-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathexport.html
2585 lines (2295 loc) · 304 KB
/
export.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>torch.export — PyTorch 2.7 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/export.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="torch.export Programming Model" href="export.programming_model.html" />
<link rel="prev" title="torch.backends" href="backends.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 current"><a class="current reference internal" href="#">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>torch.export</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/export.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="torch-export">
<span id="id1"></span><h1>torch.export<a class="headerlink" href="#torch-export" title="Permalink to this heading">¶</a></h1>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>This feature is a prototype under active development and there WILL BE
BREAKING CHANGES in the future.</p>
</div>
<div class="section" id="overview">
<h2>Overview<a class="headerlink" href="#overview" title="Permalink to this heading">¶</a></h2>
<p><a class="reference internal" href="#torch.export.export" title="torch.export.export"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.export()</span></code></a> takes a <a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.Module</span></code></a> and produces a traced graph
representing only the Tensor computation of the function in an Ahead-of-Time
(AOT) fashion, which can subsequently be executed with different outputs or
serialized.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">from</span> <span class="nn">torch.export</span> <span class="kn">import</span> <span class="n">export</span>
<span class="k">class</span> <span class="nc">Mod</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="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</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="n">y</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="o">-></span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">:</span>
<span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">sin</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cos</span><span class="p">(</span><span class="n">y</span><span class="p">)</span>
<span class="k">return</span> <span class="n">a</span> <span class="o">+</span> <span class="n">b</span>
<span class="n">example_args</span> <span class="o">=</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">10</span><span class="p">,</span> <span class="mi">10</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">10</span><span class="p">,</span> <span class="mi">10</span><span class="p">))</span>
<span class="n">exported_program</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">export</span><span class="o">.</span><span class="n">ExportedProgram</span> <span class="o">=</span> <span class="n">export</span><span class="p">(</span>
<span class="n">Mod</span><span class="p">(),</span> <span class="n">args</span><span class="o">=</span><span class="n">example_args</span>
<span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">exported_program</span><span class="p">)</span>
</pre></div>
</div>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">ExportedProgram</span><span class="p">:</span>
<span class="k">class</span> <span class="nc">GraphModule</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="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">:</span> <span class="s2">"f32[10, 10]"</span><span class="p">,</span> <span class="n">y</span><span class="p">:</span> <span class="s2">"f32[10, 10]"</span><span class="p">):</span>
<span class="c1"># code: a = torch.sin(x)</span>
<span class="n">sin</span><span class="p">:</span> <span class="s2">"f32[10, 10]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">sin</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="c1"># code: b = torch.cos(y)</span>
<span class="n">cos</span><span class="p">:</span> <span class="s2">"f32[10, 10]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">cos</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">y</span><span class="p">)</span>
<span class="c1"># code: return a + b</span>
<span class="n">add</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">10</span><span class="p">,</span> <span class="mi">10</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">add</span><span class="o">.</span><span class="n">Tensor</span><span class="p">(</span><span class="n">sin</span><span class="p">,</span> <span class="n">cos</span><span class="p">)</span>
<span class="k">return</span> <span class="p">(</span><span class="n">add</span><span class="p">,)</span>
<span class="n">Graph</span> <span class="n">signature</span><span class="p">:</span>
<span class="n">ExportGraphSignature</span><span class="p">(</span>
<span class="n">input_specs</span><span class="o">=</span><span class="p">[</span>
<span class="n">InputSpec</span><span class="p">(</span>
<span class="n">kind</span><span class="o">=<</span><span class="n">InputKind</span><span class="o">.</span><span class="n">USER_INPUT</span><span class="p">:</span> <span class="mi">1</span><span class="o">></span><span class="p">,</span>
<span class="n">arg</span><span class="o">=</span><span class="n">TensorArgument</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s1">'x'</span><span class="p">),</span>
<span class="n">target</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">persistent</span><span class="o">=</span><span class="kc">None</span>
<span class="p">),</span>
<span class="n">InputSpec</span><span class="p">(</span>
<span class="n">kind</span><span class="o">=<</span><span class="n">InputKind</span><span class="o">.</span><span class="n">USER_INPUT</span><span class="p">:</span> <span class="mi">1</span><span class="o">></span><span class="p">,</span>
<span class="n">arg</span><span class="o">=</span><span class="n">TensorArgument</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s1">'y'</span><span class="p">),</span>
<span class="n">target</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">persistent</span><span class="o">=</span><span class="kc">None</span>
<span class="p">)</span>
<span class="p">],</span>
<span class="n">output_specs</span><span class="o">=</span><span class="p">[</span>
<span class="n">OutputSpec</span><span class="p">(</span>
<span class="n">kind</span><span class="o">=<</span><span class="n">OutputKind</span><span class="o">.</span><span class="n">USER_OUTPUT</span><span class="p">:</span> <span class="mi">1</span><span class="o">></span><span class="p">,</span>
<span class="n">arg</span><span class="o">=</span><span class="n">TensorArgument</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s1">'add'</span><span class="p">),</span>
<span class="n">target</span><span class="o">=</span><span class="kc">None</span>
<span class="p">)</span>
<span class="p">]</span>
<span class="p">)</span>
<span class="n">Range</span> <span class="n">constraints</span><span class="p">:</span> <span class="p">{}</span>
</pre></div>
</div>
<p><code class="docutils literal notranslate"><span class="pre">torch.export</span></code> produces a clean intermediate representation (IR) with the
following invariants. More specifications about the IR can be found
<a class="reference internal" href="export.ir_spec.html#export-ir-spec"><span class="std std-ref">here</span></a>.</p>
<ul class="simple">
<li><p><strong>Soundness</strong>: It is guaranteed to be a sound representation of the original
program, and maintains the same calling conventions of the original program.</p></li>
<li><p><strong>Normalized</strong>: There are no Python semantics within the graph. Submodules
from the original programs are inlined to form one fully flattened
computational graph.</p></li>
<li><p><strong>Graph properties</strong>: The graph is purely functional, meaning it does not
contain operations with side effects such as mutations or aliasing. It does
not mutate any intermediate values, parameters, or buffers.</p></li>
<li><p><strong>Metadata</strong>: The graph contains metadata captured during tracing, such as a
stacktrace from user’s code.</p></li>
</ul>
<p>Under the hood, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> leverages the following latest technologies:</p>
<ul class="simple">
<li><p><strong>TorchDynamo (torch._dynamo)</strong> is an internal API that uses a CPython feature
called the Frame Evaluation API to safely trace PyTorch graphs. This
provides a massively improved graph capturing experience, with much fewer
rewrites needed in order to fully trace the PyTorch code.</p></li>
<li><p><strong>AOT Autograd</strong> provides a functionalized PyTorch graph and ensures the graph
is decomposed/lowered to the ATen operator set.</p></li>
<li><p><strong>Torch FX (torch.fx)</strong> is the underlying representation of the graph,
allowing flexible Python-based transformations.</p></li>
</ul>
<div class="section" id="existing-frameworks">
<h3>Existing frameworks<a class="headerlink" href="#existing-frameworks" title="Permalink to this heading">¶</a></h3>
<p><a class="reference internal" href="generated/torch.compile.html#torch.compile" title="torch.compile"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.compile()</span></code></a> also utilizes the same PT2 stack as <code class="docutils literal notranslate"><span class="pre">torch.export</span></code>, but
is slightly different:</p>
<ul class="simple">
<li><p><strong>JIT vs. AOT</strong>: <a class="reference internal" href="generated/torch.compile.html#torch.compile" title="torch.compile"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.compile()</span></code></a> is a JIT compiler whereas
which is not intended to be used to produce compiled artifacts outside of
deployment.</p></li>
<li><p><strong>Partial vs. Full Graph Capture</strong>: When <a class="reference internal" href="generated/torch.compile.html#torch.compile" title="torch.compile"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.compile()</span></code></a> runs into an
untraceable part of a model, it will “graph break” and fall back to running
the program in the eager Python runtime. In comparison, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> aims
to get a full graph representation of a PyTorch model, so it will error out
when something untraceable is reached. Since <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> produces a full
graph disjoint from any Python features or runtime, this graph can then be
saved, loaded, and run in different environments and languages.</p></li>
<li><p><strong>Usability tradeoff</strong>: Since <a class="reference internal" href="generated/torch.compile.html#torch.compile" title="torch.compile"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.compile()</span></code></a> is able to fallback to the
Python runtime whenever it reaches something untraceable, it is a lot more
flexible. <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> will instead require users to provide more
information or rewrite their code to make it traceable.</p></li>
</ul>
<p>Compared to <a class="reference internal" href="fx.html#torch.fx.symbolic_trace" title="torch.fx.symbolic_trace"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.fx.symbolic_trace()</span></code></a>, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> traces using
TorchDynamo which operates at the Python bytecode level, giving it the ability
to trace arbitrary Python constructs not limited by what Python operator
overloading supports. Additionally, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> keeps fine-grained track of
tensor metadata, so that conditionals on things like tensor shapes do not
fail tracing. In general, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> is expected to work on more user
programs, and produce lower-level graphs (at the <code class="docutils literal notranslate"><span class="pre">torch.ops.aten</span></code> operator
level). Note that users can still use <a class="reference internal" href="fx.html#torch.fx.symbolic_trace" title="torch.fx.symbolic_trace"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.fx.symbolic_trace()</span></code></a> as a
preprocessing step before <code class="docutils literal notranslate"><span class="pre">torch.export</span></code>.</p>
<p>Compared to <a class="reference internal" href="generated/torch.jit.script.html#torch.jit.script" title="torch.jit.script"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.jit.script()</span></code></a>, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> does not capture Python
control flow or data structures, but it supports more Python language features
than TorchScript (as it is easier to have comprehensive coverage over Python
bytecodes). The resulting graphs are simpler and only have straight line control
flow (except for explicit control flow operators).</p>
<p>Compared to <a class="reference internal" href="generated/torch.jit.trace.html#torch.jit.trace" title="torch.jit.trace"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.jit.trace()</span></code></a>, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> is sound: it is able to
trace code that performs integer computation on sizes and records all of the
side-conditions necessary to show that a particular trace is valid for other
inputs.</p>
</div>
</div>
<div class="section" id="exporting-a-pytorch-model">
<h2>Exporting a PyTorch Model<a class="headerlink" href="#exporting-a-pytorch-model" title="Permalink to this heading">¶</a></h2>
<div class="section" id="an-example">
<h3>An Example<a class="headerlink" href="#an-example" title="Permalink to this heading">¶</a></h3>
<p>The main entrypoint is through <a class="reference internal" href="#torch.export.export" title="torch.export.export"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.export()</span></code></a>, which takes a
callable (<a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.Module</span></code></a>, function, or method) and sample inputs, and
captures the computation graph into an <a class="reference internal" href="#torch.export.ExportedProgram" title="torch.export.ExportedProgram"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.export.ExportedProgram</span></code></a>. An
example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">from</span> <span class="nn">torch.export</span> <span class="kn">import</span> <span class="n">export</span>
<span class="c1"># Simple module for demonstration</span>
<span class="k">class</span> <span class="nc">M</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="o">-></span> <span class="kc">None</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">conv</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">Conv2d</span><span class="p">(</span>
<span class="n">in_channels</span><span class="o">=</span><span class="mi">3</span><span class="p">,</span> <span class="n">out_channels</span><span class="o">=</span><span class="mi">16</span><span class="p">,</span> <span class="n">kernel_size</span><span class="o">=</span><span class="mi">3</span><span class="p">,</span> <span class="n">padding</span><span class="o">=</span><span class="mi">1</span>
<span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">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">ReLU</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">maxpool</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">MaxPool2d</span><span class="p">(</span><span class="n">kernel_size</span><span class="o">=</span><span class="mi">3</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="n">x</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="o">*</span><span class="p">,</span> <span class="n">constant</span><span class="o">=</span><span class="kc">None</span><span class="p">)</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="n">a</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">a</span><span class="o">.</span><span class="n">add_</span><span class="p">(</span><span class="n">constant</span><span class="p">)</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">maxpool</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">a</span><span class="p">))</span>
<span class="n">example_args</span> <span class="o">=</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">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="mi">256</span><span class="p">,</span> <span class="mi">256</span><span class="p">),)</span>
<span class="n">example_kwargs</span> <span class="o">=</span> <span class="p">{</span><span class="s2">"constant"</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">16</span><span class="p">,</span> <span class="mi">256</span><span class="p">,</span> <span class="mi">256</span><span class="p">)}</span>
<span class="n">exported_program</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">export</span><span class="o">.</span><span class="n">ExportedProgram</span> <span class="o">=</span> <span class="n">export</span><span class="p">(</span>
<span class="n">M</span><span class="p">(),</span> <span class="n">args</span><span class="o">=</span><span class="n">example_args</span><span class="p">,</span> <span class="n">kwargs</span><span class="o">=</span><span class="n">example_kwargs</span>
<span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">exported_program</span><span class="p">)</span>
</pre></div>
</div>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">ExportedProgram</span><span class="p">:</span>
<span class="k">class</span> <span class="nc">GraphModule</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="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">p_conv_weight</span><span class="p">:</span> <span class="s2">"f32[16, 3, 3, 3]"</span><span class="p">,</span> <span class="n">p_conv_bias</span><span class="p">:</span> <span class="s2">"f32[16]"</span><span class="p">,</span> <span class="n">x</span><span class="p">:</span> <span class="s2">"f32[1, 3, 256, 256]"</span><span class="p">,</span> <span class="n">constant</span><span class="p">:</span> <span class="s2">"f32[1, 16, 256, 256]"</span><span class="p">):</span>
<span class="c1"># code: a = self.conv(x)</span>
<span class="n">conv2d</span><span class="p">:</span> <span class="s2">"f32[1, 16, 256, 256]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">conv2d</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">p_conv_weight</span><span class="p">,</span> <span class="n">p_conv_bias</span><span class="p">,</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="p">[</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">])</span>
<span class="c1"># code: a.add_(constant)</span>
<span class="n">add_</span><span class="p">:</span> <span class="s2">"f32[1, 16, 256, 256]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">add_</span><span class="o">.</span><span class="n">Tensor</span><span class="p">(</span><span class="n">conv2d</span><span class="p">,</span> <span class="n">constant</span><span class="p">)</span>
<span class="c1"># code: return self.maxpool(self.relu(a))</span>
<span class="n">relu</span><span class="p">:</span> <span class="s2">"f32[1, 16, 256, 256]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">relu</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">add_</span><span class="p">)</span>
<span class="n">max_pool2d</span><span class="p">:</span> <span class="s2">"f32[1, 16, 85, 85]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">max_pool2d</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">relu</span><span class="p">,</span> <span class="p">[</span><span class="mi">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">],</span> <span class="p">[</span><span class="mi">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">])</span>
<span class="k">return</span> <span class="p">(</span><span class="n">max_pool2d</span><span class="p">,)</span>
<span class="n">Graph</span> <span class="n">signature</span><span class="p">:</span>
<span class="n">ExportGraphSignature</span><span class="p">(</span>
<span class="n">input_specs</span><span class="o">=</span><span class="p">[</span>
<span class="n">InputSpec</span><span class="p">(</span>
<span class="n">kind</span><span class="o">=<</span><span class="n">InputKind</span><span class="o">.</span><span class="n">PARAMETER</span><span class="p">:</span> <span class="mi">2</span><span class="o">></span><span class="p">,</span>
<span class="n">arg</span><span class="o">=</span><span class="n">TensorArgument</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s1">'p_conv_weight'</span><span class="p">),</span>
<span class="n">target</span><span class="o">=</span><span class="s1">'conv.weight'</span><span class="p">,</span>
<span class="n">persistent</span><span class="o">=</span><span class="kc">None</span>
<span class="p">),</span>
<span class="n">InputSpec</span><span class="p">(</span>
<span class="n">kind</span><span class="o">=<</span><span class="n">InputKind</span><span class="o">.</span><span class="n">PARAMETER</span><span class="p">:</span> <span class="mi">2</span><span class="o">></span><span class="p">,</span>
<span class="n">arg</span><span class="o">=</span><span class="n">TensorArgument</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s1">'p_conv_bias'</span><span class="p">),</span>
<span class="n">target</span><span class="o">=</span><span class="s1">'conv.bias'</span><span class="p">,</span>
<span class="n">persistent</span><span class="o">=</span><span class="kc">None</span>
<span class="p">),</span>
<span class="n">InputSpec</span><span class="p">(</span>
<span class="n">kind</span><span class="o">=<</span><span class="n">InputKind</span><span class="o">.</span><span class="n">USER_INPUT</span><span class="p">:</span> <span class="mi">1</span><span class="o">></span><span class="p">,</span>
<span class="n">arg</span><span class="o">=</span><span class="n">TensorArgument</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s1">'x'</span><span class="p">),</span>
<span class="n">target</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">persistent</span><span class="o">=</span><span class="kc">None</span>
<span class="p">),</span>
<span class="n">InputSpec</span><span class="p">(</span>
<span class="n">kind</span><span class="o">=<</span><span class="n">InputKind</span><span class="o">.</span><span class="n">USER_INPUT</span><span class="p">:</span> <span class="mi">1</span><span class="o">></span><span class="p">,</span>
<span class="n">arg</span><span class="o">=</span><span class="n">TensorArgument</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s1">'constant'</span><span class="p">),</span>
<span class="n">target</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">persistent</span><span class="o">=</span><span class="kc">None</span>
<span class="p">)</span>
<span class="p">],</span>
<span class="n">output_specs</span><span class="o">=</span><span class="p">[</span>
<span class="n">OutputSpec</span><span class="p">(</span>
<span class="n">kind</span><span class="o">=<</span><span class="n">OutputKind</span><span class="o">.</span><span class="n">USER_OUTPUT</span><span class="p">:</span> <span class="mi">1</span><span class="o">></span><span class="p">,</span>
<span class="n">arg</span><span class="o">=</span><span class="n">TensorArgument</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s1">'max_pool2d'</span><span class="p">),</span>
<span class="n">target</span><span class="o">=</span><span class="kc">None</span>
<span class="p">)</span>
<span class="p">]</span>
<span class="p">)</span>
<span class="n">Range</span> <span class="n">constraints</span><span class="p">:</span> <span class="p">{}</span>
</pre></div>
</div>
<p>Inspecting the <code class="docutils literal notranslate"><span class="pre">ExportedProgram</span></code>, we can note the following:</p>
<ul class="simple">
<li><p>The <a class="reference internal" href="fx.html#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.fx.Graph</span></code></a> contains the computation graph of the original
program, along with records of the original code for easy debugging.</p></li>
<li><p>The graph contains only <code class="docutils literal notranslate"><span class="pre">torch.ops.aten</span></code> operators found <a class="reference external" href="https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/native_functions.yaml">here</a>
and custom operators, and is fully functional, without any inplace operators
such as <code class="docutils literal notranslate"><span class="pre">torch.add_</span></code>.</p></li>
<li><p>The parameters (weight and bias to conv) are lifted as inputs to the graph,
resulting in no <code class="docutils literal notranslate"><span class="pre">get_attr</span></code> nodes in the graph, which previously existed in
the result of <a class="reference internal" href="fx.html#torch.fx.symbolic_trace" title="torch.fx.symbolic_trace"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.fx.symbolic_trace()</span></code></a>.</p></li>
<li><p>The <a class="reference internal" href="#torch.export.ExportGraphSignature" title="torch.export.ExportGraphSignature"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.export.ExportGraphSignature</span></code></a> models the input and output
signature, along with specifying which inputs are parameters.</p></li>
<li><p>The resulting shape and dtype of tensors produced by each node in the graph is
noted. For example, the <code class="docutils literal notranslate"><span class="pre">convolution</span></code> node will result in a tensor of dtype
<code class="docutils literal notranslate"><span class="pre">torch.float32</span></code> and shape (1, 16, 256, 256).</p></li>
</ul>
</div>
<div class="section" id="non-strict-export">
<span id="id2"></span><h3>Non-Strict Export<a class="headerlink" href="#non-strict-export" title="Permalink to this heading">¶</a></h3>
<p>In PyTorch 2.3, we introduced a new mode of tracing called <strong>non-strict mode</strong>.
It’s still going through hardening, so if you run into any issues, please file
them to Github with the “oncall: export” tag.</p>
<p>In <em>non-strict mode</em>, we trace through the program using the Python interpreter.
Your code will execute exactly as it would in eager mode; the only difference is
that all Tensor objects will be replaced by ProxyTensors, which will record all
their operations into a graph.</p>
<p>In <em>strict</em> mode, which is currently the default, we first trace through the
program using TorchDynamo, a bytecode analysis engine. TorchDynamo does not
actually execute your Python code. Instead, it symbolically analyzes it and
builds a graph based on the results. This analysis allows torch.export to
provide stronger guarantees about safety, but not all Python code is supported.</p>
<p>An example of a case where one might want to use non-strict mode is if you run
into a unsupported TorchDynamo feature that might not be easily solved, and you
know the python code is not exactly needed for computation. For example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">contextlib</span>
<span class="kn">import</span> <span class="nn">torch</span>
<span class="k">class</span> <span class="nc">ContextManager</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="bp">self</span><span class="o">.</span><span class="n">count</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">def</span> <span class="fm">__enter__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">count</span> <span class="o">+=</span> <span class="mi">1</span>
<span class="k">def</span> <span class="fm">__exit__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">exc_type</span><span class="p">,</span> <span class="n">exc_value</span><span class="p">,</span> <span class="n">traceback</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">count</span> <span class="o">-=</span> <span class="mi">1</span>
<span class="k">class</span> <span class="nc">M</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="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="k">with</span> <span class="n">ContextManager</span><span class="p">():</span>
<span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">sin</span><span class="p">()</span> <span class="o">+</span> <span class="n">x</span><span class="o">.</span><span class="n">cos</span><span class="p">()</span>
<span class="n">export</span><span class="p">(</span><span class="n">M</span><span class="p">(),</span> <span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">),),</span> <span class="n">strict</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span> <span class="c1"># Non-strict traces successfully</span>
<span class="n">export</span><span class="p">(</span><span class="n">M</span><span class="p">(),</span> <span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">),))</span> <span class="c1"># Strict mode fails with torch._dynamo.exc.Unsupported: ContextManager</span>
</pre></div>
</div>
<p>In this example, the first call using non-strict mode (through the
<code class="docutils literal notranslate"><span class="pre">strict=False</span></code> flag) traces successfully whereas the second call using strict
mode (default) results with a failure, where TorchDynamo is unable to support
context managers. One option is to rewrite the code (see <a class="reference internal" href="#limitations-of-torch-export"><span class="std std-ref">Limitations of torch.export</span></a>), but seeing as the context manager does not affect the tensor
computations in the model, we can go with the non-strict mode’s result.</p>
</div>
<div class="section" id="export-for-training-and-inference">
<span id="training-export"></span><h3>Export for Training and Inference<a class="headerlink" href="#export-for-training-and-inference" title="Permalink to this heading">¶</a></h3>
<p>In PyTorch 2.5, we introduced a new API called <code class="xref py py-func docutils literal notranslate"><span class="pre">export_for_training()</span></code>.
It’s still going through hardening, so if you run into any issues, please file
them to Github with the “oncall: export” tag.</p>
<p>In this API, we produce the most generic IR that contains all ATen operators
(including both functional and non-functional) which can be used to train in
eager PyTorch Autograd. This API is intended for eager training use cases such as PT2 Quantization
and will soon be the default IR of torch.export.export. To read further about
the motivation behind this change, please refer to
<a class="reference external" href="https://dev-discuss.pytorch.org/t/why-pytorch-does-not-need-a-new-standardized-operator-set/2206">https://dev-discuss.pytorch.org/t/why-pytorch-does-not-need-a-new-standardized-operator-set/2206</a></p>
<p>When this API is combined with <code class="xref py py-func docutils literal notranslate"><span class="pre">run_decompositions()</span></code>, you should be able to get inference IR with
any desired decomposition behavior.</p>
<p>To show some examples:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">ConvBatchnorm</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="o">-></span> <span class="kc">None</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">conv</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">Conv2d</span><span class="p">(</span><span class="mi">1</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="mi">1</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</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">BatchNorm2d</span><span class="p">(</span><span class="mi">3</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="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">bn</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="p">(</span><span class="n">x</span><span class="p">,)</span>
<span class="n">mod</span> <span class="o">=</span> <span class="n">ConvBatchnorm</span><span class="p">()</span>
<span class="n">inp</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="n">ep_for_training</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">export</span><span class="o">.</span><span class="n">export_for_training</span><span class="p">(</span><span class="n">mod</span><span class="p">,</span> <span class="p">(</span><span class="n">inp</span><span class="p">,))</span>
<span class="nb">print</span><span class="p">(</span><span class="n">ep_for_training</span><span class="p">)</span>
</pre></div>
</div>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">ExportedProgram</span><span class="p">:</span>
<span class="k">class</span> <span class="nc">GraphModule</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="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">p_conv_weight</span><span class="p">:</span> <span class="s2">"f32[3, 1, 1, 1]"</span><span class="p">,</span> <span class="n">p_conv_bias</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">p_bn_weight</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">p_bn_bias</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">b_bn_running_mean</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">b_bn_running_var</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">b_bn_num_batches_tracked</span><span class="p">:</span> <span class="s2">"i64[]"</span><span class="p">,</span> <span class="n">x</span><span class="p">:</span> <span class="s2">"f32[1, 1, 3, 3]"</span><span class="p">):</span>
<span class="n">conv2d</span><span class="p">:</span> <span class="s2">"f32[1, 3, 3, 3]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">conv2d</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">p_conv_weight</span><span class="p">,</span> <span class="n">p_conv_bias</span><span class="p">)</span>
<span class="n">add_</span><span class="p">:</span> <span class="s2">"i64[]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">add_</span><span class="o">.</span><span class="n">Tensor</span><span class="p">(</span><span class="n">b_bn_num_batches_tracked</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">batch_norm</span><span class="p">:</span> <span class="s2">"f32[1, 3, 3, 3]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">batch_norm</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">conv2d</span><span class="p">,</span> <span class="n">p_bn_weight</span><span class="p">,</span> <span class="n">p_bn_bias</span><span class="p">,</span> <span class="n">b_bn_running_mean</span><span class="p">,</span> <span class="n">b_bn_running_var</span><span class="p">,</span> <span class="kc">True</span><span class="p">,</span> <span class="mf">0.1</span><span class="p">,</span> <span class="mf">1e-05</span><span class="p">,</span> <span class="kc">True</span><span class="p">)</span>
<span class="k">return</span> <span class="p">(</span><span class="n">batch_norm</span><span class="p">,)</span>
</pre></div>
</div>
<p>From the above output, you can see that <code class="xref py py-func docutils literal notranslate"><span class="pre">export_for_training()</span></code> produces pretty much the same ExportedProgram
as <code class="xref py py-func docutils literal notranslate"><span class="pre">export()</span></code> except for the operators in the graph. You can see that we captured batch_norm in the most general
form. This op is non-functional and will be lowered to different ops when running inference.</p>
<p>You can also go from this IR to an inference IR via <code class="xref py py-func docutils literal notranslate"><span class="pre">run_decompositions()</span></code> with arbitrary customizations.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Lower to core aten inference IR, but keep conv2d</span>
<span class="n">decomp_table</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">export</span><span class="o">.</span><span class="n">default_decompositions</span><span class="p">()</span>
<span class="k">del</span> <span class="n">decomp_table</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">conv2d</span><span class="o">.</span><span class="n">default</span><span class="p">]</span>
<span class="n">ep_for_inference</span> <span class="o">=</span> <span class="n">ep_for_training</span><span class="o">.</span><span class="n">run_decompositions</span><span class="p">(</span><span class="n">decomp_table</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">ep_for_inference</span><span class="p">)</span>
</pre></div>
</div>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">ExportedProgram</span><span class="p">:</span>
<span class="k">class</span> <span class="nc">GraphModule</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="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">p_conv_weight</span><span class="p">:</span> <span class="s2">"f32[3, 1, 1, 1]"</span><span class="p">,</span> <span class="n">p_conv_bias</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">p_bn_weight</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">p_bn_bias</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">b_bn_running_mean</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">b_bn_running_var</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">b_bn_num_batches_tracked</span><span class="p">:</span> <span class="s2">"i64[]"</span><span class="p">,</span> <span class="n">x</span><span class="p">:</span> <span class="s2">"f32[1, 1, 3, 3]"</span><span class="p">):</span>
<span class="n">conv2d</span><span class="p">:</span> <span class="s2">"f32[1, 3, 3, 3]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">conv2d</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">p_conv_weight</span><span class="p">,</span> <span class="n">p_conv_bias</span><span class="p">)</span>
<span class="n">add</span><span class="p">:</span> <span class="s2">"i64[]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">add</span><span class="o">.</span><span class="n">Tensor</span><span class="p">(</span><span class="n">b_bn_num_batches_tracked</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">_native_batch_norm_legit_functional</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">_native_batch_norm_legit_functional</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">conv2d</span><span class="p">,</span> <span class="n">p_bn_weight</span><span class="p">,</span> <span class="n">p_bn_bias</span><span class="p">,</span> <span class="n">b_bn_running_mean</span><span class="p">,</span> <span class="n">b_bn_running_var</span><span class="p">,</span> <span class="kc">True</span><span class="p">,</span> <span class="mf">0.1</span><span class="p">,</span> <span class="mf">1e-05</span><span class="p">)</span>
<span class="n">getitem</span><span class="p">:</span> <span class="s2">"f32[1, 3, 3, 3]"</span> <span class="o">=</span> <span class="n">_native_batch_norm_legit_functional</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
<span class="n">getitem_3</span><span class="p">:</span> <span class="s2">"f32[3]"</span> <span class="o">=</span> <span class="n">_native_batch_norm_legit_functional</span><span class="p">[</span><span class="mi">3</span><span class="p">]</span>
<span class="n">getitem_4</span><span class="p">:</span> <span class="s2">"f32[3]"</span> <span class="o">=</span> <span class="n">_native_batch_norm_legit_functional</span><span class="p">[</span><span class="mi">4</span><span class="p">]</span>
<span class="k">return</span> <span class="p">(</span><span class="n">getitem_3</span><span class="p">,</span> <span class="n">getitem_4</span><span class="p">,</span> <span class="n">add</span><span class="p">,</span> <span class="n">getitem</span><span class="p">)</span>
</pre></div>
</div>
<p>Here you can see that we kept <code class="docutils literal notranslate"><span class="pre">conv2d</span></code> op in the IR while decomposing the rest. Now the IR is a functional IR
containing core aten operators except for <code class="docutils literal notranslate"><span class="pre">conv2d</span></code>.</p>
<p>You can do even more customization by directly registering your chosen decomposition behaviors.</p>
<p>You can do even more customizations by directly registering custom decomp behaviour</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Lower to core aten inference IR, but customize conv2d</span>
<span class="n">decomp_table</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">export</span><span class="o">.</span><span class="n">default_decompositions</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">my_awesome_custom_conv2d_function</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">weight</span><span class="p">,</span> <span class="n">bias</span><span class="p">,</span> <span class="n">stride</span><span class="o">=</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="n">padding</span><span class="o">=</span><span class="p">[</span><span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span> <span class="n">dilation</span><span class="o">=</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="n">groups</span><span class="o">=</span><span class="mi">1</span><span class="p">):</span>
<span class="k">return</span> <span class="mi">2</span> <span class="o">*</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">convolution</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">weight</span><span class="p">,</span> <span class="n">bias</span><span class="p">,</span> <span class="n">stride</span><span class="p">,</span> <span class="n">padding</span><span class="p">,</span> <span class="n">dilation</span><span class="p">,</span> <span class="kc">False</span><span class="p">,</span> <span class="p">[</span><span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span> <span class="n">groups</span><span class="p">)</span>
<span class="n">decomp_table</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">conv2d</span><span class="o">.</span><span class="n">default</span><span class="p">]</span> <span class="o">=</span> <span class="n">my_awesome_conv2d_function</span>
<span class="n">ep_for_inference</span> <span class="o">=</span> <span class="n">ep_for_training</span><span class="o">.</span><span class="n">run_decompositions</span><span class="p">(</span><span class="n">decomp_table</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">ep_for_inference</span><span class="p">)</span>
</pre></div>
</div>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">ExportedProgram</span><span class="p">:</span>
<span class="k">class</span> <span class="nc">GraphModule</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="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">p_conv_weight</span><span class="p">:</span> <span class="s2">"f32[3, 1, 1, 1]"</span><span class="p">,</span> <span class="n">p_conv_bias</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">p_bn_weight</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">p_bn_bias</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">b_bn_running_mean</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">b_bn_running_var</span><span class="p">:</span> <span class="s2">"f32[3]"</span><span class="p">,</span> <span class="n">b_bn_num_batches_tracked</span><span class="p">:</span> <span class="s2">"i64[]"</span><span class="p">,</span> <span class="n">x</span><span class="p">:</span> <span class="s2">"f32[1, 1, 3, 3]"</span><span class="p">):</span>
<span class="n">convolution</span><span class="p">:</span> <span class="s2">"f32[1, 3, 3, 3]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">convolution</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">p_conv_weight</span><span class="p">,</span> <span class="n">p_conv_bias</span><span class="p">,</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="p">[</span><span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</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="kc">False</span><span class="p">,</span> <span class="p">[</span><span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">mul</span><span class="p">:</span> <span class="s2">"f32[1, 3, 3, 3]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">mul</span><span class="o">.</span><span class="n">Tensor</span><span class="p">(</span><span class="n">convolution</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
<span class="n">add</span><span class="p">:</span> <span class="s2">"i64[]"</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">add</span><span class="o">.</span><span class="n">Tensor</span><span class="p">(</span><span class="n">b_bn_num_batches_tracked</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">_native_batch_norm_legit_functional</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">aten</span><span class="o">.</span><span class="n">_native_batch_norm_legit_functional</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">mul</span><span class="p">,</span> <span class="n">p_bn_weight</span><span class="p">,</span> <span class="n">p_bn_bias</span><span class="p">,</span> <span class="n">b_bn_running_mean</span><span class="p">,</span> <span class="n">b_bn_running_var</span><span class="p">,</span> <span class="kc">True</span><span class="p">,</span> <span class="mf">0.1</span><span class="p">,</span> <span class="mf">1e-05</span><span class="p">)</span>
<span class="n">getitem</span><span class="p">:</span> <span class="s2">"f32[1, 3, 3, 3]"</span> <span class="o">=</span> <span class="n">_native_batch_norm_legit_functional</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
<span class="n">getitem_3</span><span class="p">:</span> <span class="s2">"f32[3]"</span> <span class="o">=</span> <span class="n">_native_batch_norm_legit_functional</span><span class="p">[</span><span class="mi">3</span><span class="p">]</span>
<span class="n">getitem_4</span><span class="p">:</span> <span class="s2">"f32[3]"</span> <span class="o">=</span> <span class="n">_native_batch_norm_legit_functional</span><span class="p">[</span><span class="mi">4</span><span class="p">];</span>
<span class="k">return</span> <span class="p">(</span><span class="n">getitem_3</span><span class="p">,</span> <span class="n">getitem_4</span><span class="p">,</span> <span class="n">add</span><span class="p">,</span> <span class="n">getitem</span><span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="expressing-dynamism">
<h3>Expressing Dynamism<a class="headerlink" href="#expressing-dynamism" title="Permalink to this heading">¶</a></h3>
<p>By default <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> will trace the program assuming all input shapes are
<strong>static</strong>, and specializing the exported program to those dimensions. However,
some dimensions, such as a batch dimension, can be dynamic and vary from run to
run. Such dimensions must be specified by using the
<code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.Dim()</span></code> API to create them and by passing them into
<a class="reference internal" href="#torch.export.export" title="torch.export.export"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.export()</span></code></a> through the <code class="docutils literal notranslate"><span class="pre">dynamic_shapes</span></code> argument. An example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">from</span> <span class="nn">torch.export</span> <span class="kn">import</span> <span class="n">Dim</span><span class="p">,</span> <span class="n">export</span>
<span class="k">class</span> <span class="nc">M</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">branch1</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">Sequential</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">Linear</span><span class="p">(</span><span class="mi">64</span><span class="p">,</span> <span class="mi">32</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">ReLU</span><span class="p">()</span>
<span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">branch2</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">Sequential</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">Linear</span><span class="p">(</span><span class="mi">128</span><span class="p">,</span> <span class="mi">64</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">ReLU</span><span class="p">()</span>
<span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">buffer</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">32</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="n">x1</span><span class="p">,</span> <span class="n">x2</span><span class="p">):</span>
<span class="n">out1</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">branch1</span><span class="p">(</span><span class="n">x1</span><span class="p">)</span>
<span class="n">out2</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">branch2</span><span class="p">(</span><span class="n">x2</span><span class="p">)</span>
<span class="k">return</span> <span class="p">(</span><span class="n">out1</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">buffer</span><span class="p">,</span> <span class="n">out2</span><span class="p">)</span>
<span class="n">example_args</span> <span class="o">=</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">32</span><span class="p">,</span> <span class="mi">64</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">32</span><span class="p">,</span> <span class="mi">128</span><span class="p">))</span>