-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathdistributed.checkpoint.html
2721 lines (2464 loc) · 272 KB
/
distributed.checkpoint.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Distributed Checkpoint - torch.distributed.checkpoint — PyTorch 2.7 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/distributed.checkpoint.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="Probability distributions - torch.distributions" href="distributions.html" />
<link rel="prev" title="Pipeline Parallelism" href="distributed.pipelining.html" />
<!-- Google Tag Manager -->
<script>(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':
new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],
j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src=
'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);
})(window,document,'script','dataLayer','GTM-T8XT4PS');</script>
<!-- End Google Tag Manager -->
<script src="_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Learn
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/get-started">
<span class=dropdown-title>Get Started</span>
<p>Run PyTorch locally or get started quickly with one of the supported cloud platforms</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials">
<span class="dropdown-title">Tutorials</span>
<p>Whats new in PyTorch tutorials</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/basics/intro.html">
<span class="dropdown-title">Learn the Basics</span>
<p>Familiarize yourself with PyTorch concepts and modules</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/recipes/recipes_index.html">
<span class="dropdown-title">PyTorch Recipes</span>
<p>Bite-size, ready-to-deploy PyTorch code examples</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/introyt.html">
<span class="dropdown-title">Intro to PyTorch - YouTube Series</span>
<p>Master PyTorch basics with our engaging YouTube tutorial series</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Ecosystem
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem">
<span class="dropdown-title">Tools</span>
<p>Learn about the tools and frameworks in the PyTorch Ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class=dropdown-title>Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class=dropdown-title>Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class=dropdown-title>Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem/contributor-awards-2024">
<span class="dropdown-title">Contributor Awards - 2024</span>
<p>Award winners announced at this year's PyTorch Conference</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Edge
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/edge">
<span class="dropdown-title">About PyTorch Edge</span>
<p>Build innovative and privacy-aware AI experiences for edge devices</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch-overview">
<span class="dropdown-title">ExecuTorch</span>
<p>End-to-end solution for enabling on-device inference capabilities across mobile and edge devices</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch/stable/index.html">
<span class="dropdown-title">ExecuTorch Docs</span>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p>Explore the documentation for comprehensive guidance on how to use PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/pytorch-domains">
<span class="dropdown-title">PyTorch Domains</span>
<p>Read the PyTorch Domains documentation to learn more about domain-specific libraries</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Blogs & News
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/blog/">
<span class="dropdown-title">PyTorch Blog</span>
<p>Catch up on the latest technical news and happenings</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/community-blog">
<span class="dropdown-title">Community Blog</span>
<p>Stories from the PyTorch ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/videos">
<span class="dropdown-title">Videos</span>
<p>Learn about the latest PyTorch tutorials, new, and more </p>
<a class="nav-dropdown-item" href="https://pytorch.org/community-stories">
<span class="dropdown-title">Community Stories</span>
<p>Learn how our community solves real, everyday machine learning problems with PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/newsletter">
<span class="dropdown-title">Newsletter</span>
<p>Stay up-to-date with the latest updates</p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
About
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn more about the PyTorch Foundation</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/governing-board">
<span class="dropdown-title">Governing Board</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/credits">
<span class="dropdown-title">Cloud Credit Program</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tac">
<span class="dropdown-title">Technical Advisory Council</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/staff">
<span class="dropdown-title">Staff</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/contact-us">
<span class="dropdown-title">Contact Us</span>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div class="no-dropdown">
<a href="https://pytorch.org/join" data-cta="join">
Become a Member
</a>
</div>
</li>
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="github-icon">
</a>
</div>
</li>
<!--- TODO: This block adds the search icon to the nav bar. We will enable it later.
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="search-icon">
</a>
</div>
</li>
--->
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>2.7 ▼</a>
</div>
<div id="searchBox">
<div class="searchbox" id="googleSearchBox">
<script async src="https://cse.google.com/cse.js?cx=e65585f8c3ea1440e"></script>
<div class="gcse-search"></div>
</div>
<div id="sphinxSearchBox" style="display: none;">
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
</div>
<form id="searchForm">
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="google" checked>
Google Search
</label>
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="sphinx">
Classic Search
</label>
</form>
<script>
document.addEventListener('DOMContentLoaded', function() {
const searchForm = document.getElementById('searchForm');
const googleSearchBox = document.getElementById('googleSearchBox');
const sphinxSearchBox = document.getElementById('sphinxSearchBox');
// Function to toggle search box visibility
function toggleSearchBox(searchType) {
googleSearchBox.style.display = searchType === 'google' ? 'block' : 'none';
sphinxSearchBox.style.display = searchType === 'sphinx' ? 'block' : 'none';
}
// Determine the default search type
let defaultSearchType;
const currentUrl = window.location.href;
if (currentUrl.startsWith('https://pytorch.org/docs/stable')) {
// For the stable documentation, default to Google
defaultSearchType = localStorage.getItem('searchType') || 'google';
} else {
// For any other version, including docs-preview, default to Sphinx
defaultSearchType = 'sphinx';
}
// Set the default search type
document.querySelector(`input[name="searchType"][value="${defaultSearchType}"]`).checked = true;
toggleSearchBox(defaultSearchType);
// Event listener for changes in search type
searchForm.addEventListener('change', function(event) {
const selectedSearchType = event.target.value;
localStorage.setItem('searchType', selectedSearchType);
toggleSearchBox(selectedSearchType);
});
// Set placeholder text for Google search box
window.onload = function() {
var placeholderText = "Search Docs";
var googleSearchboxText = document.querySelector("#gsc-i-id1");
if (googleSearchboxText) {
googleSearchboxText.placeholder = placeholderText;
googleSearchboxText.style.fontFamily = 'FreightSans';
googleSearchboxText.style.fontSize = "1.2rem";
googleSearchboxText.style.color = '#262626';
}
};
});
</script>
</div>
<p class="caption" role="heading"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="community/build_ci_governance.html">PyTorch Governance | Build + CI</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/design.html">PyTorch Design Philosophy</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/governance.html">PyTorch Governance | Mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/persons_of_interest.html">PyTorch Governance | Maintainers</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Developer Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="notes/amp_examples.html">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/custom_operators.html">PyTorch Custom Operators Landing Page</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.func.html">Extending torch.func with autograd.Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/fsdp.html">FSDP Notes</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/get_start_xpu.html">Getting Started on Intel GPU</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/libtorch_stable_abi.html">LibTorch Stable ABI</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/mps.html">MPS backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/numerical_accuracy.html">Numerical accuracy</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
<li class="toctree-l1"><a class="reference internal" href="deploy.html">torch::deploy</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python API</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="amp.html">torch.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="library.html">torch.library</a></li>
<li class="toctree-l1"><a class="reference internal" href="accelerator.html">torch.accelerator</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpu.html">torch.cpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html">Understanding CUDA Memory Usage</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#generating-a-snapshot">Generating a Snapshot</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#using-the-visualizer">Using the visualizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#snapshot-api-reference">Snapshot API Reference</a></li>
<li class="toctree-l1"><a class="reference internal" href="mps.html">torch.mps</a></li>
<li class="toctree-l1"><a class="reference internal" href="xpu.html">torch.xpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="mtia.html">torch.mtia</a></li>
<li class="toctree-l1"><a class="reference internal" href="mtia.memory.html">torch.mtia.memory</a></li>
<li class="toctree-l1"><a class="reference internal" href="meta.html">Meta device</a></li>
<li class="toctree-l1"><a class="reference internal" href="backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="export.html">torch.export</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.tensor.html">torch.distributed.tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="fsdp.html">torch.distributed.fsdp</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.fsdp.fully_shard.html">torch.distributed.fsdp.fully_shard</a></li>
<li class="toctree-l1"><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 current"><a class="current reference internal" href="#">torch.distributed.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.compiler.html">torch.compiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="func.html">torch.func</a></li>
<li class="toctree-l1"><a class="reference internal" href="futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.experimental.html">torch.fx.experimental</a></li>
<li class="toctree-l1"><a class="reference internal" href="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="monitor.html">torch.monitor</a></li>
<li class="toctree-l1"><a class="reference internal" href="signal.html">torch.signal</a></li>
<li class="toctree-l1"><a class="reference internal" href="special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.attention.html">torch.nn.attention</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="masked.html">torch.masked</a></li>
<li class="toctree-l1"><a class="reference internal" href="nested.html">torch.nested</a></li>
<li class="toctree-l1"><a class="reference internal" href="size.html">torch.Size</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="utils.html">torch.utils</a></li>
<li class="toctree-l1"><a class="reference internal" href="benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="deterministic.html">torch.utils.deterministic</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit_utils.html">torch.utils.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="module_tracker.html">torch.utils.module_tracker</a></li>
<li class="toctree-l1"><a class="reference internal" href="type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="config_mod.html">torch.__config__</a></li>
<li class="toctree-l1"><a class="reference internal" href="future_mod.html">torch.__future__</a></li>
<li class="toctree-l1"><a class="reference internal" href="logging.html">torch._logging</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_environment_variables.html">Torch Environment Variables</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/data">TorchData</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/torchrec">TorchRec</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/ao">torchao</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="index.html">
Docs
</a> >
</li>
<li>Distributed Checkpoint - torch.distributed.checkpoint</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/distributed.checkpoint.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="distributed-checkpoint-torch-distributed-checkpoint">
<h1>Distributed Checkpoint - torch.distributed.checkpoint<a class="headerlink" href="#distributed-checkpoint-torch-distributed-checkpoint" title="Permalink to this heading">¶</a></h1>
<p>Distributed Checkpoint (DCP) support loading and saving models from multiple ranks in parallel.
It handles load-time resharding which enables saving in one cluster topology and loading into another.</p>
<p>DCP is different than <cite>torch.save</cite> and <cite>torch.load</cite> in a few significant ways:</p>
<ul class="simple">
<li><p>It produces multiple files per checkpoint, with at least one per rank.</p></li>
<li><p>It operates in place, meaning that the model should allocate its data first and DCP uses that storage instead.</p></li>
</ul>
<p>The entrypoints to load and save a checkpoint are the following:</p>
<div class="section" id="additional-resources">
<h2>Additional resources:<a class="headerlink" href="#additional-resources" title="Permalink to this heading">¶</a></h2>
<ul class="simple">
<li><p><a class="reference external" href="https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html">Getting Started with Distributed Checkpoint (DCP)</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/tutorials/recipes/distributed_async_checkpoint_recipe.html">Asynchronous Saving with Distributed Checkpoint (DCP)</a></p></li>
<li><p><a class="reference external" href="https://github.com/pytorch/torchtitan/blob/main/docs/checkpoint.md">TorchTitan Checkpointing Docs</a></p></li>
<li><p><a class="reference external" href="https://github.com/pytorch/torchtitan/blob/main/torchtitan/components/checkpoint.py">TorchTitan DCP Implementation</a></p></li>
</ul>
<span class="target" id="module-torch.distributed.checkpoint"></span><dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.state_dict_saver.AsyncCheckpointerType">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.checkpoint.state_dict_saver.</span></span><span class="sig-name descname"><span class="pre">AsyncCheckpointerType</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">value</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/checkpoint/state_dict_saver.html#AsyncCheckpointerType"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/state_dict_saver.py#L39"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.state_dict_saver.AsyncCheckpointerType" title="Permalink to this definition">¶</a></dt>
<dd><p>Enum for async checkpointer type.</p>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.state_dict_saver.save">
<span class="sig-prename descclassname"><span class="pre">torch.distributed.checkpoint.state_dict_saver.</span></span><span class="sig-name descname"><span class="pre">save</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">checkpoint_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">storage_writer</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">planner</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">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">no_dist</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/checkpoint/state_dict_saver.html#save"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/state_dict_saver.py#L74"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.state_dict_saver.save" title="Permalink to this definition">¶</a></dt>
<dd><p>Save a distributed model in SPMD style.</p>
<p>This function is different from <code class="docutils literal notranslate"><span class="pre">torch.save()</span></code> as it handles
<code class="docutils literal notranslate"><span class="pre">ShardedTensor</span></code> , and <code class="docutils literal notranslate"><span class="pre">DTensor</span></code> by having each rank only save their local shards.</p>
<p>For each <code class="docutils literal notranslate"><span class="pre">Stateful</span></code> object (having both a <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> and a <code class="docutils literal notranslate"><span class="pre">load_state_dict</span></code>),
save will call <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> before serialization.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>There is no guarantees of Backwards Compatibility across PyTorch versions
for saved state_dicts.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>If using the <cite>process_group</cite> argument, make sure that only its ranks
call <cite>save_state_dict</cite> and that all data in state_dict belong to it.</p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>When saving checkpoint for FSDP’s <cite>ShardingStrategy.HYBRID_SHARD</cite>, only one of
the shard_group should be calling <cite>save_state_dict</cite> and the corresponding process
group needs to be passed in.</p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<dl class="simple">
<dt>If no process group is available, this function assumes the intention is to save the</dt><dd><p>state_dict in the local process.</p>
</dd>
</dl>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>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 v3.13)"><em>str</em></a><em>, </em><em>Any</em><em>]</em>) – The state_dict to save.</p></li>
<li><p><strong>checkpoint_id</strong> (<em>Union</em><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.13)"><em>str</em></a><em>, </em><a class="reference external" href="https://docs.python.org/3/library/os.html#os.PathLike" title="(in Python v3.13)"><em>os.PathLike</em></a><em>, </em><em>None</em><em>]</em>) – The ID of this checkpoint instance. The meaning of the checkpoint_id
depends on the storage. It can be a path to a folder or to a file.
It can also be a key if the storage is a key-value store.
(Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>storage_writer</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.checkpoint.StorageWriter" title="torch.distributed.checkpoint.StorageWriter"><em>StorageWriter</em></a><em>]</em>) – Instance of StorageWriter used to perform writes. If this is not
specified, DCP will automatically infer the writer based on the
checkpoint_id. If checkpoint_id is also None, an exception will
be raised. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>planner</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.checkpoint.SavePlanner" title="torch.distributed.checkpoint.SavePlanner"><em>SavePlanner</em></a><em>]</em>) – Instance of SavePlanner. If this is not specificed, the default
planner will be used. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>process_group</strong> (<em>Optional</em><em>[</em><em>ProcessGroup</em><em>]</em>) – ProcessGroup to be used for cross-rank synchronization.
(Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>no_dist</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><em>bool</em></a>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, this function will assume the intent is to load
a checkpoint without using cross-rank synchronization.
(Default: <code class="docutils literal notranslate"><span class="pre">False</span></code>)</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>Metadata object for the saved checkpoint.</p>
</dd>
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p>Metadata</p>
</dd>
</dl>
<p class="rubric">Example</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">my_model</span> <span class="o">=</span> <span class="n">MyModule</span><span class="p">()</span>
</pre></div>
</div>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">state_dict</span> <span class="o">=</span> <span class="p">{</span><span class="s2">"model"</span><span class="p">:</span> <span class="n">my_model</span><span class="p">}</span>
</pre></div>
</div>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">fs_storage_writer</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">distributed</span><span class="o">.</span><span class="n">checkpoint</span><span class="o">.</span><span class="n">FileSystemWriter</span><span class="p">(</span>
<span class="gp">... </span> <span class="s2">"/checkpoint/1"</span>
<span class="gp">... </span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">distributed</span><span class="o">.</span><span class="n">checkpoint</span><span class="o">.</span><span class="n">save</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">state_dict</span><span class="o">=</span><span class="n">state_dict</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="n">storage_writer</span><span class="o">=</span><span class="n">fs_storage_writer</span><span class="p">,</span>
<span class="gp">>>> </span><span class="p">)</span>
</pre></div>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>save_state_dict uses collectives to coordinate writes across ranks.
For NCCL-based process groups, internal tensor representations of
objects must be moved to the GPU device before communication takes place.
In this case, the device used is given by <code class="docutils literal notranslate"><span class="pre">torch.cuda.current_device()</span></code>
and it is the user’s responsibility to ensure that this is set so that
each rank has an individual GPU, via <code class="docutils literal notranslate"><span class="pre">torch.cuda.set_device()</span></code>.</p>
</div>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.state_dict_saver.async_save">
<span class="sig-prename descclassname"><span class="pre">torch.distributed.checkpoint.state_dict_saver.</span></span><span class="sig-name descname"><span class="pre">async_save</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">checkpoint_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">storage_writer</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">planner</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">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">async_checkpointer_type</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">AsyncCheckpointerType.THREAD</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/checkpoint/state_dict_saver.html#async_save"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/state_dict_saver.py#L185"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.state_dict_saver.async_save" title="Permalink to this definition">¶</a></dt>
<dd><p>Asynchronous version of <code class="docutils literal notranslate"><span class="pre">save</span></code>. This code first de-stages the state_dict on to the
staging storage (defaults to CPU memory), and then calls the <cite>save</cite> in a separate thread.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>This feature is experimental and subject to change.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>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 v3.13)"><em>str</em></a><em>, </em><em>Any</em><em>]</em>) – The state_dict to save.</p></li>
<li><p><strong>checkpoint_id</strong> (<em>Union</em><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.13)"><em>str</em></a><em>, </em><a class="reference external" href="https://docs.python.org/3/library/os.html#os.PathLike" title="(in Python v3.13)"><em>os.PathLike</em></a><em>, </em><em>None</em><em>]</em>) – The ID of this checkpoint instance. The meaning of the checkpoint_id
depends on the storage. It can be a path to a folder or to a file.
It can also be a key if the storage is a key-value store.
(Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>storage_writer</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.checkpoint.StorageWriter" title="torch.distributed.checkpoint.StorageWriter"><em>StorageWriter</em></a><em>]</em>) – Instance of StorageWriter used to perform ‘stage’ and ‘save’. If
this is not specified, DCP will automatically infer the writer based on the
checkpoint_id. If checkpoint_id is also None, an exception will
be raised. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>planner</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.checkpoint.SavePlanner" title="torch.distributed.checkpoint.SavePlanner"><em>SavePlanner</em></a><em>]</em>) – Instance of SavePlanner. If this is not specificed, the default
planner will be used. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>process_group</strong> (<em>Optional</em><em>[</em><em>ProcessGroup</em><em>]</em>) – ProcessGroup to be used for cross-rank synchronization.
(Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>A future holding the resultant Metadata object from <cite>save</cite>.</p>
</dd>
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p><a class="reference internal" href="futures.html#torch.futures.Future" title="torch.futures.Future">Future</a></p>
</dd>
</dl>
<p class="rubric">Example</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">my_model</span> <span class="o">=</span> <span class="n">MyModule</span><span class="p">()</span>
</pre></div>
</div>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">state_dict</span> <span class="o">=</span> <span class="p">{</span><span class="s2">"model"</span><span class="p">:</span> <span class="n">my_model</span><span class="p">}</span>
</pre></div>
</div>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">fs_storage_writer</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">distributed</span><span class="o">.</span><span class="n">checkpoint</span><span class="o">.</span><span class="n">FileSystemWriter</span><span class="p">(</span>
<span class="gp">... </span> <span class="s2">"/checkpoint/1"</span>
<span class="gp">... </span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">checkpoint_future</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">distributed</span><span class="o">.</span><span class="n">checkpoint</span><span class="o">.</span><span class="n">async_save</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">state_dict</span><span class="o">=</span><span class="n">state_dict</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="n">storage_writer</span><span class="o">=</span><span class="n">fs_storage_writer</span><span class="p">,</span>
<span class="gp">>>> </span><span class="p">)</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># ... do some work ...</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="n">checkpoint_future</span><span class="o">.</span><span class="n">result</span><span class="p">()</span>
</pre></div>
</div>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.state_dict_saver.save_state_dict">
<span class="sig-prename descclassname"><span class="pre">torch.distributed.checkpoint.state_dict_saver.</span></span><span class="sig-name descname"><span class="pre">save_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">storage_writer</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">coordinator_rank</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">no_dist</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">planner</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/checkpoint/state_dict_saver.html#save_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/state_dict_saver.py#L46"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.state_dict_saver.save_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>This method is deprecated. Please switch to ‘save’.</p>
<dl class="field-list simple">
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p><em>Metadata</em></p>
</dd>
</dl>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.state_dict_loader.load">
<span class="sig-prename descclassname"><span class="pre">torch.distributed.checkpoint.state_dict_loader.</span></span><span class="sig-name descname"><span class="pre">load</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">checkpoint_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">storage_reader</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">planner</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">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">no_dist</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/checkpoint/state_dict_loader.html#load"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/state_dict_loader.py#L51"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.state_dict_loader.load" title="Permalink to this definition">¶</a></dt>
<dd><p>Load a checkpoint into a distributed state dict in SPMD style.</p>
<p>Each rank must have the same keys in their <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> provided to this
API. Mismatched keys may result in hangs or errors. If unsure, you can use
the <code class="docutils literal notranslate"><span class="pre">utils._assert_same_keys</span></code> API to check (but may incur communication
costs).</p>
<p>Each rank will try to read the least amount of data necessary
to fulfill the requested <cite>state_dict</cite>. When loading <code class="xref py py-class docutils literal notranslate"><span class="pre">ShardedTensor</span></code>
or <code class="xref py py-class docutils literal notranslate"><span class="pre">DTensor</span></code> instances, each rank only reads data for their local shards.</p>
<p>For each <code class="docutils literal notranslate"><span class="pre">Stateful</span></code> object (having both a <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> and a <code class="docutils literal notranslate"><span class="pre">load_state_dict</span></code>),
load will first call <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> before attempting deserialization, followed by
<code class="docutils literal notranslate"><span class="pre">load_state_dict</span></code> once the deserialization is complete.
For each non-<code class="docutils literal notranslate"><span class="pre">Stateful</span></code> object, load will deserailize the object, and then replace
it in the <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> with the deserialized object.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>All tensors in <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> must be allocated on their
destination device <em>prior to</em> calling this function.</p>
<p>All non-tensor data is loaded using <cite>torch.load()</cite> and modified in place
on state_dict.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Users must call <cite>load_state_dict</cite> on the root module to ensure load
pos-processing and non-tensor data properly propagates.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>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 v3.13)"><em>str</em></a><em>, </em><em>Any</em><em>]</em>) – The state_dict to load the checkpoint into.</p></li>
<li><p><strong>checkpoint_id</strong> (<em>Union</em><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.13)"><em>str</em></a><em>, </em><a class="reference external" href="https://docs.python.org/3/library/os.html#os.PathLike" title="(in Python v3.13)"><em>os.PathLike</em></a><em>, </em><em>None</em><em>]</em>) – The ID of this checkpoint instance. The meaning of the checkpoint_id
depends on the storage. It can be a path to a folder or to a file.
It can also be a key if the storage is a key-value store.
(Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>storage_reader</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.checkpoint.StorageReader" title="torch.distributed.checkpoint.StorageReader"><em>StorageReader</em></a><em>]</em>) – Instance of StorageWriter used to perform reads. If this is not
specified, DCP will automatically infer the reader based on the
checkpoint_id. If checkpoint_id is also None, an exception will
be raised. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>planner</strong> (<em>Optional</em><em>[</em><a class="reference internal" href="#torch.distributed.checkpoint.LoadPlanner" title="torch.distributed.checkpoint.LoadPlanner"><em>LoadPlanner</em></a><em>]</em>) – Instance of LoadPlanner. If this is not specificed, the default
planner will be used. (Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>process_group</strong> (<em>Optional</em><em>[</em><em>ProcessGroup</em><em>]</em>) – ProcessGroup to be used for cross-rank synchronization.
(Default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>no_dist</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><em>bool</em></a>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, this function will assume the intent is to load
a checkpoint without using cross-rank synchronization. (Default: <code class="docutils literal notranslate"><span class="pre">False</span></code>)</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>None.</p>
</dd>
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p>None</p>
</dd>
</dl>
<dl>
<dt>Examples</dt><dd><div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">my_model</span> <span class="o">=</span> <span class="n">MyModule</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">optimizer</span> <span class="o">=</span> <span class="n">Adagrad</span><span class="p">(</span><span class="n">my_model</span><span class="o">.</span><span class="n">parameters</span><span class="p">())</span>
<span class="gp">>>> </span><span class="n">model_state_dict</span> <span class="o">=</span> <span class="n">my_model</span><span class="o">.</span><span class="n">state_dict</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">fs_storage_reader</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">distributed</span><span class="o">.</span><span class="n">checkpoint</span><span class="o">.</span><span class="n">FileSystemReader</span><span class="p">(</span>
<span class="gp">... </span> <span class="s2">"/checkpoint/1"</span>
<span class="gp">... </span><span class="p">)</span>
</pre></div>
</div>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">distributed</span><span class="o">.</span><span class="n">checkpoint</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span>
<span class="gp">>>> </span> <span class="n">state_dict</span><span class="o">=</span><span class="n">model_state_dict</span><span class="p">,</span>
<span class="gp">>>> </span> <span class="n">storage_reader</span><span class="o">=</span><span class="n">fs_storage_reader</span><span class="p">,</span>
<span class="gp">>>> </span><span class="p">)</span>
</pre></div>
</div>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="c1"># module.load_state_dict() function might have customized steps</span>
<span class="gp">>>> </span><span class="c1"># to flush the state_dict, must call it to</span>
<span class="gp">>>> </span><span class="c1"># ensure correct behavior.</span>
<span class="gp">>>> </span><span class="n">my_model</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">model_state_dict</span><span class="p">)</span>
</pre></div>
</div>
</dd>
</dl>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>load_state_dict uses collectives to coordinate reads across ranks.
For NCCL-based process groups, internal tensor representations of
objects must be moved to the GPU device before communication takes place.
In this case, the device used is given by <code class="docutils literal notranslate"><span class="pre">torch.cuda.current_device()</span></code>
and it is the user’s responsibility to ensure that this is set so that each
rank has an individual GPU, via <code class="docutils literal notranslate"><span class="pre">torch.cuda.set_device()</span></code>.</p>
</div>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.state_dict_loader.load_state_dict">
<span class="sig-prename descclassname"><span class="pre">torch.distributed.checkpoint.state_dict_loader.</span></span><span class="sig-name descname"><span class="pre">load_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">storage_reader</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">coordinator_rank</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">no_dist</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">planner</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/checkpoint/state_dict_loader.html#load_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/state_dict_loader.py#L24"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.state_dict_loader.load_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>This method is deprecated. Please switch to ‘load’.</p>
<dl class="field-list simple">
</dl>
</dd></dl>
<p>The following module is also useful for additional customization of the staging mechanisms used for asynchronous checkpointing (<cite>torch.distributed.checkpoint.async_save</cite>):</p>
<span class="target" id="module-torch.distributed.checkpoint.staging"></span><dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.staging.AsyncStager">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.checkpoint.staging.</span></span><span class="sig-name descname"><span class="pre">AsyncStager</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/checkpoint/staging.html#AsyncStager"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/staging.py#L11"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.staging.AsyncStager" title="Permalink to this definition">¶</a></dt>
<dd><p>This protocol is meant to provide customization and extensibility for dcp.async_save, allowing users
to customize how data is staged previous to executing the usual dcp.save path in parallel.
The expected order of operations (concretely defined in <cite>torch.distributed.state_dict_saver.async_save</cite>)
is the following:</p>
<ol class="arabic simple">
<li><dl class="simple">
<dt>AsyncStager.stage_data(state_dict):</dt><dd><p>This call gives the AsyncStager the opportunity to ‘stage’
the state_dict. The expectation and purpose of staging in this context is to create a “training-safe”
representation of the state dict, meaning that any updates to module data after staging is complete
should not be reflected in the state dict returned from this method. For example, in the default
case a copy of the entire state dict is created on CPU RAM and returned here, allowing users
to continue training without risking changes to data which is being serialized.</p>
</dd>
</dl>
</li>
<li><dl class="simple">
<dt>dcp.save is called on the state_dict returned from stage in parallel. This call is responsible</dt><dd><p>for serializing the state_dict and writing it to storage.</p>
</dd>
</dl>
</li>
<li><dl class="simple">
<dt>If AsyncStager.should_synchronize_after_execute is True, this method will be called immediately after</dt><dd><p>the serialization thread starts and before returning from dcp.async_save. If this is set to False,
the assumption is the user has defined a custom synchronization point for the the purpose of further
optimizing save latency in the training loop (for example, by overlapping staging with the
forward/backward pass), and it is the respondsibility of the user to call <cite>AsyncStager.synchronize_staging</cite>
at the appropriate time.</p>
</dd>
</dl>
</li>
</ol>
<dl class="py property">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.staging.AsyncStager.should_synchronize_after_execute">
<em class="property"><span class="pre">property</span><span class="w"> </span></em><span class="sig-name descname"><span class="pre">should_synchronize_after_execute</span></span><em class="property"><span class="p"><span class="pre">:</span></span><span class="w"> </span><a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><span class="pre">bool</span></a></em><a class="headerlink" href="#torch.distributed.checkpoint.staging.AsyncStager.should_synchronize_after_execute" title="Permalink to this definition">¶</a></dt>
<dd><p>Whether to synchronize after executing the stage.</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.staging.AsyncStager.stage">
<span class="sig-name descname"><span class="pre">stage</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/checkpoint/staging.html#AsyncStager.stage"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/staging.py#L50"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.staging.AsyncStager.stage" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns a “staged” copy of <cite>state_dict</cite>. The expectation of the staged copy is that it is
innoculated from any updates incurred after the stage call is complete.</p>
<dl class="field-list simple">
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.13)">dict</a>[<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.13)">str</a>, <a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Union" title="(in Python v3.13)"><em>Union</em></a>[~StatefulT, <a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Any" title="(in Python v3.13)"><em>Any</em></a>]]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.staging.AsyncStager.synchronize_staging">
<span class="sig-name descname"><span class="pre">synchronize_staging</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/checkpoint/staging.html#AsyncStager.synchronize_staging"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/staging.py#L59"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.staging.AsyncStager.synchronize_staging" title="Permalink to this definition">¶</a></dt>
<dd><p>In the case <cite>stage</cite> is async in some way, this method should be called to ensure staging
is complete and it is safe to begin modifying the original <cite>state_dict</cite></p>
<dl class="field-list simple">
</dl>
</dd></dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.staging.BlockingAsyncStager">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.checkpoint.staging.</span></span><span class="sig-name descname"><span class="pre">BlockingAsyncStager</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">cache_staged_state_dict</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">type_check</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/checkpoint/staging.html#BlockingAsyncStager"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/staging.py#L66"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.staging.BlockingAsyncStager" title="Permalink to this definition">¶</a></dt>
<dd><p>An implementation of AsyncStager which stages the state_dict on CPU RAM and blocks until the copy is complete.
This implementation also provides an option to optimize stage latency using pinned memory.</p>
<p>N.B. synchronize_staging is a no-op in this case.</p>
<dl class="field-list simple">
</dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.staging.BlockingAsyncStager.stage">
<span class="sig-name descname"><span class="pre">stage</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/checkpoint/staging.html#BlockingAsyncStager.stage"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/staging.py#L98"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.staging.BlockingAsyncStager.stage" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns a copy of <cite>state_dict</cite> on the CPU.</p>
<dl class="field-list simple">
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.13)">dict</a>[<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.13)">str</a>, <a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Union" title="(in Python v3.13)"><em>Union</em></a>[~StatefulT, <a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Any" title="(in Python v3.13)"><em>Any</em></a>]]</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.staging.BlockingAsyncStager.synchronize_staging">
<span class="sig-name descname"><span class="pre">synchronize_staging</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/checkpoint/staging.html#BlockingAsyncStager.synchronize_staging"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/staging.py#L112"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.staging.BlockingAsyncStager.synchronize_staging" title="Permalink to this definition">¶</a></dt>
<dd><p>No-op function, since staging is blocking.</p>
<dl class="field-list simple">
</dl>
</dd></dl>
</dd></dl>
<p>In addition to the above entrypoints, <cite>Stateful</cite> objects, as described below, provide additional customization during saving/loading
.. automodule:: torch.distributed.checkpoint.stateful</p>
<dl class="py class">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.stateful.Stateful">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.distributed.checkpoint.stateful.</span></span><span class="sig-name descname"><span class="pre">Stateful</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/checkpoint/stateful.html#Stateful"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/stateful.py#L8"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.stateful.Stateful" title="Permalink to this definition">¶</a></dt>
<dd><p>Stateful protocol for objects that can be checkpointed and restored.</p>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.stateful.Stateful.load_state_dict">
<span class="sig-name descname"><span class="pre">load_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/checkpoint/stateful.html#Stateful.load_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/stateful.py#L31"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.stateful.Stateful.load_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Restore the object’s state from the provided state_dict.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>state_dict</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.13)"><em>dict</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.13)"><em>str</em></a><em>, </em><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Any" title="(in Python v3.13)"><em>Any</em></a><em>]</em>) – The state dict to restore from</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.distributed.checkpoint.stateful.Stateful.state_dict">
<span class="sig-name descname"><span class="pre">state_dict</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/distributed/checkpoint/stateful.html#Stateful.state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/distributed/checkpoint/stateful.py#L14"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.distributed.checkpoint.stateful.Stateful.state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Objects should return their state_dict representation as a dictionary.
The output of this function will be checkpointed, and later restored in
<cite>load_state_dict()</cite>.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Because of the inplace nature of restoring a checkpoint, this function
is also called during <cite>torch.distributed.checkpoint.load</cite>.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Returns</dt>
<dd class="field-odd"><p>The objects state dict</p>
</dd>
<dt class="field-even">Return type</dt>
<dd class="field-even"><p>Dict</p>
</dd>
</dl>
</dd></dl>
</dd></dl>
<p>This <a class="reference external" href="https://github.com/pytorch/pytorch/blob/main/torch/distributed/checkpoint/examples/fsdp_checkpoint_example.py">example</a> shows how to use Pytorch Distributed Checkpoint to save a FSDP model.</p>
<p>The following types define the IO interface used during checkpoint:</p>