-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathserialization.html
1723 lines (1523 loc) · 149 KB
/
serialization.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>Serialization semantics — PyTorch 2.7 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/notes/serialization.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="Windows FAQ" href="windows.html" />
<link rel="prev" title="Reproducibility" href="randomness.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 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"><a class="reference internal" href="fsdp.html">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="libtorch_stable_abi.html">LibTorch Stable ABI</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 current"><a class="current reference internal" href="#">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>
<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>Serialization semantics</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/notes/serialization.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="serialization-semantics">
<h1><a class="toc-backref" href="#id2">Serialization semantics</a><a class="headerlink" href="#serialization-semantics" title="Permalink to this heading">¶</a></h1>
<p>This note describes how you can save and load PyTorch tensors and module states
in Python, and how to serialize Python modules so they can be loaded in C++.</p>
<div class="contents topic" id="table-of-contents">
<p class="topic-title">Table of Contents</p>
<ul class="simple">
<li><p><a class="reference internal" href="#serialization-semantics" id="id2">Serialization semantics</a></p>
<ul>
<li><p><a class="reference internal" href="#saving-and-loading-tensors" id="id3">Saving and loading tensors</a></p></li>
<li><p><a class="reference internal" href="#saving-and-loading-tensors-preserves-views" id="id4">Saving and loading tensors preserves views</a></p></li>
<li><p><a class="reference internal" href="#saving-and-loading-torch-nn-modules" id="id5">Saving and loading torch.nn.Modules</a></p></li>
<li><p><a class="reference internal" href="#serialized-file-format-for-torch-save" id="id6">Serialized file format for <code class="docutils literal notranslate"><span class="pre">torch.save</span></code></a></p></li>
<li><p><a class="reference internal" href="#torch-load-with-weights-only-true" id="id7"><code class="docutils literal notranslate"><span class="pre">torch.load</span></code> with <code class="docutils literal notranslate"><span class="pre">weights_only=True</span></code></a></p>
<ul>
<li><p><a class="reference internal" href="#troubleshooting-weights-only" id="id8">Troubleshooting <code class="docutils literal notranslate"><span class="pre">weights_only</span></code></a></p>
<ul>
<li><p><a class="reference internal" href="#getting-unsafe-globals" id="id9">Getting unsafe globals</a></p></li>
<li><p><a class="reference internal" href="#environment-variables" id="id10">Environment Variables</a></p></li>
</ul>
</li>
</ul>
</li>
<li><p><a class="reference internal" href="#serializing-torch-nn-modules-and-loading-them-in-c" id="id11">Serializing torch.nn.Modules and loading them in C++</a></p></li>
<li><p><a class="reference internal" href="#saving-and-loading-scriptmodules-across-pytorch-versions" id="id12">Saving and loading ScriptModules across PyTorch versions</a></p>
<ul>
<li><p><a class="reference internal" href="#torch-div-performing-integer-division" id="id13">torch.div performing integer division</a></p></li>
<li><p><a class="reference internal" href="#torch-full-always-inferring-a-float-dtype" id="id14">torch.full always inferring a float dtype</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#utility-functions" id="id15">Utility functions</a></p></li>
<li><p><a class="reference internal" href="#module-torch.utils.serialization" id="id16">Config</a></p></li>
</ul>
</li>
</ul>
</div>
<div class="section" id="saving-and-loading-tensors">
<span id="saving-loading-tensors"></span><h2><a class="toc-backref" href="#id3">Saving and loading tensors</a><a class="headerlink" href="#saving-and-loading-tensors" title="Permalink to this heading">¶</a></h2>
<p><a class="reference internal" href="../generated/torch.save.html#torch.save" title="torch.save"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.save()</span></code></a> and <a class="reference internal" href="../generated/torch.load.html#torch.load" title="torch.load"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.load()</span></code></a> let you easily save and load tensors:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">t</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mf">1.</span><span class="p">,</span> <span class="mf">2.</span><span class="p">])</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">t</span><span class="p">,</span> <span class="s1">'tensor.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'tensor.pt'</span><span class="p">)</span>
<span class="go">tensor([1., 2.])</span>
</pre></div>
</div>
<p>By convention, PyTorch files are typically written with a ‘.pt’ or ‘.pth’ extension.</p>
<p><a class="reference internal" href="../generated/torch.save.html#torch.save" title="torch.save"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.save()</span></code></a> and <a class="reference internal" href="../generated/torch.load.html#torch.load" title="torch.load"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.load()</span></code></a> use Python’s pickle by default,
so you can also save multiple tensors as part of Python objects like tuples,
lists, and dicts:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">d</span> <span class="o">=</span> <span class="p">{</span><span class="s1">'a'</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mf">1.</span><span class="p">,</span> <span class="mf">2.</span><span class="p">]),</span> <span class="s1">'b'</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mf">3.</span><span class="p">,</span> <span class="mf">4.</span><span class="p">])}</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">d</span><span class="p">,</span> <span class="s1">'tensor_dict.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'tensor_dict.pt'</span><span class="p">)</span>
<span class="go">{'a': tensor([1., 2.]), 'b': tensor([3., 4.])}</span>
</pre></div>
</div>
<p>Custom data structures that include PyTorch tensors can also be saved if the
data structure is pickle-able.</p>
</div>
<div class="section" id="saving-and-loading-tensors-preserves-views">
<span id="preserve-storage-sharing"></span><h2><a class="toc-backref" href="#id4">Saving and loading tensors preserves views</a><a class="headerlink" href="#saving-and-loading-tensors-preserves-views" title="Permalink to this heading">¶</a></h2>
<p>Saving tensors preserves their view relationships:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">numbers</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">10</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">evens</span> <span class="o">=</span> <span class="n">numbers</span><span class="p">[</span><span class="mi">1</span><span class="p">::</span><span class="mi">2</span><span class="p">]</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">([</span><span class="n">numbers</span><span class="p">,</span> <span class="n">evens</span><span class="p">],</span> <span class="s1">'tensors.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">loaded_numbers</span><span class="p">,</span> <span class="n">loaded_evens</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'tensors.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">loaded_evens</span> <span class="o">*=</span> <span class="mi">2</span>
<span class="gp">>>> </span><span class="n">loaded_numbers</span>
<span class="go">tensor([ 1, 4, 3, 8, 5, 12, 7, 16, 9])</span>
</pre></div>
</div>
<p>Behind the scenes, these tensors share the same “storage.” See
<a class="reference external" href="https://pytorch.org/docs/main/tensor_view.html">Tensor Views</a> for more
on views and storage.</p>
<p>When PyTorch saves tensors it saves their storage objects and tensor
metadata separately. This is an implementation detail that may change in the
future, but it typically saves space and lets PyTorch easily
reconstruct the view relationships between the loaded tensors. In the above
snippet, for example, only a single storage is written to ‘tensors.pt’.</p>
<p>In some cases, however, saving the current storage objects may be unnecessary
and create prohibitively large files. In the following snippet a storage much
larger than the saved tensor is written to a file:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">large</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1000</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">small</span> <span class="o">=</span> <span class="n">large</span><span class="p">[</span><span class="mi">0</span><span class="p">:</span><span class="mi">5</span><span class="p">]</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">small</span><span class="p">,</span> <span class="s1">'small.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">loaded_small</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'small.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">loaded_small</span><span class="o">.</span><span class="n">storage</span><span class="p">()</span><span class="o">.</span><span class="n">size</span><span class="p">()</span>
<span class="go">999</span>
</pre></div>
</div>
<p>Instead of saving only the five values in the <cite>small</cite> tensor to ‘small.pt,’
the 999 values in the storage it shares with <cite>large</cite> were saved and loaded.</p>
<p>When saving tensors with fewer elements than their storage objects, the size of
the saved file can be reduced by first cloning the tensors. Cloning a tensor
produces a new tensor with a new storage object containing only the values
in the tensor:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">large</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1000</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">small</span> <span class="o">=</span> <span class="n">large</span><span class="p">[</span><span class="mi">0</span><span class="p">:</span><span class="mi">5</span><span class="p">]</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">small</span><span class="o">.</span><span class="n">clone</span><span class="p">(),</span> <span class="s1">'small.pt'</span><span class="p">)</span> <span class="c1"># saves a clone of small</span>
<span class="gp">>>> </span><span class="n">loaded_small</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'small.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">loaded_small</span><span class="o">.</span><span class="n">storage</span><span class="p">()</span><span class="o">.</span><span class="n">size</span><span class="p">()</span>
<span class="go">5</span>
</pre></div>
</div>
<p>Since the cloned tensors are independent of each other, however, they have
none of the view relationships the original tensors did. If both file size and
view relationships are important when saving tensors smaller than their
storage objects, then care must be taken to construct new tensors that minimize
the size of their storage objects but still have the desired view relationships
before saving.</p>
</div>
<div class="section" id="saving-and-loading-torch-nn-modules">
<span id="saving-loading-python-modules"></span><h2><a class="toc-backref" href="#id5">Saving and loading torch.nn.Modules</a><a class="headerlink" href="#saving-and-loading-torch-nn-modules" title="Permalink to this heading">¶</a></h2>
<p>See also: <a class="reference external" href="https://pytorch.org/tutorials/beginner/saving_loading_models.html">Tutorial: Saving and loading modules</a></p>
<p>In PyTorch, a module’s state is frequently serialized using a ‘state dict.’
A module’s state dict contains all of its parameters and persistent buffers:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">bn</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm1d</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="n">track_running_stats</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="nb">list</span><span class="p">(</span><span class="n">bn</span><span class="o">.</span><span class="n">named_parameters</span><span class="p">())</span>
<span class="go">[('weight', Parameter containing: tensor([1., 1., 1.], requires_grad=True)),</span>
<span class="go"> ('bias', Parameter containing: tensor([0., 0., 0.], requires_grad=True))]</span>
<span class="gp">>>> </span><span class="nb">list</span><span class="p">(</span><span class="n">bn</span><span class="o">.</span><span class="n">named_buffers</span><span class="p">())</span>
<span class="go">[('running_mean', tensor([0., 0., 0.])),</span>
<span class="go"> ('running_var', tensor([1., 1., 1.])),</span>
<span class="go"> ('num_batches_tracked', tensor(0))]</span>
<span class="gp">>>> </span><span class="n">bn</span><span class="o">.</span><span class="n">state_dict</span><span class="p">()</span>
<span class="go">OrderedDict([('weight', tensor([1., 1., 1.])),</span>
<span class="go"> ('bias', tensor([0., 0., 0.])),</span>
<span class="go"> ('running_mean', tensor([0., 0., 0.])),</span>
<span class="go"> ('running_var', tensor([1., 1., 1.])),</span>
<span class="go"> ('num_batches_tracked', tensor(0))])</span>
</pre></div>
</div>
<p>Instead of saving a module directly, for compatibility reasons it is recommended
to instead save only its state dict. Python modules even have a function,
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.load_state_dict" title="torch.nn.Module.load_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">load_state_dict()</span></code></a>, to restore their states from a state dict:</p>
<div class="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">save</span><span class="p">(</span><span class="n">bn</span><span class="o">.</span><span class="n">state_dict</span><span class="p">(),</span> <span class="s1">'bn.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">bn_state_dict</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'bn.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">new_bn</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm1d</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="n">track_running_stats</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">new_bn</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">bn_state_dict</span><span class="p">)</span>
<span class="go"><All keys matched successfully></span>
</pre></div>
</div>
<p>Note that the state dict is first loaded from its file with <a class="reference internal" href="../generated/torch.load.html#torch.load" title="torch.load"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.load()</span></code></a>
and the state then restored with <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.load_state_dict" title="torch.nn.Module.load_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">load_state_dict()</span></code></a>.</p>
<p>Even custom modules and modules containing other modules have state dicts and
can use this pattern:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># A module with two linear layers</span>
<span class="o">>>></span> <span class="k">class</span> <span class="nc">MyModule</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">l0</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">l1</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="n">out0</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">l0</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="n">out0_relu</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">functional</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">out0</span><span class="p">)</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">l1</span><span class="p">(</span><span class="n">out0_relu</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">m</span> <span class="o">=</span> <span class="n">MyModule</span><span class="p">()</span>
<span class="o">>>></span> <span class="n">m</span><span class="o">.</span><span class="n">state_dict</span><span class="p">()</span>
<span class="n">OrderedDict</span><span class="p">([(</span><span class="s1">'l0.weight'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([[</span> <span class="mf">0.1400</span><span class="p">,</span> <span class="mf">0.4563</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.0271</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.4406</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.3289</span><span class="p">,</span> <span class="mf">0.2827</span><span class="p">,</span> <span class="mf">0.4588</span><span class="p">,</span> <span class="mf">0.2031</span><span class="p">]])),</span>
<span class="p">(</span><span class="s1">'l0.bias'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span> <span class="mf">0.0300</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.1316</span><span class="p">])),</span>
<span class="p">(</span><span class="s1">'l1.weight'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([[</span><span class="mf">0.6533</span><span class="p">,</span> <span class="mf">0.3413</span><span class="p">]])),</span>
<span class="p">(</span><span class="s1">'l1.bias'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span><span class="o">-</span><span class="mf">0.1112</span><span class="p">]))])</span>
<span class="o">>>></span> <span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">state_dict</span><span class="p">(),</span> <span class="s1">'mymodule.pt'</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">m_state_dict</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'mymodule.pt'</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">new_m</span> <span class="o">=</span> <span class="n">MyModule</span><span class="p">()</span>
<span class="o">>>></span> <span class="n">new_m</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">m_state_dict</span><span class="p">)</span>
<span class="o"><</span><span class="n">All</span> <span class="n">keys</span> <span class="n">matched</span> <span class="n">successfully</span><span class="o">></span>
</pre></div>
</div>
</div>
<div class="section" id="serialized-file-format-for-torch-save">
<span id="serialized-file-format"></span><h2><a class="toc-backref" href="#id6">Serialized file format for <code class="docutils literal notranslate"><span class="pre">torch.save</span></code></a><a class="headerlink" href="#serialized-file-format-for-torch-save" title="Permalink to this heading">¶</a></h2>
<p>Since PyTorch 1.6.0, <code class="docutils literal notranslate"><span class="pre">torch.save</span></code> defaults to returning an uncompressed ZIP64
archive unless the user sets <code class="docutils literal notranslate"><span class="pre">_use_new_zipfile_serialization=False</span></code>.</p>
<p>In this archive, the files are ordered as such</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>checkpoint.pth
├── data.pkl
├── byteorder # added in PyTorch 2.1.0
├── data/
│ ├── 0
│ ├── 1
│ ├── 2
│ └── …
└── version
</pre></div>
</div>
<dl class="simple">
<dt>The entries are as follows:</dt><dd><ul class="simple">
<li><p><code class="docutils literal notranslate"><span class="pre">data.pkl</span></code> is the result of pickling the object passed to <code class="docutils literal notranslate"><span class="pre">torch.save</span></code>
excluding <code class="docutils literal notranslate"><span class="pre">torch.Storage</span></code> objects that it contains</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">byteorder</span></code> contains a string with the <code class="docutils literal notranslate"><span class="pre">sys.byteorder</span></code> when saving (“little” or “big”)</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">data/</span></code> contains all the storages in the object, where each storage is a separate file</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">version</span></code> contains a version number at save time that can be used at load time</p></li>
</ul>
</dd>
</dl>
<p>When saving, PyTorch will ensure that the local file header of each file is padded
to an offset that is a multiple of 64 bytes, ensuring that the offset of each file
is 64-byte aligned.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Tensors on certain devices such as XLA are serialized as pickled numpy arrays. As
such, their storages are not serialized. In these cases <code class="docutils literal notranslate"><span class="pre">data/</span></code> might not exist
in the checkpoint.</p>
</div>
</div>
<div class="section" id="torch-load-with-weights-only-true">
<span id="weights-only"></span><h2><a class="toc-backref" href="#id7"><code class="docutils literal notranslate"><span class="pre">torch.load</span></code> with <code class="docutils literal notranslate"><span class="pre">weights_only=True</span></code></a><a class="headerlink" href="#torch-load-with-weights-only-true" title="Permalink to this heading">¶</a></h2>
<p>Starting in version 2.6, <code class="docutils literal notranslate"><span class="pre">torch.load</span></code> will use <code class="docutils literal notranslate"><span class="pre">weights_only=True</span></code> if the <code class="docutils literal notranslate"><span class="pre">pickle_module</span></code>
argument is not passed.</p>
<p>As discussed in the documentation for <a class="reference internal" href="../generated/torch.load.html#torch.load" title="torch.load"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.load()</span></code></a>, <code class="docutils literal notranslate"><span class="pre">weights_only=True</span></code> restricts
the unpickler used in <code class="docutils literal notranslate"><span class="pre">torch.load</span></code> to only executing functions/building classes required for
<code class="docutils literal notranslate"><span class="pre">state_dicts</span></code> of plain <code class="docutils literal notranslate"><span class="pre">torch.Tensors</span></code> as well as some other primitive types. Further,
unlike the default <code class="docutils literal notranslate"><span class="pre">Unpickler</span></code> provided by the <code class="docutils literal notranslate"><span class="pre">pickle</span></code> module, the <code class="docutils literal notranslate"><span class="pre">weights_only</span></code> Unpickler
is not allowed to dynamically import anything during unpickling.</p>
<p>As mentioned above, saving a module’s <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> is a best practice when using <code class="docutils literal notranslate"><span class="pre">torch.save</span></code>. If loading an old
checkpoint that contains an <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code>, we recommend <code class="docutils literal notranslate"><span class="pre">weights_only=False</span></code>. When loading a checkpoint that contains
tensor subclasses, there will likely be functions/classes that need to be allowlisted, see below for further details.</p>
<p>If the <code class="docutils literal notranslate"><span class="pre">weights_only</span></code> Unpickler encounters a function or class that is not allowlisted
by default within the pickle file, you should see an actionable error like such</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span>_pickle.UnpicklingError: Weights only load failed. This file can still be loaded,
to do so you have two options, do those steps only if you trust the source of the checkpoint.
1. Re-running `torch.load` with `weights_only` set to `False` will likely succeed,
but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
2. Alternatively, to load with `weights_only=True` please check the recommended
steps in the following error message.
WeightsUnpickler error: Unsupported global: GLOBAL {__module__}.{__name__} was not an allowed global by
default. Please use `torch.serialization.add_safe_globals([{__name__}])` or the
`torch.serialization.safe_globals([{__name__}])` context manager to allowlist this global
if you trust this class/function.
</pre></div>
</div>
<p>Please follow the steps in the error message and allowlist the functions or classes only if you trust them.</p>
<p>To get all GLOBALs (functions/classes) in the checkpoint that are not yet allowlisted you can use
<a class="reference internal" href="#torch.serialization.get_unsafe_globals_in_checkpoint" title="torch.serialization.get_unsafe_globals_in_checkpoint"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.serialization.get_unsafe_globals_in_checkpoint()</span></code></a> which will return a list of strings of the form
<code class="docutils literal notranslate"><span class="pre">{__module__}.{__name__}</span></code>. If you trust these functions/classes, you can import them and allowlist them per
the error message either via <a class="reference internal" href="#torch.serialization.add_safe_globals" title="torch.serialization.add_safe_globals"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.serialization.add_safe_globals()</span></code></a> or the context manager
<a class="reference internal" href="#torch.serialization.safe_globals" title="torch.serialization.safe_globals"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.serialization.safe_globals</span></code></a>.</p>
<p>To access the list of user-allowlisted functions/classes you can use <a class="reference internal" href="#torch.serialization.get_safe_globals" title="torch.serialization.get_safe_globals"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.serialization.get_safe_globals()</span></code></a> and
to clear the current list see <a class="reference internal" href="#torch.serialization.clear_safe_globals" title="torch.serialization.clear_safe_globals"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.serialization.clear_safe_globals()</span></code></a>.</p>
<div class="section" id="troubleshooting-weights-only">
<h3><a class="toc-backref" href="#id8">Troubleshooting <code class="docutils literal notranslate"><span class="pre">weights_only</span></code></a><a class="headerlink" href="#troubleshooting-weights-only" title="Permalink to this heading">¶</a></h3>
<div class="section" id="getting-unsafe-globals">
<h4><a class="toc-backref" href="#id9">Getting unsafe globals</a><a class="headerlink" href="#getting-unsafe-globals" title="Permalink to this heading">¶</a></h4>
<p>A caveat is that <a class="reference internal" href="#torch.serialization.get_unsafe_globals_in_checkpoint" title="torch.serialization.get_unsafe_globals_in_checkpoint"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.serialization.get_unsafe_globals_in_checkpoint()</span></code></a> analyzes the checkpoint statically,
some types might be built dynamically during the unpickling process and hence will not be reported by
<a class="reference internal" href="#torch.serialization.get_unsafe_globals_in_checkpoint" title="torch.serialization.get_unsafe_globals_in_checkpoint"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.serialization.get_unsafe_globals_in_checkpoint()</span></code></a>. One such example is <code class="docutils literal notranslate"><span class="pre">dtypes</span></code> in numpy. In
<code class="docutils literal notranslate"><span class="pre">numpy</span> <span class="pre"><</span> <span class="pre">1.25</span></code> after allowlisting all the functions/classes reported by
<a class="reference internal" href="#torch.serialization.get_unsafe_globals_in_checkpoint" title="torch.serialization.get_unsafe_globals_in_checkpoint"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.serialization.get_unsafe_globals_in_checkpoint()</span></code></a> you might see an error like</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span>WeightsUnpickler error: Can only build Tensor, Parameter, OrderedDict or types allowlisted via `add_safe_globals`,
but got <class 'numpy.dtype[float32]'>
</pre></div>
</div>
<p>This can be allowlisted via <code class="docutils literal notranslate"><span class="pre">{add_}safe_globals([type(np.dtype(np.float32))])</span></code>.</p>
<p>In <code class="docutils literal notranslate"><span class="pre">numpy</span> <span class="pre">>=1.25</span></code> you would see</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span>WeightsUnpickler error: Can only build Tensor, Parameter, OrderedDict or types allowlisted via `add_safe_globals`,
but got <class 'numpy.dtypes.Float32DType'>
</pre></div>
</div>
<p>This can be allowlisted via <code class="docutils literal notranslate"><span class="pre">{add_}safe_globals([np.dtypes.Float32DType])</span></code>.</p>
</div>
<div class="section" id="environment-variables">
<h4><a class="toc-backref" href="#id10">Environment Variables</a><a class="headerlink" href="#environment-variables" title="Permalink to this heading">¶</a></h4>
<p>There are two environment variables that will influence the behavior of <code class="docutils literal notranslate"><span class="pre">torch.load</span></code>. These can be helpful
if one does not have access to the <code class="docutils literal notranslate"><span class="pre">torch.load</span></code> callsites.</p>
<ul class="simple">
<li><p><code class="docutils literal notranslate"><span class="pre">TORCH_FORCE_WEIGHTS_ONLY_LOAD=1</span></code> will override all <code class="docutils literal notranslate"><span class="pre">torch.load</span></code> callsites to use <code class="docutils literal notranslate"><span class="pre">weights_only=True</span></code>.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1</span></code> will make <code class="docutils literal notranslate"><span class="pre">torch.load</span></code> callsites use <code class="docutils literal notranslate"><span class="pre">weights_only=False</span></code> <strong>only</strong>
if <code class="docutils literal notranslate"><span class="pre">weights_only</span></code> was not passed as an argument.</p></li>
</ul>
</div>
</div>
</div>
<div class="section" id="serializing-torch-nn-modules-and-loading-them-in-c">
<span id="serializing-python-modules"></span><h2><a class="toc-backref" href="#id11">Serializing torch.nn.Modules and loading them in C++</a><a class="headerlink" href="#serializing-torch-nn-modules-and-loading-them-in-c" title="Permalink to this heading">¶</a></h2>
<p>See also: <a class="reference external" href="https://pytorch.org/tutorials/advanced/cpp_export.html">Tutorial: Loading a TorchScript Model in C++</a></p>
<p>ScriptModules can be serialized as a TorchScript program and loaded
using <a class="reference internal" href="../generated/torch.jit.load.html#torch.jit.load" title="torch.jit.load"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.jit.load()</span></code></a>.
This serialization encodes all the modules’ methods, submodules, parameters,
and attributes, and it allows the serialized program to be loaded in C++
(i.e. without Python).</p>
<p>The distinction between <a class="reference internal" href="../generated/torch.jit.save.html#torch.jit.save" title="torch.jit.save"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.jit.save()</span></code></a> and <a class="reference internal" href="../generated/torch.save.html#torch.save" title="torch.save"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.save()</span></code></a> may not
be immediately clear. <a class="reference internal" href="../generated/torch.save.html#torch.save" title="torch.save"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.save()</span></code></a> saves Python objects with pickle.
This is especially useful for prototyping, researching, and training.
<a class="reference internal" href="../generated/torch.jit.save.html#torch.jit.save" title="torch.jit.save"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.jit.save()</span></code></a>, on the other hand, serializes ScriptModules to a format
that can be loaded in Python or C++. This is useful when saving and loading C++
modules or for running modules trained in Python with C++, a common practice
when deploying PyTorch models.</p>
<p>To script, serialize and load a module in Python:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">scripted_module</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">script</span><span class="p">(</span><span class="n">MyModule</span><span class="p">())</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">scripted_module</span><span class="p">,</span> <span class="s1">'mymodule.pt'</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'mymodule.pt'</span><span class="p">)</span>
<span class="go">RecursiveScriptModule( original_name=MyModule</span>
<span class="go"> (l0): RecursiveScriptModule(original_name=Linear)</span>
<span class="go"> (l1): RecursiveScriptModule(original_name=Linear) )</span>
</pre></div>
</div>
<p>Traced modules can also be saved with <a class="reference internal" href="../generated/torch.jit.save.html#torch.jit.save" title="torch.jit.save"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.jit.save()</span></code></a>, with the caveat
that only the traced code path is serialized. The following example demonstrates
this:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># A module with control flow</span>
<span class="o">>>></span> <span class="k">class</span> <span class="nc">ControlFlowModule</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">l0</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">l1</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="k">if</span> <span class="nb">input</span><span class="o">.</span><span class="n">dim</span><span class="p">()</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
<span class="k">return</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
<span class="n">out0</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">l0</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="n">out0_relu</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">functional</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">out0</span><span class="p">)</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">l1</span><span class="p">(</span><span class="n">out0_relu</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">traced_module</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">trace</span><span class="p">(</span><span class="n">ControlFlowModule</span><span class="p">(),</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">))</span>
<span class="o">>>></span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">traced_module</span><span class="p">,</span> <span class="s1">'controlflowmodule_traced.pt'</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">loaded</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'controlflowmodule_traced.pt'</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">loaded</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">4</span><span class="p">)))</span>
<span class="n">tensor</span><span class="p">([[</span><span class="o">-</span><span class="mf">0.1571</span><span class="p">],</span> <span class="p">[</span><span class="o">-</span><span class="mf">0.3793</span><span class="p">]],</span> <span class="n">grad_fn</span><span class="o">=<</span><span class="n">AddBackward0</span><span class="o">></span><span class="p">)</span>
<span class="o">>>></span> <span class="n">scripted_module</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">script</span><span class="p">(</span><span class="n">ControlFlowModule</span><span class="p">(),</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">))</span>
<span class="o">>>></span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">scripted_module</span><span class="p">,</span> <span class="s1">'controlflowmodule_scripted.pt'</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">loaded</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'controlflowmodule_scripted.pt'</span><span class="p">)</span>
<span class="o">>></span> <span class="n">loaded</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">4</span><span class="p">))</span>
<span class="n">tensor</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
</pre></div>
</div>
<p>The above module has an if statement that is not triggered by the traced inputs,
and so is not part of the traced module and not serialized with it.
The scripted module, however, contains the if statement and is serialized with it.
See the <a class="reference external" href="https://pytorch.org/docs/stable/jit.html">TorchScript documentation</a>
for more on scripting and tracing.</p>
<p>Finally, to load the module in C++:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">torch</span><span class="p">::</span><span class="n">jit</span><span class="p">::</span><span class="n">script</span><span class="p">::</span><span class="n">Module</span> <span class="n">module</span><span class="p">;</span>
<span class="gp">>>> </span><span class="n">module</span> <span class="o">=</span> <span class="n">torch</span><span class="p">::</span><span class="n">jit</span><span class="p">::</span><span class="n">load</span><span class="p">(</span><span class="s1">'controlflowmodule_scripted.pt'</span><span class="p">);</span>
</pre></div>
</div>
<p>See the <a class="reference external" href="https://pytorch.org/cppdocs/">PyTorch C++ API documentation</a>
for details about how to use PyTorch modules in C++.</p>
</div>
<div class="section" id="saving-and-loading-scriptmodules-across-pytorch-versions">
<span id="saving-loading-across-versions"></span><h2><a class="toc-backref" href="#id12">Saving and loading ScriptModules across PyTorch versions</a><a class="headerlink" href="#saving-and-loading-scriptmodules-across-pytorch-versions" title="Permalink to this heading">¶</a></h2>
<p>The PyTorch Team recommends saving and loading modules with the same version of
PyTorch. Older versions of PyTorch may not support newer modules, and newer
versions may have removed or modified older behavior. These changes are
explicitly described in
PyTorch’s <a class="reference external" href="https://github.com/pytorch/pytorch/releases">release notes</a>,
and modules relying on functionality that has changed may need to be updated
to continue working properly. In limited cases, detailed below, PyTorch will
preserve the historic behavior of serialized ScriptModules so they do not require
an update.</p>
<div class="section" id="torch-div-performing-integer-division">
<h3><a class="toc-backref" href="#id13">torch.div performing integer division</a><a class="headerlink" href="#torch-div-performing-integer-division" title="Permalink to this heading">¶</a></h3>
<p>In PyTorch 1.5 and earlier <a class="reference internal" href="../generated/torch.div.html#torch.div" title="torch.div"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.div()</span></code></a> would perform floor division when
given two integer inputs:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># PyTorch 1.5 (and earlier)</span>
<span class="o">>>></span> <span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="mi">5</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="mi">3</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">a</span> <span class="o">/</span> <span class="n">b</span>
<span class="n">tensor</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
</pre></div>
</div>
<p>In PyTorch 1.7, however, <a class="reference internal" href="../generated/torch.div.html#torch.div" title="torch.div"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.div()</span></code></a> will always perform a true division
of its inputs, just like division in Python 3:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># PyTorch 1.7</span>
<span class="o">>>></span> <span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="mi">5</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="mi">3</span><span class="p">)</span>
<span class="o">>>></span> <span class="n">a</span> <span class="o">/</span> <span class="n">b</span>
<span class="n">tensor</span><span class="p">(</span><span class="mf">1.6667</span><span class="p">)</span>
</pre></div>
</div>
<p>The behavior of <a class="reference internal" href="../generated/torch.div.html#torch.div" title="torch.div"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.div()</span></code></a> is preserved in serialized ScriptModules.
That is, ScriptModules serialized with versions of PyTorch before 1.6 will continue
to see <a class="reference internal" href="../generated/torch.div.html#torch.div" title="torch.div"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.div()</span></code></a> perform floor division when given two integer inputs
even when loaded with newer versions of PyTorch. ScriptModules using <a class="reference internal" href="../generated/torch.div.html#torch.div" title="torch.div"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.div()</span></code></a>
and serialized on PyTorch 1.6 and later cannot be loaded in earlier versions of
PyTorch, however, since those earlier versions do not understand the new behavior.</p>
</div>
<div class="section" id="torch-full-always-inferring-a-float-dtype">
<h3><a class="toc-backref" href="#id14">torch.full always inferring a float dtype</a><a class="headerlink" href="#torch-full-always-inferring-a-float-dtype" title="Permalink to this heading">¶</a></h3>
<p>In PyTorch 1.5 and earlier <a class="reference internal" href="../generated/torch.full.html#torch.full" title="torch.full"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.full()</span></code></a> always returned a float tensor,
regardless of the fill value it’s given:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># PyTorch 1.5 and earlier</span>
<span class="o">>>></span> <span class="n">torch</span><span class="o">.</span><span class="n">full</span><span class="p">((</span><span class="mi">3</span><span class="p">,),</span> <span class="mi">1</span><span class="p">)</span> <span class="c1"># Note the integer fill value...</span>
<span class="n">tensor</span><span class="p">([</span><span class="mf">1.</span><span class="p">,</span> <span class="mf">1.</span><span class="p">,</span> <span class="mf">1.</span><span class="p">])</span> <span class="c1"># ...but float tensor!</span>
</pre></div>
</div>
<p>In PyTorch 1.7, however, <a class="reference internal" href="../generated/torch.full.html#torch.full" title="torch.full"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.full()</span></code></a> will infer the returned tensor’s
dtype from the fill value:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># PyTorch 1.7</span>
<span class="o">>>></span> <span class="n">torch</span><span class="o">.</span><span class="n">full</span><span class="p">((</span><span class="mi">3</span><span class="p">,),</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">tensor</span><span class="p">([</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">])</span>
<span class="o">>>></span> <span class="n">torch</span><span class="o">.</span><span class="n">full</span><span class="p">((</span><span class="mi">3</span><span class="p">,),</span> <span class="kc">True</span><span class="p">)</span>
<span class="n">tensor</span><span class="p">([</span><span class="kc">True</span><span class="p">,</span> <span class="kc">True</span><span class="p">,</span> <span class="kc">True</span><span class="p">])</span>
<span class="o">>>></span> <span class="n">torch</span><span class="o">.</span><span class="n">full</span><span class="p">((</span><span class="mi">3</span><span class="p">,),</span> <span class="mf">1.</span><span class="p">)</span>
<span class="n">tensor</span><span class="p">([</span><span class="mf">1.</span><span class="p">,</span> <span class="mf">1.</span><span class="p">,</span> <span class="mf">1.</span><span class="p">])</span>
<span class="o">>>></span> <span class="n">torch</span><span class="o">.</span><span class="n">full</span><span class="p">((</span><span class="mi">3</span><span class="p">,),</span> <span class="mi">1</span> <span class="o">+</span> <span class="mi">1</span><span class="n">j</span><span class="p">)</span>
<span class="n">tensor</span><span class="p">([</span><span class="mf">1.</span><span class="o">+</span><span class="mf">1.</span><span class="n">j</span><span class="p">,</span> <span class="mf">1.</span><span class="o">+</span><span class="mf">1.</span><span class="n">j</span><span class="p">,</span> <span class="mf">1.</span><span class="o">+</span><span class="mf">1.</span><span class="n">j</span><span class="p">])</span>
</pre></div>