-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathfsdp.html
1052 lines (877 loc) · 63 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 name="robots" content="noindex">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>FSDP Notes — PyTorch 2.6 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/notes/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="stylesheet" href="../_static/css/custom.css" type="text/css" />
<link rel="index" title="Index" href="../genindex.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="Getting Started on Intel GPU" href="get_start_xpu.html" />
<link rel="prev" title="Frequently Asked Questions" href="faq.html" />
<!-- Google Tag Manager -->
<script>(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':
new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],
j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src=
'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);
})(window,document,'script','dataLayer','GTM-T8XT4PS');</script>
<!-- End Google Tag Manager -->
<script src="../_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Learn
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/get-started">
<span class=dropdown-title>Get Started</span>
<p>Run PyTorch locally or get started quickly with one of the supported cloud platforms</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials">
<span class="dropdown-title">Tutorials</span>
<p>Whats new in PyTorch tutorials</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/basics/intro.html">
<span class="dropdown-title">Learn the Basics</span>
<p>Familiarize yourself with PyTorch concepts and modules</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/recipes/recipes_index.html">
<span class="dropdown-title">PyTorch Recipes</span>
<p>Bite-size, ready-to-deploy PyTorch code examples</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/introyt.html">
<span class="dropdown-title">Intro to PyTorch - YouTube Series</span>
<p>Master PyTorch basics with our engaging YouTube tutorial series</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Ecosystem
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem">
<span class="dropdown-title">Tools</span>
<p>Learn about the tools and frameworks in the PyTorch Ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class=dropdown-title>Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class=dropdown-title>Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class=dropdown-title>Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem/contributor-awards-2023">
<span class="dropdown-title">Contributor Awards - 2023</span>
<p>Award winners announced at this year's PyTorch Conference</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Edge
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/edge">
<span class="dropdown-title">About PyTorch Edge</span>
<p>Build innovative and privacy-aware AI experiences for edge devices</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch-overview">
<span class="dropdown-title">ExecuTorch</span>
<p>End-to-end solution for enabling on-device inference capabilities across mobile and edge devices</p>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p>Explore the documentation for comprehensive guidance on how to use PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/pytorch-domains">
<span class="dropdown-title">PyTorch Domains</span>
<p>Read the PyTorch Domains documentation to learn more about domain-specific libraries</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Blogs & News
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/blog/">
<span class="dropdown-title">PyTorch Blog</span>
<p>Catch up on the latest technical news and happenings</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/community-blog">
<span class="dropdown-title">Community Blog</span>
<p>Stories from the PyTorch ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/videos">
<span class="dropdown-title">Videos</span>
<p>Learn about the latest PyTorch tutorials, new, and more </p>
<a class="nav-dropdown-item" href="https://pytorch.org/community-stories">
<span class="dropdown-title">Community Stories</span>
<p>Learn how our community solves real, everyday machine learning problems with PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
About
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn more about the PyTorch Foundation</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/governing-board">
<span class="dropdown-title">Governing Board</span>
<p></p>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div class="no-dropdown">
<a href="https://pytorch.org/join" data-cta="join">
Become a Member
</a>
</div>
</li>
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="github-icon">
</a>
</div>
</li>
<!--- TODO: This block adds the search icon to the nav bar. We will enable it later.
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="search-icon">
</a>
</div>
</li>
--->
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>2.6 ▼</a>
</div>
<div id="searchBox">
<div class="searchbox" id="googleSearchBox">
<script async src="https://cse.google.com/cse.js?cx=e65585f8c3ea1440e"></script>
<div class="gcse-search"></div>
</div>
<div id="sphinxSearchBox" style="display: none;">
<div role="search">
<form id="rtd-search-form" class="wy-form" action="../search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
</div>
<form id="searchForm">
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="google" checked>
Google Search
</label>
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="sphinx">
Classic Search
</label>
</form>
<script>
document.addEventListener('DOMContentLoaded', function() {
const searchForm = document.getElementById('searchForm');
const googleSearchBox = document.getElementById('googleSearchBox');
const sphinxSearchBox = document.getElementById('sphinxSearchBox');
// Function to toggle search box visibility
function toggleSearchBox(searchType) {
googleSearchBox.style.display = searchType === 'google' ? 'block' : 'none';
sphinxSearchBox.style.display = searchType === 'sphinx' ? 'block' : 'none';
}
// Determine the default search type
let defaultSearchType;
const currentUrl = window.location.href;
if (currentUrl.startsWith('https://pytorch.org/docs/stable')) {
// For the stable documentation, default to Google
defaultSearchType = localStorage.getItem('searchType') || 'google';
} else {
// For any other version, including docs-preview, default to Sphinx
defaultSearchType = 'sphinx';
}
// Set the default search type
document.querySelector(`input[name="searchType"][value="${defaultSearchType}"]`).checked = true;
toggleSearchBox(defaultSearchType);
// Event listener for changes in search type
searchForm.addEventListener('change', function(event) {
const selectedSearchType = event.target.value;
localStorage.setItem('searchType', selectedSearchType);
toggleSearchBox(selectedSearchType);
});
// Set placeholder text for Google search box
window.onload = function() {
var placeholderText = "Search Docs";
var googleSearchboxText = document.querySelector("#gsc-i-id1");
if (googleSearchboxText) {
googleSearchboxText.placeholder = placeholderText;
googleSearchboxText.style.fontFamily = 'FreightSans';
googleSearchboxText.style.fontSize = "1.2rem";
googleSearchboxText.style.color = '#262626';
}
};
});
</script>
<!--temporarily add a link to survey -->
<div class="survey-link">
<p><i class="fas fa-poll" aria-hidden="true">  </i><a href="https://forms.gle/tdrnwJhaQ9tUePxz9">Share Your Feedback</a> about our new search</p>
</div>
</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 class="current">
<li class="toctree-l1"><a class="reference internal" href="amp_examples.html">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="custom_operators.html">PyTorch Custom Operators Landing Page</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="extending.func.html">Extending torch.func with autograd.Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">FSDP Notes</a></li>
<li class="toctree-l1"><a class="reference internal" href="get_start_xpu.html">Getting Started on Intel GPU</a></li>
<li class="toctree-l1"><a class="reference internal" href="gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="mps.html">MPS backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="numerical_accuracy.html">Numerical accuracy</a></li>
<li class="toctree-l1"><a class="reference internal" href="randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="windows.html">Windows FAQ</a></li>
</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>
<li class="toctree-l1"><a class="reference internal" href="../torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="../amp.html">torch.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="../autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="../library.html">torch.library</a></li>
<li class="toctree-l1"><a class="reference internal" href="../accelerator.html">torch.accelerator</a></li>
<li class="toctree-l1"><a class="reference internal" href="../cpu.html">torch.cpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="../cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_cuda_memory.html">Understanding CUDA Memory Usage</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_cuda_memory.html#generating-a-snapshot">Generating a Snapshot</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_cuda_memory.html#using-the-visualizer">Using the visualizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_cuda_memory.html#snapshot-api-reference">Snapshot API Reference</a></li>
<li class="toctree-l1"><a class="reference internal" href="../mps.html">torch.mps</a></li>
<li class="toctree-l1"><a class="reference internal" href="../xpu.html">torch.xpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="../mtia.html">torch.mtia</a></li>
<li class="toctree-l1"><a class="reference internal" href="../mtia.memory.html">torch.mtia.memory</a></li>
<li class="toctree-l1"><a class="reference internal" href="../meta.html">Meta device</a></li>
<li class="toctree-l1"><a class="reference internal" href="../backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="../export.html">torch.export</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.tensor.html">torch.distributed.tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fsdp.html">torch.distributed.fsdp</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.fsdp.fully_shard.html">torch.distributed.fsdp.fully_shard</a></li>
<li class="toctree-l1"><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.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.pipelining.html">torch.distributed.pipelining</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.checkpoint.html">torch.distributed.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch.compiler.html">torch.compiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="../func.html">torch.func</a></li>
<li class="toctree-l1"><a class="reference internal" href="../futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fx.experimental.html">torch.fx.experimental</a></li>
<li class="toctree-l1"><a class="reference internal" href="../hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="../jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="../linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="../monitor.html">torch.monitor</a></li>
<li class="toctree-l1"><a class="reference internal" href="../signal.html">torch.signal</a></li>
<li class="toctree-l1"><a class="reference internal" href="../special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="../package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="../profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.attention.html">torch.nn.attention</a></li>
<li class="toctree-l1"><a class="reference internal" href="../onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="../optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="../complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="../ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="../quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="../rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="../random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="../masked.html">torch.masked</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nested.html">torch.nested</a></li>
<li class="toctree-l1"><a class="reference internal" href="../size.html">torch.Size</a></li>
<li class="toctree-l1"><a class="reference internal" href="../sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="../storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="../testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="../utils.html">torch.utils</a></li>
<li class="toctree-l1"><a class="reference internal" href="../benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="../bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="../checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="../cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="../data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="../deterministic.html">torch.utils.deterministic</a></li>
<li class="toctree-l1"><a class="reference internal" href="../jit_utils.html">torch.utils.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="../dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="../mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="../model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="../module_tracker.html">torch.utils.module_tracker</a></li>
<li class="toctree-l1"><a class="reference internal" href="../type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="../named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="../name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="../config_mod.html">torch.__config__</a></li>
<li class="toctree-l1"><a class="reference internal" href="../future_mod.html">torch.__future__</a></li>
<li class="toctree-l1"><a class="reference internal" href="../logging.html">torch._logging</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch_environment_variables.html">Torch Environment Variables</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/data">TorchData</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/torchrec">TorchRec</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</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>FSDP Notes</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/notes/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">
<!-- Google Tag Manager (noscript) -->
<noscript><iframe src="https://www.googletagmanager.com/ns.html?id=GTM-T8XT4PS"
height="0" width="0" style="display:none;visibility:hidden"></iframe></noscript>
<!-- End Google Tag Manager (noscript) -->
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<div class="section" id="fsdp-notes">
<span id="id1"></span><h1>FSDP Notes<a class="headerlink" href="#fsdp-notes" title="Permalink to this heading">¶</a></h1>
<div class="section" id="fsdp-prefetch-nuances">
<span id="fsdp-prefetch"></span><h2>FSDP Prefetch Nuances<a class="headerlink" href="#fsdp-prefetch-nuances" title="Permalink to this heading">¶</a></h2>
<p>For overlapping <code class="docutils literal notranslate"><span class="pre">forward</span></code> all-gathers with <code class="docutils literal notranslate"><span class="pre">forward</span></code> compute, there are two possible mechanisms:</p>
<ol class="arabic simple">
<li><p>Implicit forward prefetching (always enabled)</p></li>
<li><p>Explicit forward prefetching (<code class="docutils literal notranslate"><span class="pre">forward_prefetch=True</span></code>)</p></li>
</ol>
<p>Implicit <code class="docutils literal notranslate"><span class="pre">forward</span></code> prefetching refers to relying on issuing the all-gathers from a separate CUDA
stream to allow for overlapping an all-gather with <code class="docutils literal notranslate"><span class="pre">forward</span></code> compute issued before it (from the CPU
perspective). For example, if we have layer 0 all-gather -> layer 0 <code class="docutils literal notranslate"><span class="pre">forward</span></code> compute -> layer 1
all-gather -> …, then layer 1 all-gather can overlap with layer 0 <code class="docutils literal notranslate"><span class="pre">forward</span></code> compute even though the
CPU thread issued it afterwards. (The 1st all-gather will not be able to overlap with anything.)</p>
<p>Explicit <code class="docutils literal notranslate"><span class="pre">forward</span></code> prefetching refers to changing the CPU thread’s issue order: e.g. layer 0
all-gather -> layer 1 all-gather -> layer 0 <code class="docutils literal notranslate"><span class="pre">forward</span></code> compute -> …. In eager mode, there is no way to
know in general which layer is the next layer (e.g. layer 1 in the example) when still executing on
layer 0. Therefore, explicit <code class="docutils literal notranslate"><span class="pre">forward</span></code> prefetching should only be used for models whose execution
order is fixed from iteration to iteration (which we sometimes call “static graph”). An example of a
model that does not satisfy this constraint is <a class="reference external" href="https://pytorch.org/blog/scaling-multimodal-foundation-models-in-torchmultimodal-with-pytorch-distributed/">FLAVA</a>).</p>
<p>Explicit <code class="docutils literal notranslate"><span class="pre">forward</span></code> prefetching only saves the time taken to issue a layer’s <code class="docutils literal notranslate"><span class="pre">forward</span></code> compute kernels at
the cost that the next all-gather’s output tensor must be allocated while the current one is still
in use. By issuing the next all- gather before the current <code class="docutils literal notranslate"><span class="pre">forward</span></code> compute kernels, the next
all-gather can start sooner on GPU. For most LLM workloads, this is not the case, so there is no
motivation for enabling <code class="docutils literal notranslate"><span class="pre">forward_prefetch=True</span></code>.</p>
<p>In contrast, for <code class="docutils literal notranslate"><span class="pre">backward</span></code>, we must use explicit <code class="docutils literal notranslate"><span class="pre">backward</span></code> prefetching or else there will be 0 overlap
of communication and computation. The reason is because we use a single NCCL process group for both
all-gather and reduce-scatter (partially because in earlier NCCL versions, it was not safe to use
multiple concurrently on the same device over the same ranks). A single NCCL process group means a
single internal NCCL stream on which reduce-scatters and all-gathers run serially. As such, unless
we explicitly reorder the CPU issue order to be next all-gather -> current reduce-scatter, then the
current reduce-scatter would block the next all-gather and hence the next <code class="docutils literal notranslate"><span class="pre">backward</span></code> computation,
preventing the current reduce-scatter from overlapping.</p>
</div>
<div class="section" id="communication-payload-size">
<span id="fsdp-comms-payload-size"></span><h2>Communication payload size<a class="headerlink" href="#communication-payload-size" title="Permalink to this heading">¶</a></h2>
<p>In FSDP the communications are:</p>
<ol class="arabic simple">
<li><p>all-gather on parameters in <code class="docutils literal notranslate"><span class="pre">forward</span></code></p></li>
<li><p>all-gather on parameters in <code class="docutils literal notranslate"><span class="pre">backward</span></code></p></li>
<li><p>reduce-scatter on gradients in <code class="docutils literal notranslate"><span class="pre">backward</span></code></p></li>
</ol>
<p>If activation checkpointing (<a class="reference internal" href="../checkpoint.html#torch.utils.checkpoint.checkpoint" title="torch.utils.checkpoint.checkpoint"><code class="xref py py-func docutils literal notranslate"><span class="pre">checkpoint()</span></code></a>) is used there is no
additional communication since the parameters are prefetched anyway during <code class="docutils literal notranslate"><span class="pre">backward</span></code>.</p>
<p>In the FSDP design, the communication payload per rank is determined as follows: Each call to
<code class="xref py py-class docutils literal notranslate"><span class="pre">FullyShardedDataParallel</span></code> creates one communication group consisting of the parameters in
<code class="docutils literal notranslate"><span class="pre">module.parameters()</span></code> except any already assigned to a nested <code class="xref py py-class docutils literal notranslate"><span class="pre">FullyShardedDataParallel</span></code>
instance. For example, for Llama, if you apply <code class="xref py py-class docutils literal notranslate"><span class="pre">FullyShardedDataParallel</span></code> to every
transformer block and also to the root module, then there is one communication group for each
transformer block and finally one communication group with the initial embedding and final linear.
Each communication group corresponds to a single all-gather call and single reduce-scatter call. In
that way, how you apply <code class="xref py py-class docutils literal notranslate"><span class="pre">FullyShardedDataParallel</span></code> determines the communication size. In
general, applying FSDP to each transformer block is a good heuristic for LLMs, and it is hard to do
better than that given the current design.</p>
<p>Let’s consider an example where we have a Transformer-based model sharded over 8 GPUs, where the
sharding happens at the transformer block-level only, and each transformer block contains 1.6B
parameters and the parameters are in fp32 (4 bytes each). Which means that once sharded, each
transformer block will contain 0.2B parameters on each rank.</p>
<ul class="simple">
<li><p>The <code class="docutils literal notranslate"><span class="pre">forward</span></code> pass will communicate in chunks of <code class="docutils literal notranslate"><span class="pre">0.2*4</span> <span class="pre">=</span> <span class="pre">0.8GB</span></code> in all-gather</p></li>
<li><p>The <code class="docutils literal notranslate"><span class="pre">backward</span></code> pass will communicate 2 times <code class="docutils literal notranslate"><span class="pre">0.8GB</span></code> each (1x all-gather and 1x reduce-scatter)</p></li>
</ul>
<p>In other words there will be 3 communications with a payload of <code class="docutils literal notranslate"><span class="pre">0.8GB</span></code> each. If the model was
comprised of 10 transformer blocks there would be a total of 30 communications for a total of
<code class="docutils literal notranslate"><span class="pre">30*0.8=24GB</span></code>.</p>
<p>To formalize the payload size per communication per rank is
<code class="docutils literal notranslate"><span class="pre">total_transformer_block_params_in_B*dtype_bytes/num_gpus</span></code> (GBs).</p>
<p>Please note that in this example we didn’t include the additional communications required for the
embedding, which should be accounted for as well. And the math would depend on whether the input and
output embeddings are tied or not. If they aren’t tied there will be 2x more communications.</p>
</div>
<div class="section" id="fsdp-buffers-sizes">
<span id="id2"></span><h2>FSDP buffers sizes<a class="headerlink" href="#fsdp-buffers-sizes" title="Permalink to this heading">¶</a></h2>
<p>First, let’s cover the buffers allocated for communications:</p>
<p><code class="docutils literal notranslate"><span class="pre">forward</span></code> currently requires 2x all-gather buffer size. Here is why:</p>
<p>As explained in <a class="reference internal" href="#fsdp-prefetch"><span class="std std-ref">FSDP Prefetch Nuances</span></a> in the case of explicit <code class="docutils literal notranslate"><span class="pre">forward</span></code> prefetching
(<code class="docutils literal notranslate"><span class="pre">forward_prefetch=True`)</span> <span class="pre">case</span> <span class="pre">of</span> <span class="pre">layer</span> <span class="pre">0</span> <span class="pre">all-gather</span> <span class="pre">-></span> <span class="pre">layer</span> <span class="pre">0</span> <span class="pre">forward</span> <span class="pre">compute</span> <span class="pre">-></span> <span class="pre">layer</span> <span class="pre">1</span>
<span class="pre">all-gather</span> <span class="pre">there</span> <span class="pre">is</span> <span class="pre">a</span> <span class="pre">need</span> <span class="pre">for</span> <span class="pre">2</span> <span class="pre">all-gather-sized</span> <span class="pre">buffers,</span> <span class="pre">because</span> <span class="pre">one</span> <span class="pre">buffer</span> <span class="pre">is</span> <span class="pre">used</span> <span class="pre">in</span> <span class="pre">the</span> <span class="pre">current</span> <span class="pre">``forward</span></code> while the other is used to do the prefetching.</p>
<p>While the implicit <code class="docutils literal notranslate"><span class="pre">forward</span></code> prefetching (<code class="docutils literal notranslate"><span class="pre">forward_prefetch=False</span></code>, default) case of the same sequence in theory should need only 1 buffer, in reality it’s still 2x all-gather-sized buffers. The reason is that in the flat-parameter FSDP design, we do not copy-out of the all-gather buffer. The parameters used for compute are directly viewed into the all-gather buffer (in fact, the main benefit of the “flat parameter” is exactly this reason). In that case, while ‘layer 1 all-gather’ is overlapping with ‘layer 0 forward compute’, the ‘layer 0 forward compute’ is using the parameters viewed into the ‘layer 0 all-gather’ buffer.</p>
<p>A natural question then is, when would you want <code class="docutils literal notranslate"><span class="pre">forward_prefetch=False</span></code>? For static-graph models (like most LLMs), there is a major technical reason. It is more that, practically, we added this option quickly for some CPU-bound internal models and have not tested every code path with it in unit testing, so we are less confident in it. <code class="docutils literal notranslate"><span class="pre">forward_prefetching=False</span></code> can be slightly easier to reason about since we do not have to check the recorded forward order as a possible ‘failure mode’; a module’s all-gather can always be found under its own <code class="docutils literal notranslate"><span class="pre">record_function</span></code> label in its profiler trace.</p>
<p><code class="docutils literal notranslate"><span class="pre">backward</span></code> currently requires at least 2x all-gather buffer size and potentially a bit more. Here is why:</p>
<p>The current FSDP design uses <code class="docutils literal notranslate"><span class="pre">recordStream</span></code> to manage allocations produced in one stream consumed in another, which can lead to more memory usage than expected. How much more can be “non-deterministic” in that it depends on GPU kernel timing relative to the CPU. The <code class="docutils literal notranslate"><span class="pre">limit_all_gathers=True</span></code> argument is a mitigation to that - for more details refer to this discussion is <a class="reference external" href="https://dev-discuss.pytorch.org/t/fsdp-cudacachingallocator-an-outsider-newb-perspective/1486/1">FSDP & CUDACachingAllocator</a>.</p>
<p>The way existing FSDP works with autograd:</p>
<ul class="simple">
<li><p>Existing FSDP all-gathers the <code class="docutils literal notranslate"><span class="pre">flat_param</span></code>, which is the autograd leaf.</p></li>
<li><p>It calls <code class="docutils literal notranslate"><span class="pre">torch.split</span></code> to get 1D views into the <code class="docutils literal notranslate"><span class="pre">flat_param</span></code> corresponding to its constituent original parameters.</p></li>
<li><p>It calls <code class="docutils literal notranslate"><span class="pre">torch.view</span></code> on each 1D split to view back to ND.</p></li>
<li><p>This means that in <code class="docutils literal notranslate"><span class="pre">backward</span></code>, we end up with <code class="docutils literal notranslate"><span class="pre">ViewBackward</span></code> (ND -> 1D) and <code class="docutils literal notranslate"><span class="pre">SplitWithSizesBackward</span></code> (which is a concat). In particular, each individual gradient is computed as a separate allocation, and an explicit concat happens to construct the reduce-scatter input buffer. This implies actually a 2x buffer size for reduce-scatter at that peak memory point.</p></li>
</ul>
<p>In summary, for <code class="docutils literal notranslate"><span class="pre">backward</span></code>, it is about 2x buffer size for reduce-scatter plus any <code class="docutils literal notranslate"><span class="pre">recordStream</span></code> effects.</p>
<p>Second, let’s discuss the additional buffers:</p>
<p>Once the sharded parameters are gathered from all ranks, they require an additional buffer of <cite>total_transformer_block_params_in_B*dtype_bytes</cite> for the full parameters - so continuing the example from earlier if each transformer block is 1.6B parameters and the parameters are in fp32, then it’d be <cite>1.6*4=6.4GB</cite> buffer.</p>
<p>And there is a need for 2 of those buffers, since there is one currently being used and another being prefetched.</p>
<p>To summarize, we have:</p>
<ol class="arabic simple">
<li><p>2 times communication buffers of <code class="docutils literal notranslate"><span class="pre">total_transformer_block_params_in_B*dtype_bytes/num_gpus</span></code></p></li>
<li><p>2 times unsharded transformer block parameters buffer <code class="docutils literal notranslate"><span class="pre">``total_transformer_block_params_in_B*dtype_bytes</span></code></p></li>
</ol>
<p>or if you have been following the example:</p>
<ol class="arabic simple">
<li><p><code class="docutils literal notranslate"><span class="pre">2*1.6*4/8=1.6GB</span></code></p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">2**1.6*4=12.8GB</span></code></p></li>
</ol>
<p>and the total of <code class="docutils literal notranslate"><span class="pre">14.4GB</span></code>.</p>
<p>Now let’s briefly discuss what happens to the embeddings as we have left those out from the calculations:</p>
<p>Given the rule we discussed that you included in the note starting with “the communication buffer
size is determined as follows”, we can analyze as follows:</p>
<ul class="simple">
<li><p>Suppose we apply FSDP to the root module (e.g. the <code class="docutils literal notranslate"><span class="pre">Transformer</span></code> class). Suppose we further apply FSDP to each transformer block (e.g. the <code class="docutils literal notranslate"><span class="pre">TransformerBlock</span></code> class).</p></li>
<li><p>Most commonly, the embedding and final linear projection are direct children of the root <code class="docutils literal notranslate"><span class="pre">Transformer</span></code> class.</p></li>
<li><p>Following our rule, that means that the embedding and final linear projection are assigned to the root <code class="docutils literal notranslate"><span class="pre">Transformer</span></code>’s flat parameter.</p></li>
<li><p>We have _another_ special rule, which is that the root does not free its parameters after forward because they will be anyways immediately all-gathered in backward.</p></li>
<li><p>Putting this together, this means that the root’s flat parameter including the embedding and final projection are all-gathered to begin forward and kept in GPU memory until the end of backward.</p></li>
<li><p>If the embedding and final linear are not weight-tied, then we _could_ further apply FSDP to the embedding and to the final linear. For weight-tied parameters, we require them to be part of the same flat parameter (or else it would get double-counted). That would allow the embedding to be freed after its usage in forward and only all-gathered toward the end of backward.</p></li>
<li><p>Hopefully, this gives a better sense – each FSDP module gets assigned parameters in its <code class="docutils literal notranslate"><span class="pre">module.parameters</span></code> except those already assigned to another nested FSDP module, and the FSDP module’s <code class="docutils literal notranslate"><span class="pre">forward</span></code> defines the ‘live’ interval for its parameters. Hence, the nested <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code> structure can affect the all-gather/free schedule and hence the memory/throughput performance.</p></li>
</ul>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="get_start_xpu.html" class="btn btn-neutral float-right" title="Getting Started on Intel GPU" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="faq.html" class="btn btn-neutral" title="Frequently Asked Questions" accesskey="p" rel="prev"><img src="../_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright 2024, PyTorch Contributors.
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
<script>
var match = window.location.href.match(/\/_[a-zA-Z0-9_]*.html|_dynamo/gi);
var url = window.location.href.lastIndexOf(match[match.length-1]);
if (url)
{
var div = '<div class="admonition note"><p class="admonition-title">Note</p><p><i class="fa fa-exclamation-circle" aria-hidden="true"> </i> This page describes an internal API which is not intended to be used outside of the PyTorch codebase and can be modified or removed without notice.</p></div>'
document.getElementById("pytorch-article").insertAdjacentHTML('afterBegin', div)
}
</script>
</div>
<div class="pytorch-content-right" id="pytorch-content-right">
<div class="pytorch-right-menu" id="pytorch-right-menu">
<div class="pytorch-side-scroll" id="pytorch-side-scroll-right">
<ul>
<li><a class="reference internal" href="#">FSDP Notes</a><ul>
<li><a class="reference internal" href="#fsdp-prefetch-nuances">FSDP Prefetch Nuances</a></li>
<li><a class="reference internal" href="#communication-payload-size">Communication payload size</a></li>
<li><a class="reference internal" href="#fsdp-buffers-sizes">FSDP buffers sizes</a></li>
</ul>
</li>
</ul>
</div>
</div>
</div>
</section>
</div>
<script type="text/javascript" id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script data-url_root="../" id="documentation_options" src="../_static/documentation_options.js"></script>
<script src="../_static/jquery.js"></script>
<script src="../_static/underscore.js"></script>
<script src="../_static/_sphinx_javascript_frameworks_compat.js"></script>
<script src="../_static/doctools.js"></script>
<script src="../_static/sphinx_highlight.js"></script>
<script src="../_static/clipboard.min.js"></script>
<script src="../_static/copybutton.js"></script>
<script type="text/javascript" src="../_static/js/vendor/popper.min.js"></script>
<script type="text/javascript" src="../_static/js/vendor/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/list.js/1.5.0/list.min.js"></script>
<script type="text/javascript" src="../_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<script script type="text/javascript">
var collapsedSections = ['Developer Notes', 'Language Bindings', 'Libraries', 'Community'];
</script>
<img height="1" width="1" style="border-style:none;" alt="" src="https://www.googleadservices.com/pagead/conversion/795629140/?label=txkmCPmdtosBENSssfsC&guid=ON&script=0"/>
<!-- Begin Footer -->
<div class="container-fluid docs-tutorials-resources" id="docs-tutorials-resources">
<div class="container">
<div class="row">
<div class="col-md-4 text-center">
<h2>Docs</h2>
<p>Access comprehensive developer documentation for PyTorch</p>
<a class="with-right-arrow" href="https://pytorch.org/docs/stable/index.html">View Docs</a>
</div>
<div class="col-md-4 text-center">
<h2>Tutorials</h2>
<p>Get in-depth tutorials for beginners and advanced developers</p>
<a class="with-right-arrow" href="https://pytorch.org/tutorials">View Tutorials</a>
</div>
<div class="col-md-4 text-center">
<h2>Resources</h2>
<p>Find development resources and get your questions answered</p>
<a class="with-right-arrow" href="https://pytorch.org/resources">View Resources</a>
</div>
</div>
</div>
</div>
<footer class="site-footer">
<div class="container footer-container">
<div class="footer-logo-wrapper">
<a href="https://pytorch.org/" class="footer-logo"></a>
</div>
<div class="footer-links-wrapper">
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/">PyTorch</a></li>
<li><a href="https://pytorch.org/get-started">Get Started</a></li>
<li><a href="https://pytorch.org/features">Features</a></li>
<li><a href="https://pytorch.org/ecosystem">Ecosystem</a></li>
<li><a href="https://pytorch.org/blog/">Blog</a></li>
<li><a href="https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md">Contributing</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/resources">Resources</a></li>
<li><a href="https://pytorch.org/tutorials">Tutorials</a></li>
<li><a href="https://pytorch.org/docs/stable/index.html">Docs</a></li>
<li><a href="https://discuss.pytorch.org" target="_blank">Discuss</a></li>
<li><a href="https://github.com/pytorch/pytorch/issues" target="_blank">Github Issues</a></li>
<li><a href="https://pytorch.org/assets/brand-guidelines/PyTorch-Brand-Guidelines.pdf" target="_blank">Brand Guidelines</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title">Stay up to date</li>
<li><a href="https://www.facebook.com/pytorch" target="_blank">Facebook</a></li>
<li><a href="https://twitter.com/pytorch" target="_blank">Twitter</a></li>
<li><a href="https://www.youtube.com/pytorch" target="_blank">YouTube</a></li>
<li><a href="https://www.linkedin.com/company/pytorch" target="_blank">LinkedIn</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title">PyTorch Podcasts</li>
<li><a href="https://open.spotify.com/show/6UzHKeiy368jKfQMKKvJY5" target="_blank">Spotify</a></li>
<li><a href="https://podcasts.apple.com/us/podcast/pytorch-developer-podcast/id1566080008" target="_blank">Apple</a></li>
<li><a href="https://www.google.com/podcasts?feed=aHR0cHM6Ly9mZWVkcy5zaW1wbGVjYXN0LmNvbS9PQjVGa0lsOA%3D%3D" target="_blank">Google</a></li>
<li><a href="https://music.amazon.com/podcasts/7a4e6f0e-26c2-49e9-a478-41bd244197d0/PyTorch-Developer-Podcast?" target="_blank">Amazon</a></li>
</ul>
</div>
</div>
<div class="privacy-policy">
<ul>
<li class="privacy-policy-links"><a href="https://www.linuxfoundation.org/terms/" target="_blank">Terms</a></li>
<li class="privacy-policy-links">|</li>
<li class="privacy-policy-links"><a href="https://www.linuxfoundation.org/privacy-policy/" target="_blank">Privacy</a></li>
</ul>
</div>
<div class="copyright">
<p>© Copyright The Linux Foundation. The PyTorch Foundation is a project of The Linux Foundation.
For web site terms of use, trademark policy and other policies applicable to The PyTorch Foundation please see
<a href="https://www.linuxfoundation.org/policies/">www.linuxfoundation.org/policies/</a>. The PyTorch Foundation supports the PyTorch open source
project, which has been established as PyTorch Project a Series of LF Projects, LLC. For policies applicable to the PyTorch Project a Series of LF Projects, LLC,
please see <a href="https://www.lfprojects.org/policies/">www.lfprojects.org/policies/</a>.</p>
</div>
</div>
</footer>
<div class="cookie-banner-wrapper">
<div class="container">
<p class="gdpr-notice">To analyze traffic and optimize your experience, we serve cookies on this site. By clicking or navigating, you agree to allow our usage of cookies. As the current maintainers of this site, Facebook’s Cookies Policy applies. Learn more, including about available controls: <a href="https://www.facebook.com/policies/cookies/">Cookies Policy</a>.</p>
<img class="close-button" src="../_static/images/pytorch-x.svg">
</div>
</div>
<!-- End Footer -->
<!-- Begin Mobile Menu -->
<div class="mobile-main-menu">
<div class="container-fluid">
<div class="container">
<div class="mobile-main-menu-header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<a class="main-menu-close-button" href="#" data-behavior="close-mobile-menu"></a>
</div>
</div>
</div>
<div class="mobile-main-menu-links-container">
<div class="main-menu">
<ul>
<li class="resources-mobile-menu-title">
<a>Learn</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li>
<a href="https://pytorch.org/tutorials/beginner/basics/intro.html">Learn the Basics</a>
</li>
<li>
<a href="https://pytorch.org/tutorials/recipes/recipes_index.html">PyTorch Recipes</a>
</li>
<li>
<a href="https://pytorch.org/tutorials/beginner/introyt.html">Introduction to PyTorch - YouTube Series</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a>Ecosystem</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/ecosystem">Tools</a>
</li>
<li>
<a href="https://pytorch.org/#community-module">Community</a>
</li>
<li>
<a href="https://discuss.pytorch.org/">Forums</a>
</li>
<li>
<a href="https://pytorch.org/resources">Developer Resources</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem/contributor-awards-2023">Contributor Awards - 2023</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a>Edge</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/edge">About PyTorch Edge</a>
</li>
<li>
<a href="https://pytorch.org/executorch-overview">ExecuTorch</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a>Docs</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/docs/stable/index.html">PyTorch</a>
</li>
<li>
<a href="https://pytorch.org/pytorch-domains">PyTorch Domains</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a>Blog & News</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/blog/">PyTorch Blog</a>
</li>
<li>
<a href="https://pytorch.org/community-blog">Community Blog</a>
</li>