forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfsdp.html
1792 lines (1573 loc) · 154 KB
/
fsdp.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="generator" content="Docutils 0.18.1: http://docutils.sourceforge.net/" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>FullyShardedDataParallel — PyTorch master documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/fsdp.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="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="TorchElastic Kubernetes" href="elastic/kubernetes.html" />
<!--
Search engines should not index the master version of documentation.
Stable documentation are built without release == 'master'.
-->
<meta name="robots" content="noindex">
<!-- Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-117752657-2"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-117752657-2');
</script>
<!-- End Google Analytics -->
<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>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<a href="https://pytorch.org/mobile">Mobile</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="active docs-active">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-orange-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/audio/stable/index.html">
<span class="dropdown-title">torchaudio</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/text/stable/index.html">
<span class="dropdown-title">torchtext</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/vision/stable/index.html">
<span class="dropdown-title">torchvision</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torcharrow">
<span class="dropdown-title">torcharrow</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/data">
<span class="dropdown-title">TorchData</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torchrec">
<span class="dropdown-title">TorchRec</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/serve/">
<span class="dropdown-title">TorchServe</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torchx/">
<span class="dropdown-title">TorchX</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/xla">
<span class="dropdown-title">PyTorch on XLA Devices</span>
<p></p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-arrow">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class="dropdown-title">About</span>
<p>Learn about PyTorch’s features and capabilities</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn about the PyTorch foundation</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class="dropdown-title">Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered.</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/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/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/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class="dropdown-title">Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/hub">
<span class="dropdown-title">Models (Beta)</span>
<p>Discover, publish, and reuse pre-trained models</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">GitHub</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>master (1.14.0a0+git876b702 ) ▼</a>
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<div>
<a style="color:#F05732" href="https://pytorch.org/docs/stable/fsdp.html">
You are viewing unstable developer preview docs.
Click here to view docs for latest stable release.
</a>
</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">CUDA Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/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">torch.compile</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="dynamo/index.html">TorchDynamo Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/installation.html">Installing TorchDynamo</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/get-started.html">Getting Started</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/guards-overview.html">Guards Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/custom-backends.html">Custom Backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/deep-dive.html">TorchDynamo Deeper Dive</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/troubleshooting.html">TorchDynamo Troubleshooting</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo/faq.html">Frequently Asked Questions</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="cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">torch.distributed.fsdp</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.tensor.parallel.html">torch.distributed.tensor.parallel</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="_dynamo.html">torch._dynamo</a></li>
<li class="toctree-l1"><a class="reference internal" href="fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="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="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx_diagnostics.html">torch.onnx diagnostics</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="pipeline.html">Pipeline Parallelism</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="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="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="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="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>
</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="http://pytorch.org/xla/">PyTorch on XLA Devices</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>FullyShardedDataParallel</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/fsdp.rst.txt" rel="nofollow"><img src="_static/images/view-page-source-icon.svg"></a>
</li>
</ul>
</div>
</div>
<div class="pytorch-shortcuts-wrapper" id="pytorch-shortcuts-wrapper">
Shortcuts
</div>
</div>
<section data-toggle="wy-nav-shift" id="pytorch-content-wrap" class="pytorch-content-wrap">
<div class="pytorch-content-left">
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<section id="module-torch.distributed.fsdp">
<span id="fullyshardeddataparallel"></span><h1>FullyShardedDataParallel<a class="headerlink" href="#module-torch.distributed.fsdp" title="Permalink to this heading">¶</a></h1>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.fsdp.</span></span><span class="sig-name descname"><span class="pre">FullyShardedDataParallel</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">process_group</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">sharding_strategy</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">cpu_offload</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">auto_wrap_policy</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">backward_prefetch</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">BackwardPrefetch.BACKWARD_PRE</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">mixed_precision</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">ignored_modules</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">param_init_fn</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">device_id</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">sync_module_states</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">forward_prefetch</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">limit_all_gathers</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">use_orig_params</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/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel" title="Permalink to this definition">¶</a></dt>
<dd><p>A wrapper for sharding Module parameters across data parallel workers. This
is inspired by <a class="reference external" href="https://arxiv.org/abs/2004.13336">Xu et al.</a> as well as the ZeRO Stage 3 from <a class="reference external" href="https://www.deepspeed.ai/">DeepSpeed</a>.
FullyShardedDataParallel is commonly shortened to FSDP.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.distributed.fsdp</span> <span class="kn">import</span> <span class="n">FullyShardedDataParallel</span> <span class="k">as</span> <span class="n">FSDP</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">set_device</span><span class="p">(</span><span class="n">device_id</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">sharded_module</span> <span class="o">=</span> <span class="n">FSDP</span><span class="p">(</span><span class="n">my_module</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">optim</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">optim</span><span class="o">.</span><span class="n">Adam</span><span class="p">(</span><span class="n">sharded_module</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.0001</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">x</span> <span class="o">=</span> <span class="n">sharded_module</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="o">=</span><span class="mi">3</span><span class="p">,</span> <span class="n">z</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">([</span><span class="mi">1</span><span class="p">]))</span>
<span class="gp">>>> </span><span class="n">loss</span> <span class="o">=</span> <span class="n">x</span><span class="o">.</span><span class="n">sum</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="n">optim</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>The optimizer must be initialized <em>after</em> the module has been wrapped,
since FSDP will shard parameters in-place and this will break any
previously initialized optimizers.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>If the destination CUDA device has ID <code class="docutils literal notranslate"><span class="pre">dev_id</span></code>, either (1)
<code class="docutils literal notranslate"><span class="pre">module</span></code> should already be placed on that device, (2) the device
should be set using <code class="docutils literal notranslate"><span class="pre">torch.cuda.set_device(dev_id)</span></code>, or (3)
<code class="docutils literal notranslate"><span class="pre">dev_id</span></code> should be passed into the <code class="docutils literal notranslate"><span class="pre">device_id</span></code> constructor
argument. This FSDP instance’s compute device will be that destination
device. For (1) and (3), the FSDP initialization always occurs on GPU.
For (2), the FSDP initialization happens on <code class="docutils literal notranslate"><span class="pre">module</span></code> ‘s current
device, which may be CPU.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>FSDP currently does not support gradient accumulation outside
<code class="docutils literal notranslate"><span class="pre">no_sync()</span></code> when using CPU offloading. Trying to do so yields
incorrect results since FSDP will use the newly-reduced gradient
instead of accumulating with any existing gradient.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Changing the original parameter variable names after construction will
lead to undefined behavior.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Passing in <cite>sync_module_states=True</cite> flag requires module to be put
on GPU, or to use <code class="docutils literal notranslate"><span class="pre">device_id</span></code> argument to specify a CUDA device that
FSDP will move module to. This is because <code class="docutils literal notranslate"><span class="pre">sync_module_states=True</span></code>
requires GPU communication.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>As of PyTorch 1.12, FSDP only offers limited support for shared parameters
(for example, setting one <code class="docutils literal notranslate"><span class="pre">Linear</span></code> layer’s weight to another’s). In
particular, modules that share parameters must be wrapped as part of the
same FSDP unit. If enhanced shared parameter support is needed for your
use case, please ping <a class="reference external" href="https://github.com/pytorch/pytorch/issues/77724">https://github.com/pytorch/pytorch/issues/77724</a></p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Inputs into FSDP <code class="docutils literal notranslate"><span class="pre">forward</span></code> function will be moved to compute device
(same device FSDP module is on) before running <code class="docutils literal notranslate"><span class="pre">forward</span></code>, so user does
not have to manually move inputs from CPU -> GPU.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>module</strong> (<a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><em>nn.Module</em></a>) – This is the module to be wrapped with FSDP.</p></li>
<li><p><strong>process_group</strong> (<em>Optional</em><em>[</em><em>ProcessGroup</em><em>]</em>) – This is the process group used for collective communications.</p></li>
<li><p><strong>sharding_strategy</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.fsdp.ShardingStrategy" title="torch.distributed.fsdp.ShardingStrategy"><em>ShardingStrategy</em></a><em>]</em>) – This configures the sharding strategy used by FSDP, which may trade
off memory saving and communication overhead. See
<a class="reference internal" href="#torch.distributed.fsdp.ShardingStrategy" title="torch.distributed.fsdp.ShardingStrategy"><code class="xref py py-class docutils literal notranslate"><span class="pre">ShardingStrategy</span></code></a> for details. (Default: <code class="docutils literal notranslate"><span class="pre">FULL_SHARD</span></code>)</p></li>
<li><p><strong>cpu_offload</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.fsdp.CPUOffload" title="torch.distributed.fsdp.CPUOffload"><em>CPUOffload</em></a><em>]</em>) – This configures CPU offloading. If this is set to <code class="docutils literal notranslate"><span class="pre">None</span></code>, then
no CPU offloading happens. See <a class="reference internal" href="#torch.distributed.fsdp.CPUOffload" title="torch.distributed.fsdp.CPUOffload"><code class="xref py py-class docutils literal notranslate"><span class="pre">CPUOffload</span></code></a> for details.
(Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>auto_wrap_policy</strong> (<em>Optional</em><em>[</em><em>Union</em><em>[</em><em>Callable</em><em>[</em><em>[</em><a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><em>nn.Module</em></a><em>, </em><a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.11)"><em>bool</em></a><em>, </em><a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.11)"><em>int</em></a><em>]</em><em>, </em><a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.11)"><em>bool</em></a><em>]</em><em>, </em><em>_FSDPPolicy</em><em>]</em><em>]</em>) – <p>This is either <code class="docutils literal notranslate"><span class="pre">None</span></code>, an <code class="docutils literal notranslate"><span class="pre">_FSDPPolicy</span></code>, or a callable of
a fixed signature. If it is <code class="docutils literal notranslate"><span class="pre">None</span></code>, then <code class="docutils literal notranslate"><span class="pre">module</span></code> is wrapped
with only a top-level FSDP instance without any nested wrapping. If
it is an <code class="docutils literal notranslate"><span class="pre">_FSDPPolicy</span></code>, then the wrapping follows the given
policy. <code class="docutils literal notranslate"><span class="pre">ModuleWrapPolicy</span></code> in <code class="docutils literal notranslate"><span class="pre">torch.distributed.fsdp.wrap.py</span></code>
is an example. If it is a callable, then it should take in three
arguments <code class="docutils literal notranslate"><span class="pre">module:</span> <span class="pre">nn.Module</span></code>, <code class="docutils literal notranslate"><span class="pre">recurse:</span> <span class="pre">bool</span></code>, and
<code class="docutils literal notranslate"><span class="pre">nonwrapped_numel:</span> <span class="pre">int</span></code> and should return a <code class="docutils literal notranslate"><span class="pre">bool</span></code> specifying
whether the passed-in <code class="docutils literal notranslate"><span class="pre">module</span></code> should be wrapped if
<code class="docutils literal notranslate"><span class="pre">recurse=False</span></code> or if the traversal should continue down the
subtree if <code class="docutils literal notranslate"><span class="pre">recurse=True</span></code>. Additional custom arguments may be
added to the callable. The <code class="docutils literal notranslate"><span class="pre">size_based_auto_wrap_policy</span></code> in
<code class="docutils literal notranslate"><span class="pre">torch.distributed.fsdp.wrap.py</span></code> gives an example callable that
wraps a module if the parameters in its subtree exceed 100M numel.
A good practice is to print the model after wrapping and adjust as
needed.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="k">def</span> <span class="nf">custom_auto_wrap_policy</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">module</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="gp">>>> </span> <span class="n">recurse</span><span class="p">:</span> <span class="nb">bool</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="n">nonwrapped_numel</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="c1"># Additional custom arguments</span>
<span class="gp">>>> </span> <span class="n">min_num_params</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="mf">1e8</span><span class="p">),</span>
<span class="gp">>>> </span><span class="p">)</span> <span class="o">-></span> <span class="nb">bool</span><span class="p">:</span>
<span class="gp">>>> </span> <span class="k">return</span> <span class="n">nonwrapped_numel</span> <span class="o">>=</span> <span class="n">min_num_params</span>
<span class="gp">>>> </span><span class="c1"># Configure a custom `min_num_params`</span>
<span class="gp">>>> </span><span class="n">my_auto_wrap_policy</span> <span class="o">=</span> <span class="n">functools</span><span class="o">.</span><span class="n">partial</span><span class="p">(</span><span class="n">custom_auto_wrap_policy</span><span class="p">,</span> <span class="n">min_num_params</span><span class="o">=</span><span class="nb">int</span><span class="p">(</span><span class="mf">1e5</span><span class="p">))</span>
</pre></div>
</div>
</p></li>
<li><p><strong>backward_prefetch</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.fsdp.BackwardPrefetch" title="torch.distributed.fsdp.BackwardPrefetch"><em>BackwardPrefetch</em></a><em>]</em>) – This configures explicit backward prefetching of all-gathers. See
<a class="reference internal" href="#torch.distributed.fsdp.BackwardPrefetch" title="torch.distributed.fsdp.BackwardPrefetch"><code class="xref py py-class docutils literal notranslate"><span class="pre">BackwardPrefetch</span></code></a> for details. (Default: <code class="docutils literal notranslate"><span class="pre">BACKWARD_PRE</span></code>)</p></li>
<li><p><strong>mixed_precision</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.fsdp.MixedPrecision" title="torch.distributed.fsdp.MixedPrecision"><em>MixedPrecision</em></a><em>]</em>) – This configures native mixed precision for FSDP. If this is set to
<code class="docutils literal notranslate"><span class="pre">None</span></code>, then no mixed precision is used. Otherwise, parameter,
buffer, and gradient reduction dtypes can be set. See
<a class="reference internal" href="#torch.distributed.fsdp.MixedPrecision" title="torch.distributed.fsdp.MixedPrecision"><code class="xref py py-class docutils literal notranslate"><span class="pre">MixedPrecision</span></code></a> for details. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>ignored_modules</strong> (<em>Optional</em><em>[</em><em>Iterable</em><em>[</em><a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><em>torch.nn.Module</em></a><em>]</em><em>]</em>) – Modules whose
own parameters and child modules’ parameters and buffers are
ignored by this instance. None of the modules directly in
<code class="docutils literal notranslate"><span class="pre">ignored_modules</span></code> should be <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel" title="torch.distributed.fsdp.FullyShardedDataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">FullyShardedDataParallel</span></code></a>
instances, and any child modules that are already-constructed
<a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel" title="torch.distributed.fsdp.FullyShardedDataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">FullyShardedDataParallel</span></code></a> instances will not be ignored if
they are nested under this instance. This argument may be used to
avoid sharding specific parameters at module granularity when using an
<code class="docutils literal notranslate"><span class="pre">auto_wrap_policy</span></code> or if parameters’ sharding is not managed by
FSDP. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>param_init_fn</strong> (<em>Optional</em><em>[</em><em>Callable</em><em>[</em><em>[</em><a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><em>nn.Module</em></a><em>]</em><em>, </em><em>None</em><em>]</em><em>]</em>) – <p>A <code class="docutils literal notranslate"><span class="pre">Callable[torch.nn.Module]</span> <span class="pre">-></span> <span class="pre">None</span></code> that
specifies how modules that are currently on the meta device should be initialized
onto an actual device. Note that as of v1.12, we detect modules on the meta
device via <code class="docutils literal notranslate"><span class="pre">is_meta</span></code> check and apply a default initialization that calls
<code class="docutils literal notranslate"><span class="pre">reset_parameters</span></code> method on the passed in <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code> if <code class="docutils literal notranslate"><span class="pre">param_init_fn</span></code>
is not specified, otherwise we run <code class="docutils literal notranslate"><span class="pre">param_init_fn</span></code> to initialize the passed
in <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code>. In particular, this means that if <code class="docutils literal notranslate"><span class="pre">is_meta=True</span></code> for any
module parameters for modules that will be wrapped with FSDP and <code class="docutils literal notranslate"><span class="pre">param_init_fn</span></code>
is not specified, we assume your module properly implements a <code class="docutils literal notranslate"><span class="pre">reset_parameters()</span></code>
and will throw errors if not. Note that additionally, we offer support for modules
initialized with torchdistX’s (<a class="reference external" href="https://github.com/pytorch/torchdistX">https://github.com/pytorch/torchdistX</a>)
<code class="docutils literal notranslate"><span class="pre">deferred_init</span></code> API. In this case, deferred modules would be initialized
by a default initialization function that calls torchdistX’s
<code class="docutils literal notranslate"><span class="pre">materialize_module</span></code>, or the passed in <code class="docutils literal notranslate"><span class="pre">param_init_fn</span></code>, if it is not
<code class="docutils literal notranslate"><span class="pre">None</span></code>. The same <code class="docutils literal notranslate"><span class="pre">Callable</span></code> is applied to initialize all meta modules.
Note that this initialization function is applied before doing any FSDP sharding
logic.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">module</span> <span class="o">=</span> <span class="n">MyModule</span><span class="p">(</span><span class="n">device</span><span class="o">=</span><span class="s2">"meta"</span><span class="p">)</span>
<span class="gp">>>> </span><span class="k">def</span> <span class="nf">my_init_fn</span><span class="p">(</span><span class="n">module</span><span class="p">):</span>
<span class="gp">>>> </span> <span class="c1"># responsible for initializing a module, such as with reset_parameters</span>
<span class="gp">>>> </span> <span class="o">...</span>
<span class="gp">>>> </span><span class="n">fsdp_model</span> <span class="o">=</span> <span class="n">FSDP</span><span class="p">(</span><span class="n">module</span><span class="p">,</span> <span class="n">param_init_fn</span><span class="o">=</span><span class="n">my_init_fn</span><span class="p">,</span> <span class="n">auto_wrap_policy</span><span class="o">=</span><span class="n">size_based_auto_wrap_policy</span><span class="p">)</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="nb">next</span><span class="p">(</span><span class="n">fsdp_model</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="p">)</span> <span class="c1"># current CUDA device</span>
<span class="gp">>>> </span><span class="c1"># With torchdistX</span>
<span class="gp">>>> </span><span class="n">module</span> <span class="o">=</span> <span class="n">deferred_init</span><span class="o">.</span><span class="n">deferred_init</span><span class="p">(</span><span class="n">MyModule</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="c1"># Will initialize via deferred_init.materialize_module().</span>
<span class="gp">>>> </span><span class="n">fsdp_model</span> <span class="o">=</span> <span class="n">FSDP</span><span class="p">(</span><span class="n">module</span><span class="p">,</span> <span class="n">auto_wrap_policy</span><span class="o">=</span><span class="n">size_based_auto_wrap_policy</span><span class="p">)</span>
</pre></div>
</div>
</p></li>
<li><p><strong>device_id</strong> (<em>Optional</em><em>[</em><em>Union</em><em>[</em><a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.11)"><em>int</em></a><em>, </em><a class="reference internal" href="tensor_attributes.html#torch.device" title="torch.device"><em>torch.device</em></a><em>]</em><em>]</em>) – An <code class="docutils literal notranslate"><span class="pre">int</span></code> or <code class="docutils literal notranslate"><span class="pre">torch.device</span></code>
describing the CUDA device the FSDP module should be moved to determining where
initialization such as sharding takes place. If this argument is not specified
and <code class="docutils literal notranslate"><span class="pre">module</span></code> is on CPU, we issue a warning mentioning that this argument can
be specified for faster initialization. If specified, resulting FSDP instances
will reside on this device, including moving ignored modules’ parameters if
needed. Note that if <code class="docutils literal notranslate"><span class="pre">device_id</span></code> is specified but <code class="docutils literal notranslate"><span class="pre">module</span></code> is already on a
different CUDA device, an error will be thrown. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>sync_module_states</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.11)"><em>bool</em></a>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, each individually wrapped FSDP unit will broadcast
module parameters from rank 0 to ensure they are the same across all ranks after
initialization. This helps ensure model parameters are the same across ranks
before starting training, but adds communication overhead to <code class="docutils literal notranslate"><span class="pre">__init__</span></code>, as at least
one broadcast is triggered per individually wrapped FSDP unit.
This can also help load checkpoints taken by <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> and to be loaded by
<code class="docutils literal notranslate"><span class="pre">load_state_dict</span></code> in a memory efficient way. See documentation for
<code class="xref py py-class docutils literal notranslate"><span class="pre">FullStateDictConfig</span></code> for an example of this. (Default: <code class="docutils literal notranslate"><span class="pre">False</span></code>)</p></li>
<li><p><strong>forward_prefetch</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.11)"><em>bool</em></a>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, then FSDP <em>explicitly</em> prefetches
the next upcoming all-gather while executing in the forward pass.
This may improve communication and computation overlap for CPU
bound workloads. This should only be used for static graph models
since the forward order is fixed based on the first iteration’s
execution. (Default: <code class="docutils literal notranslate"><span class="pre">False</span></code>)</p></li>
<li><p><strong>limit_all_gathers</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.11)"><em>bool</em></a>) – If <code class="docutils literal notranslate"><span class="pre">False</span></code>, then FSDP allows the CPU
thread to schedule all-gathers without any extra synchronization.
If <code class="docutils literal notranslate"><span class="pre">True</span></code>, then FSDP explicitly synchronizes the CPU thread to
prevent too many in-flight all-gathers. This <code class="docutils literal notranslate"><span class="pre">bool</span></code> only affects
the sharded strategies that schedule all-gathers. Enabling this can
help lower the number of CUDA malloc retries.</p></li>
</ul>
</dd>
</dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.apply">
<span class="sig-name descname"><span class="pre">apply</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">fn</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.apply"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.apply" title="Permalink to this definition">¶</a></dt>
<dd><p>Applies <code class="docutils literal notranslate"><span class="pre">fn</span></code> recursively to every submodule (as returned by <code class="docutils literal notranslate"><span class="pre">.children()</span></code>)
as well as self. Typical use includes initializing the parameters of a model
(see also <a class="reference internal" href="nn.init.html#nn-init-doc"><span class="std std-ref">torch.nn.init</span></a>).</p>
<p>Compared to <code class="docutils literal notranslate"><span class="pre">torch.nn.Module.apply</span></code>, this version additionally gathers
the full parameters before applying <code class="docutils literal notranslate"><span class="pre">fn</span></code>. It should not be called from
within another <code class="docutils literal notranslate"><span class="pre">summon_full_params</span></code> context.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>fn</strong> (<code class="xref py py-class docutils literal notranslate"><span class="pre">Module</span></code> -> None) – function to be applied to each submodule</p>
</dd>
<dt class="field-even">Returns<span class="colon">:</span></dt>
<dd class="field-even"><p>self</p>
</dd>
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p><p id="torch.nn.Module"/><a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module">Module</a></p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.clip_grad_norm_">
<span class="sig-name descname"><span class="pre">clip_grad_norm_</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">max_norm</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">norm_type</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">2.0</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.clip_grad_norm_"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.clip_grad_norm_" title="Permalink to this definition">¶</a></dt>
<dd><p>Clips the gradient norm of all parameters. The norm is computed over
all parameters’ gradients as viewed as a single vector, and the
gradients are modified in-place.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>max_norm</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#float" title="(in Python v3.11)"><em>float</em></a><em> or </em><a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.11)"><em>int</em></a>) – max norm of the gradients</p></li>
<li><p><strong>norm_type</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#float" title="(in Python v3.11)"><em>float</em></a><em> or </em><a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.11)"><em>int</em></a>) – type of the used p-norm. Can be <code class="docutils literal notranslate"><span class="pre">'inf'</span></code>
for infinity norm.</p></li>
</ul>
</dd>
<dt class="field-even">Returns<span class="colon">:</span></dt>
<dd class="field-even"><p>Total norm of the parameters (viewed as a single vector).</p>
</dd>
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p><a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>Tensor</em></a></p>
</dd>
</dl>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>If every FSDP instance uses <code class="docutils literal notranslate"><span class="pre">NO_SHARD</span></code>, meaning that no
gradients are sharded across ranks, then you may directly use
<a class="reference internal" href="generated/torch.nn.utils.clip_grad_norm_.html#torch.nn.utils.clip_grad_norm_" title="torch.nn.utils.clip_grad_norm_"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.nn.utils.clip_grad_norm_()</span></code></a>.</p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>If at least some FSDP instance uses a sharded strategy (i.e.
one other than <code class="docutils literal notranslate"><span class="pre">NO_SHARD</span></code>), then you should use this method
instead of <a class="reference internal" href="generated/torch.nn.utils.clip_grad_norm_.html#torch.nn.utils.clip_grad_norm_" title="torch.nn.utils.clip_grad_norm_"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.nn.utils.clip_grad_norm_()</span></code></a> since this method
handles the fact that gradients are sharded across ranks.</p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>The total norm returned will have the “largest” dtype across
all parameters/gradients as defined by PyTorch’s type promotion
semantics. For example, if <em>all</em> parameters/gradients use a low
precision dtype, then the returned norm’s dtype will be that low
precision dtype, but if there exists at least one parameter/
gradient using FP32, then the returned norm’s dtype will be FP32.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>This needs to be called on all ranks since it uses
collective communications.</p>
</div>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.flatten_sharded_optim_state_dict">
<em class="property"><span class="pre">static</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">flatten_sharded_optim_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">sharded_optim_state_dict</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">model</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">optim_input</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">optim</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.flatten_sharded_optim_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.flatten_sharded_optim_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>The API is similar to <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.shard_full_optim_state_dict" title="torch.distributed.fsdp.FullyShardedDataParallel.shard_full_optim_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">shard_full_optim_state_dict()</span></code></a>. The only
difference is that the input <code class="docutils literal notranslate"><span class="pre">sharded_optim_state_dict</span></code> should be
returned from <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.sharded_optim_state_dict" title="torch.distributed.fsdp.FullyShardedDataParallel.sharded_optim_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">sharded_optim_state_dict()</span></code></a>. Therefore, there will
be all-gather calls on each rank to gather <code class="docutils literal notranslate"><span class="pre">ShardedTensor</span></code> s.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>sharded_optim_state_dict</strong> (<em>Dict</em><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.11)"><em>str</em></a><em>, </em><em>Any</em><em>]</em>) – Optimizer state dict
corresponding to the unflattened parameters and holding the
sharded optimizer state.</p></li>
<li><p><strong>model</strong> (<a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><em>torch.nn.Module</em></a>) – Refer to :meth:<code class="docutils literal notranslate"><span class="pre">shard_full_optim_state_dict</span></code>.</p></li>
</ul>
</dd>
<dt class="field-even">Returns<span class="colon">:</span></dt>
<dd class="field-even"><p>Refer to <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.shard_full_optim_state_dict" title="torch.distributed.fsdp.FullyShardedDataParallel.shard_full_optim_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">shard_full_optim_state_dict()</span></code></a>.</p>
</dd>
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Dict" title="(in Python v3.11)"><em>Dict</em></a>[<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.11)">str</a>, <a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Any" title="(in Python v3.11)"><em>Any</em></a>]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.forward">
<span class="sig-name descname"><span class="pre">forward</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.forward"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.forward" title="Permalink to this definition">¶</a></dt>
<dd><p>Runs the forward pass for the wrapped module, inserting FSDP-specific
pre- and post-forward sharding logic.</p>
<dl class="field-list simple">
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Any" title="(in Python v3.11)"><em>Any</em></a></p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.fsdp_modules">
<em class="property"><span class="pre">static</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">fsdp_modules</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">root_only</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/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.fsdp_modules"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.fsdp_modules" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns all nested FSDP instances, possibly including <code class="docutils literal notranslate"><span class="pre">module</span></code> itself
and only including FSDP root modules if <code class="docutils literal notranslate"><span class="pre">root_only=True</span></code>.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>module</strong> (<a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><em>torch.nn.Module</em></a>) – Root module, which may or may not be an
<code class="docutils literal notranslate"><span class="pre">FSDP</span></code> module.</p></li>
<li><p><strong>root_only</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.11)"><em>bool</em></a>) – Whether to return only FSDP root modules.
(Default: <code class="docutils literal notranslate"><span class="pre">False</span></code>)</p></li>
</ul>
</dd>
<dt class="field-even">Returns<span class="colon">:</span></dt>
<dd class="field-even"><p>FSDP modules that are nested in
the input <code class="docutils literal notranslate"><span class="pre">module</span></code>.</p>
</dd>
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p>List[<a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel" title="torch.distributed.fsdp.FullyShardedDataParallel">FullyShardedDataParallel</a>]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.full_optim_state_dict">
<em class="property"><span class="pre">static</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">full_optim_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">model</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">optim</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">optim_input</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">rank0_only</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">group</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.full_optim_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.full_optim_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Consolidates the full optimizer state on rank 0 and returns it
as a <a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a> following the convention of
<a class="reference internal" href="generated/torch.optim.Optimizer.state_dict.html#torch.optim.Optimizer.state_dict" title="torch.optim.Optimizer.state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.optim.Optimizer.state_dict()</span></code></a>, i.e. with keys <code class="docutils literal notranslate"><span class="pre">"state"</span></code>
and <code class="docutils literal notranslate"><span class="pre">"param_groups"</span></code>. The flattened parameters in <code class="docutils literal notranslate"><span class="pre">FSDP</span></code> modules
contained in <code class="docutils literal notranslate"><span class="pre">model</span></code> are mapped back to their unflattened parameters.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>This needs to be called on all ranks since it uses
collective communications. However, if <code class="docutils literal notranslate"><span class="pre">rank0_only=True</span></code>, then
the state dict is only populated on rank 0, and all other ranks
return an empty <a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a>.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Unlike <code class="docutils literal notranslate"><span class="pre">torch.optim.Optimizer.state_dict()</span></code>, this method
uses full parameter names as keys instead of parameter IDs.</p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Like in <a class="reference internal" href="generated/torch.optim.Optimizer.state_dict.html#torch.optim.Optimizer.state_dict" title="torch.optim.Optimizer.state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.optim.Optimizer.state_dict()</span></code></a>, the tensors
contained in the optimizer state dict are not cloned, so there may
be aliasing surprises. For best practices, consider saving the
returned optimizer state dict immediately, e.g. using
<code class="docutils literal notranslate"><span class="pre">torch.save()</span></code>.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>model</strong> (<a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><em>torch.nn.Module</em></a>) – Root module (which may or may not be a
<a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel" title="torch.distributed.fsdp.FullyShardedDataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">FullyShardedDataParallel</span></code></a> instance) whose parameters
were passed into the optimizer <code class="docutils literal notranslate"><span class="pre">optim</span></code>.</p></li>
<li><p><strong>optim</strong> (<a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.Optimizer"><em>torch.optim.Optimizer</em></a>) – Optimizer for <code class="docutils literal notranslate"><span class="pre">model</span></code> ‘s
parameters.</p></li>
<li><p><strong>optim_input</strong> (<em>Optional</em><em>[</em><em>Union</em><em>[</em><em>List</em><em>[</em><em>Dict</em><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.11)"><em>str</em></a><em>, </em><em>Any</em><em>]</em><em>]</em><em>, </em><em>Iterable</em><em>[</em><em>torch.nn.Parameter</em><em>]</em><em>]</em><em>]</em>) – Input passed into the optimizer <code class="docutils literal notranslate"><span class="pre">optim</span></code> representing either a
<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#list" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">list</span></code></a> of parameter groups or an iterable of parameters;
if <code class="docutils literal notranslate"><span class="pre">None</span></code>, then this method assumes the input was
<code class="docutils literal notranslate"><span class="pre">model.parameters()</span></code>. This argument is deprecated, and there
is no need to pass it in anymore. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>rank0_only</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.11)"><em>bool</em></a>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, saves the populated <a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a>
only on rank 0; if <code class="docutils literal notranslate"><span class="pre">False</span></code>, saves it on all ranks. (Default:
<code class="docutils literal notranslate"><span class="pre">True</span></code>)</p></li>
<li><p><strong>group</strong> (<em>dist.ProcessGroup</em>) – Model’s process group or <code class="docutils literal notranslate"><span class="pre">None</span></code> if using
the default process group. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
</ul>
</dd>
<dt class="field-even">Returns<span class="colon">:</span></dt>
<dd class="field-even"><p>A <a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a> containing the optimizer state for
<code class="docutils literal notranslate"><span class="pre">model</span></code> ‘s original unflattened parameters and including keys
“state” and “param_groups” following the convention of
<a class="reference internal" href="generated/torch.optim.Optimizer.state_dict.html#torch.optim.Optimizer.state_dict" title="torch.optim.Optimizer.state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">torch.optim.Optimizer.state_dict()</span></code></a>. If <code class="docutils literal notranslate"><span class="pre">rank0_only=True</span></code>,
then nonzero ranks return an empty <a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a>.</p>
</dd>
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p>Dict[<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.11)">str</a>, Any]</p>
</dd>
</dl>
</dd></dl>
<dl class="py property">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.module">
<em class="property"><span class="pre">property</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">module</span></span><em class="property"><span class="p"><span class="pre">:</span></span><span class="w"> </span><a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.modules.module.Module"><span class="pre">Module</span></a></em><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.module" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns the wrapped module (like <code class="xref py py-class docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code>).</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.named_buffers">
<span class="sig-name descname"><span class="pre">named_buffers</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.named_buffers"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.named_buffers" title="Permalink to this definition">¶</a></dt>
<dd><p>Overrides <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.named_buffers" title="torch.distributed.fsdp.FullyShardedDataParallel.named_buffers"><code class="xref py py-meth docutils literal notranslate"><span class="pre">named_buffers()</span></code></a> to intercept buffer names and
remove all occurrences of the FSDP-specific flattened buffer prefix
when inside the <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.summon_full_params" title="torch.distributed.fsdp.FullyShardedDataParallel.summon_full_params"><code class="xref py py-meth docutils literal notranslate"><span class="pre">summon_full_params()</span></code></a> context manager.</p>
<dl class="field-list simple">
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Iterator" title="(in Python v3.11)"><em>Iterator</em></a>[<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Tuple" title="(in Python v3.11)"><em>Tuple</em></a>[<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.11)">str</a>, <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>Tensor</em></a>]]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.named_parameters">
<span class="sig-name descname"><span class="pre">named_parameters</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.named_parameters"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.named_parameters" title="Permalink to this definition">¶</a></dt>
<dd><p>Overrides <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.named_parameters" title="torch.distributed.fsdp.FullyShardedDataParallel.named_parameters"><code class="xref py py-meth docutils literal notranslate"><span class="pre">named_parameters()</span></code></a> to intercept parameter names and
remove all occurrences of the FSDP-specific flattened parameter prefix
when inside the <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.summon_full_params" title="torch.distributed.fsdp.FullyShardedDataParallel.summon_full_params"><code class="xref py py-meth docutils literal notranslate"><span class="pre">summon_full_params()</span></code></a> context manager.</p>
<dl class="field-list simple">
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Iterator" title="(in Python v3.11)"><em>Iterator</em></a>[<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Tuple" title="(in Python v3.11)"><em>Tuple</em></a>[<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.11)">str</a>, <a class="reference internal" href="generated/torch.nn.parameter.Parameter.html#torch.nn.parameter.Parameter" title="torch.nn.parameter.Parameter"><em>Parameter</em></a>]]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.no_sync">
<span class="sig-name descname"><span class="pre">no_sync</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.no_sync"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.no_sync" title="Permalink to this definition">¶</a></dt>
<dd><p>A context manager to disable gradient synchronizations across FSDP
instances. Within this context, gradients will be accumulated in module
variables, which will later be synchronized in the first
forward-backward pass after exiting the context. This should only be
used on the root FSDP instance and will recursively apply to all
children FSDP instances.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>This likely results in higher memory usage because FSDP will
accumulate the full model gradients (instead of gradient shards)
until the eventual sync.</p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>When used with CPU offloading, the gradients will not be
offloaded to CPU when inside the context manager. Instead, they
will only be offloaded right after the eventual sync.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Return type<span class="colon">:</span></dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Generator" title="(in Python v3.11)"><em>Generator</em></a></p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.register_comm_hook">
<span class="sig-name descname"><span class="pre">register_comm_hook</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">hook</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.register_comm_hook"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.register_comm_hook" title="Permalink to this definition">¶</a></dt>
<dd><p>Registers a communication hook which is an enhancement that provides a
flexible hook to users where they can specify how FSDP aggregates gradients
across multiple workers.
This hook can be used to implement several algorithms like
<a class="reference external" href="https://arxiv.org/abs/1803.05880">GossipGrad</a> and gradient compression
which involve different communication strategies for
parameter syncs while training with <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel" title="torch.distributed.fsdp.FullyShardedDataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">FullyShardedDataParallel</span></code></a>.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>FSDP communication hook should be registered before running an initial forward pass
and only once.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>state</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#object" title="(in Python v3.11)"><em>object</em></a>) – <p>Passed to the hook to maintain any state information during the training process.
Examples include error feedback in gradient compression,
peers to communicate with next in <a class="reference external" href="https://arxiv.org/abs/1803.05880">GossipGrad</a>, etc.
It is locally stored by each worker
and shared by all the gradient tensors on the worker.</p>
</p></li>
<li><p><strong>hook</strong> (<em>Callable</em>) – Callable, which has one of the following signatures:
1) <code class="docutils literal notranslate"><span class="pre">hook:</span> <span class="pre">Callable[torch.Tensor]</span> <span class="pre">-></span> <span class="pre">None</span></code>:
This function takes in a Python tensor, which represents
the full, flattened, unsharded gradient with respect to all variables
corresponding to the model this FSDP unit is wrapping
(that are not wrapped by other FSDP sub-units).
It then performs all necessary processing and returns <code class="docutils literal notranslate"><span class="pre">None</span></code>;
2) <code class="docutils literal notranslate"><span class="pre">hook:</span> <span class="pre">Callable[torch.Tensor,</span> <span class="pre">torch.Tensor]</span> <span class="pre">-></span> <span class="pre">None</span></code>:
This function takes in two Python tensors, the first one represents
the full, flattened, unsharded gradient with respect to all variables
corresponding to the model this FSDP unit is wrapping
(that are not wrapped by other FSDP sub-units). The latter
represents a pre-sized tensor to store a chunk of a sharded gradient after
reduction.
In both cases, callable performs all necessary processing and returns <code class="docutils literal notranslate"><span class="pre">None</span></code>.
Callables with signature 1 are expected to handle gradient communication for a <cite>NO_SHARD</cite> case.
Callables with signature 2 are expected to handle gradient communication for sharded cases.</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.rekey_optim_state_dict">
<em class="property"><span class="pre">static</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">rekey_optim_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">optim_state_dict</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">optim_state_key_type</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">model</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">optim_input</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">optim</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.rekey_optim_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.rekey_optim_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Re-keys the optimizer state dict <code class="docutils literal notranslate"><span class="pre">optim_state_dict</span></code> to use the key
type <code class="docutils literal notranslate"><span class="pre">optim_state_key_type</span></code>. This can be used to achieve
compatibility between optimizer state dicts from models with FSDP
instances and ones without.</p>
<p>To re-key an FSDP full optimizer state dict (i.e. from
<a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.full_optim_state_dict" title="torch.distributed.fsdp.FullyShardedDataParallel.full_optim_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">full_optim_state_dict()</span></code></a>) to use parameter IDs and be loadable to
a non-wrapped model:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">wrapped_model</span><span class="p">,</span> <span class="n">wrapped_optim</span> <span class="o">=</span> <span class="o">...</span>
<span class="gp">>>> </span><span class="n">full_osd</span> <span class="o">=</span> <span class="n">FSDP</span><span class="o">.</span><span class="n">full_optim_state_dict</span><span class="p">(</span><span class="n">wrapped_model</span><span class="p">,</span> <span class="n">wrapped_optim</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nonwrapped_model</span><span class="p">,</span> <span class="n">nonwrapped_optim</span> <span class="o">=</span> <span class="o">...</span>
<span class="gp">>>> </span><span class="n">rekeyed_osd</span> <span class="o">=</span> <span class="n">FSDP</span><span class="o">.</span><span class="n">rekey_optim_state_dict</span><span class="p">(</span><span class="n">full_osd</span><span class="p">,</span> <span class="n">OptimStateKeyType</span><span class="o">.</span><span class="n">PARAM_ID</span><span class="p">,</span> <span class="n">nonwrapped_model</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nonwrapped_optim</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">rekeyed_osd</span><span class="p">)</span>
</pre></div>
</div>
<p>To re-key a normal optimizer state dict from a non-wrapped model to be
loadable to a wrapped model:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">nonwrapped_model</span><span class="p">,</span> <span class="n">nonwrapped_optim</span> <span class="o">=</span> <span class="o">...</span>
<span class="gp">>>> </span><span class="n">osd</span> <span class="o">=</span> <span class="n">nonwrapped_optim</span><span class="o">.</span><span class="n">state_dict</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">rekeyed_osd</span> <span class="o">=</span> <span class="n">FSDP</span><span class="o">.</span><span class="n">rekey_optim_state_dict</span><span class="p">(</span><span class="n">osd</span><span class="p">,</span> <span class="n">OptimStateKeyType</span><span class="o">.</span><span class="n">PARAM_NAME</span><span class="p">,</span> <span class="n">nonwrapped_model</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">wrapped_model</span><span class="p">,</span> <span class="n">wrapped_optim</span> <span class="o">=</span> <span class="o">...</span>
<span class="gp">>>> </span><span class="n">sharded_osd</span> <span class="o">=</span> <span class="n">FSDP</span><span class="o">.</span><span class="n">shard_full_optim_state_dict</span><span class="p">(</span><span class="n">rekeyed_osd</span><span class="p">,</span> <span class="n">wrapped_model</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">wrapped_optim</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">sharded_osd</span><span class="p">)</span>
</pre></div>
</div>
<dl class="field-list simple">
<dt class="field-odd">Returns<span class="colon">:</span></dt>
<dd class="field-odd"><p>The optimizer state dict re-keyed using the
parameter keys specified by <code class="docutils literal notranslate"><span class="pre">optim_state_key_type</span></code>.</p>
</dd>
<dt class="field-even">Return type<span class="colon">:</span></dt>
<dd class="field-even"><p>Dict[<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.11)">str</a>, Any]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.fsdp.FullyShardedDataParallel.scatter_full_optim_state_dict">
<em class="property"><span class="pre">static</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">scatter_full_optim_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">full_optim_state_dict</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">model</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">optim_input</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">optim</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">group</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/fsdp/fully_sharded_data_parallel.html#FullyShardedDataParallel.scatter_full_optim_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.fsdp.FullyShardedDataParallel.scatter_full_optim_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Scatters the full optimizer state dict from rank 0 to all other ranks,
returning the sharded optimizer state dict on each rank. The return
value is the same as <a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.shard_full_optim_state_dict" title="torch.distributed.fsdp.FullyShardedDataParallel.shard_full_optim_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">shard_full_optim_state_dict()</span></code></a>, and on rank
0, the first argument should be the return value of
<a class="reference internal" href="#torch.distributed.fsdp.FullyShardedDataParallel.full_optim_state_dict" title="torch.distributed.fsdp.FullyShardedDataParallel.full_optim_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">full_optim_state_dict()</span></code></a>.</p>
<p>Example:</p>