="n">loss</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="gp">>>> </span><span class="o">...</span>
</pre></div>
</div>
</dd></dl>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>The loss_parallel API is experimental and subject to change.</p>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="distributed.optim.html" class="btn btn-neutral float-right" title="Distributed Optimizers" accesskey="n" rel="next">Next <img src="_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="distributed.fsdp.fully_shard.html" class="btn btn-neutral" title="torch.distributed.fsdp.fully_shard" accesskey="p" rel="prev"><img src="_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright PyTorch Contributors.
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
<script>
var match = window.location.href.match(/\/_[a-zA-Z0-9_]*.html|_dynamo/gi);
var url = window.location.href.lastIndexOf(match[match.length-1]);
if (url)
{
var div = '<div class="admonition note"><p class="admonition-title">Note</p><p><i class="fa fa-exclamation-circle" aria-hidden="true"> </i> This page describes an internal API which is not intended to be used outside of the PyTorch codebase and can be modified or removed without notice.</p></div>'
document.getElementById("pytorch-article").insertAdjacentHTML('afterBegin', div)
}
</script>