-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathexport.html
2000 lines (1725 loc) · 216 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 name="robots" content="noindex">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>torch.export — PyTorch 2.2 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="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="torch.export IR Specification" href="export.ir_spec.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>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option 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>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch">
<span class="dropdown-title">ExecuTorch</span>
</a>
</div>
</div>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="active docs-active">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-orange-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/audio/stable/index.html">
<span class="dropdown-title">torchaudio</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/text/stable/index.html">
<span class="dropdown-title">torchtext</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/vision/stable/index.html">
<span class="dropdown-title">torchvision</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torcharrow">
<span class="dropdown-title">torcharrow</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/data">
<span class="dropdown-title">TorchData</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torchrec">
<span class="dropdown-title">TorchRec</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/serve/">
<span class="dropdown-title">TorchServe</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torchx/">
<span class="dropdown-title">TorchX</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/xla">
<span class="dropdown-title">PyTorch on XLA Devices</span>
<p></p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-arrow">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class="dropdown-title">About</span>
<p>Learn about PyTorch’s features and capabilities</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn about the PyTorch foundation</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class="dropdown-title">Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered.</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/community-stories">
<span class="dropdown-title">Community Stories</span>
<p>Learn how our community solves real, everyday machine learning problems with PyTorch.</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class="dropdown-title">Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class="dropdown-title">Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/hub">
<span class="dropdown-title">Models (Beta)</span>
<p>Discover, publish, and reuse pre-trained models</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">GitHub</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>2.2 ▼</a>
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption" role="heading"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="community/build_ci_governance.html">PyTorch Governance | Build + CI</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/design.html">PyTorch Design Philosophy</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/governance.html">PyTorch Governance | Mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/persons_of_interest.html">PyTorch Governance | Maintainers</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Developer Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="notes/amp_examples.html">CUDA Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/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/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/mps.html">MPS backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/numerical_accuracy.html">Numerical accuracy</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">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="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="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.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.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.tensor.parallel.html">torch.distributed.tensor.parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.checkpoint.html">torch.distributed.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="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="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="monitor.html">torch.monitor</a></li>
<li class="toctree-l1"><a class="reference internal" href="signal.html">torch.signal</a></li>
<li class="toctree-l1"><a class="reference internal" href="special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="pipeline.html">Pipeline Parallelism</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="masked.html">torch.masked</a></li>
<li class="toctree-l1"><a class="reference internal" href="nested.html">torch.nested</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="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="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="logging.html">torch._logging</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/data">TorchData</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/torchrec">TorchRec</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="index.html">
Docs
</a> >
</li>
<li>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 an arbitrary Python callable (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>, a function or a method) 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">def</span> <span class="nf">f</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">f</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">arg0_1</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="n">arg1_1</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="c1"># code: a = torch.sin(x)</span>
<span class="n">sin</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">sin</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">arg0_1</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="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">cos</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">arg1_1</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">parameters</span><span class="o">=</span><span class="p">[],</span>
<span class="n">buffers</span><span class="o">=</span><span class="p">[],</span>
<span class="n">user_inputs</span><span class="o">=</span><span class="p">[</span><span class="s1">'arg0_1'</span><span class="p">,</span> <span class="s1">'arg1_1'</span><span class="p">],</span>
<span class="n">user_outputs</span><span class="o">=</span><span class="p">[</span><span class="s1">'add'</span><span class="p">],</span>
<span class="n">inputs_to_parameters</span><span class="o">=</span><span class="p">{},</span>
<span class="n">inputs_to_buffers</span><span class="o">=</span><span class="p">{},</span>
<span class="n">buffers_to_mutate</span><span class="o">=</span><span class="p">{},</span>
<span class="n">backward_signature</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">assertion_dep_token</span><span class="o">=</span><span class="kc">None</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>
<span class="n">Equality</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>Defined Operator Set</strong>: The graph produced contains only a small defined
<a class="reference internal" href="torch.compiler_ir.html#torch-compiler-ir"><span class="std std-ref">Core ATen IR</span></a> opset and registered custom
operators.</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 small defined Core 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">arg0_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">16</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="mi">3</span><span class="p">],</span> <span class="n">arg1_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">16</span><span class="p">],</span> <span class="n">arg2_1</span><span class="p">:</span> <span class="n">f32</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">arg3_1</span><span class="p">:</span> <span class="n">f32</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="c1"># code: a = self.conv(x)</span>
<span class="n">convolution</span><span class="p">:</span> <span class="n">f32</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="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">arg2_1</span><span class="p">,</span> <span class="n">arg0_1</span><span class="p">,</span> <span class="n">arg1_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="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="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="c1"># code: a.add_(constant)</span>
<span class="n">add</span><span class="p">:</span> <span class="n">f32</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="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">convolution</span><span class="p">,</span> <span class="n">arg3_1</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="n">f32</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="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_with_indices</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_with_indices</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="p">);</span>
<span class="n">getitem</span><span class="p">:</span> <span class="n">f32</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">85</span><span class="p">,</span> <span class="mi">85</span><span class="p">]</span> <span class="o">=</span> <span class="n">max_pool2d_with_indices</span><span class="p">[</span><span class="mi">0</span><span class="p">];</span>
<span class="k">return</span> <span class="p">(</span><span class="n">getitem</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">parameters</span><span class="o">=</span><span class="p">[</span><span class="s1">'L__self___conv.weight'</span><span class="p">,</span> <span class="s1">'L__self___conv.bias'</span><span class="p">],</span>
<span class="n">buffers</span><span class="o">=</span><span class="p">[],</span>
<span class="n">user_inputs</span><span class="o">=</span><span class="p">[</span><span class="s1">'arg2_1'</span><span class="p">,</span> <span class="s1">'arg3_1'</span><span class="p">],</span>
<span class="n">user_outputs</span><span class="o">=</span><span class="p">[</span><span class="s1">'getitem'</span><span class="p">],</span>
<span class="n">inputs_to_parameters</span><span class="o">=</span><span class="p">{</span>
<span class="s1">'arg0_1'</span><span class="p">:</span> <span class="s1">'L__self___conv.weight'</span><span class="p">,</span>
<span class="s1">'arg1_1'</span><span class="p">:</span> <span class="s1">'L__self___conv.bias'</span><span class="p">,</span>
<span class="p">},</span>
<span class="n">inputs_to_buffers</span><span class="o">=</span><span class="p">{},</span>
<span class="n">buffers_to_mutate</span><span class="o">=</span><span class="p">{},</span>
<span class="n">backward_signature</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">assertion_dep_token</span><span class="o">=</span><span class="kc">None</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>
<span class="n">Equality</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 in the
<a class="reference internal" href="torch.compiler_ir.html#torch-compiler-ir"><span class="std std-ref">Core ATen IR</span></a> opset 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="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
<a class="reference internal" href="#torch.export.Dim" title="torch.export.Dim"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.Dim()</span></code></a> 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>
<span class="c1"># Create a dynamic batch size</span>
<span class="n">batch</span> <span class="o">=</span> <span class="n">Dim</span><span class="p">(</span><span class="s2">"batch"</span><span class="p">)</span>
<span class="c1"># Specify that the first dimension of each input is that batch size</span>
<span class="n">dynamic_shapes</span> <span class="o">=</span> <span class="p">{</span><span class="s2">"x1"</span><span class="p">:</span> <span class="p">{</span><span class="mi">0</span><span class="p">:</span> <span class="n">batch</span><span class="p">},</span> <span class="s2">"x2"</span><span class="p">:</span> <span class="p">{</span><span class="mi">0</span><span class="p">:</span> <span class="n">batch</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">dynamic_shapes</span><span class="o">=</span><span class="n">dynamic_shapes</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">arg0_1</span><span class="p">:</span> <span class="n">f32</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">arg1_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">32</span><span class="p">],</span> <span class="n">arg2_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">64</span><span class="p">,</span> <span class="mi">128</span><span class="p">],</span> <span class="n">arg3_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">64</span><span class="p">],</span> <span class="n">arg4_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">32</span><span class="p">],</span> <span class="n">arg5_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="n">s0</span><span class="p">,</span> <span class="mi">64</span><span class="p">],</span> <span class="n">arg6_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="n">s0</span><span class="p">,</span> <span class="mi">128</span><span class="p">]):</span>
<span class="c1"># code: out1 = self.branch1(x1)</span>
<span class="n">permute</span><span class="p">:</span> <span class="n">f32</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="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">permute</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">arg0_1</span><span class="p">,</span> <span class="p">[</span><span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">]);</span>
<span class="n">addmm</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="n">s0</span><span class="p">,</span> <span class="mi">32</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">addmm</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">arg1_1</span><span class="p">,</span> <span class="n">arg5_1</span><span class="p">,</span> <span class="n">permute</span><span class="p">);</span>
<span class="n">relu</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="n">s0</span><span class="p">,</span> <span class="mi">32</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">relu</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">addmm</span><span class="p">);</span>
<span class="c1"># code: out2 = self.branch2(x2)</span>
<span class="n">permute_1</span><span class="p">:</span> <span class="n">f32</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="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">permute</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">arg2_1</span><span class="p">,</span> <span class="p">[</span><span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">]);</span>
<span class="n">addmm_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="n">s0</span><span class="p">,</span> <span class="mi">64</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">addmm</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">arg3_1</span><span class="p">,</span> <span class="n">arg6_1</span><span class="p">,</span> <span class="n">permute_1</span><span class="p">);</span>
<span class="n">relu_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="n">s0</span><span class="p">,</span> <span class="mi">64</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">relu</span><span class="o">.</span><span class="n">default</span><span class="p">(</span><span class="n">addmm_1</span><span class="p">);</span> <span class="n">addmm_1</span> <span class="o">=</span> <span class="kc">None</span>
<span class="c1"># code: return (out1 + self.buffer, out2)</span>
<span class="n">add</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="n">s0</span><span class="p">,</span> <span class="mi">32</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">relu</span><span class="p">,</span> <span class="n">arg4_1</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">relu_1</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">parameters</span><span class="o">=</span><span class="p">[</span>
<span class="s1">'branch1.0.weight'</span><span class="p">,</span>
<span class="s1">'branch1.0.bias'</span><span class="p">,</span>
<span class="s1">'branch2.0.weight'</span><span class="p">,</span>
<span class="s1">'branch2.0.bias'</span><span class="p">,</span>
<span class="p">],</span>
<span class="n">buffers</span><span class="o">=</span><span class="p">[</span><span class="s1">'L__self___buffer'</span><span class="p">],</span>
<span class="n">user_inputs</span><span class="o">=</span><span class="p">[</span><span class="s1">'arg5_1'</span><span class="p">,</span> <span class="s1">'arg6_1'</span><span class="p">],</span>
<span class="n">user_outputs</span><span class="o">=</span><span class="p">[</span><span class="s1">'add'</span><span class="p">,</span> <span class="s1">'relu_1'</span><span class="p">],</span>
<span class="n">inputs_to_parameters</span><span class="o">=</span><span class="p">{</span>
<span class="s1">'arg0_1'</span><span class="p">:</span> <span class="s1">'branch1.0.weight'</span><span class="p">,</span>
<span class="s1">'arg1_1'</span><span class="p">:</span> <span class="s1">'branch1.0.bias'</span><span class="p">,</span>
<span class="s1">'arg2_1'</span><span class="p">:</span> <span class="s1">'branch2.0.weight'</span><span class="p">,</span>
<span class="s1">'arg3_1'</span><span class="p">:</span> <span class="s1">'branch2.0.bias'</span><span class="p">,</span>
<span class="p">},</span>
<span class="n">inputs_to_buffers</span><span class="o">=</span><span class="p">{</span><span class="s1">'arg4_1'</span><span class="p">:</span> <span class="s1">'L__self___buffer'</span><span class="p">},</span>
<span class="n">buffers_to_mutate</span><span class="o">=</span><span class="p">{},</span>
<span class="n">backward_signature</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">assertion_dep_token</span><span class="o">=</span><span class="kc">None</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><span class="n">s0</span><span class="p">:</span> <span class="n">RangeConstraint</span><span class="p">(</span><span class="n">min_val</span><span class="o">=</span><span class="mi">2</span><span class="p">,</span> <span class="n">max_val</span><span class="o">=</span><span class="mi">9223372036854775806</span><span class="p">)}</span>
<span class="n">Equality</span> <span class="n">constraints</span><span class="p">:</span> <span class="p">[(</span><span class="n">InputDim</span><span class="p">(</span><span class="n">input_name</span><span class="o">=</span><span class="s1">'arg5_1'</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">),</span> <span class="n">InputDim</span><span class="p">(</span><span class="n">input_name</span><span class="o">=</span><span class="s1">'arg6_1'</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">))]</span>
</pre></div>
</div>
<p>Some additional things to note:</p>
<ul class="simple">
<li><p>Through the <a class="reference internal" href="#torch.export.Dim" title="torch.export.Dim"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.Dim()</span></code></a> API and the <code class="docutils literal notranslate"><span class="pre">dynamic_shapes</span></code> argument, we specified the first
dimension of each input to be dynamic. Looking at the inputs <code class="docutils literal notranslate"><span class="pre">arg5_1</span></code> and
<code class="docutils literal notranslate"><span class="pre">arg6_1</span></code>, they have a symbolic shape of (s0, 64) and (s0, 128), instead of
the (32, 64) and (32, 128) shaped tensors that we passed in as example inputs.
<code class="docutils literal notranslate"><span class="pre">s0</span></code> is a symbol representing that this dimension can be a range
of values.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">exported_program.range_constraints</span></code> describes the ranges of each symbol
appearing in the graph. In this case, we see that <code class="docutils literal notranslate"><span class="pre">s0</span></code> has the range
[2, inf]. For technical reasons that are difficult to explain here, they are
assumed to be not 0 or 1. This is not a bug, and does not necessarily mean
that the exported program will not work for dimensions 0 or 1. See
<a class="reference external" href="https://docs.google.com/document/d/16VPOa3d-Liikf48teAOmxLc92rgvJdfosIy-yoT38Io/edit?fbclid=IwAR3HNwmmexcitV0pbZm_x1a4ykdXZ9th_eJWK-3hBtVgKnrkmemz6Pm5jRQ#heading=h.ez923tomjvyk">The 0/1 Specialization Problem</a>
for an in-depth discussion of this topic.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">exported_program.equality_constraints</span></code> describes which dimensions are
required to be equal. Since we specified in the constraints that the first
dimension of each argument is equivalent,
(<code class="docutils literal notranslate"><span class="pre">dynamic_dim(example_args[0],</span> <span class="pre">0)</span> <span class="pre">==</span> <span class="pre">dynamic_dim(example_args[1],</span> <span class="pre">0)</span></code>),
we see in the equality constraints the tuple specifying that <code class="docutils literal notranslate"><span class="pre">arg5_1</span></code>
dimension 0 and <code class="docutils literal notranslate"><span class="pre">arg6_1</span></code> dimension 0 are equal.</p></li>
</ul>
<p>(A legacy mechanism for specifying dynamic shapes
involves marking and constraining dynamic dimensions with the
<a class="reference internal" href="#torch.export.dynamic_dim" title="torch.export.dynamic_dim"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.dynamic_dim()</span></code></a> API and 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">constraints</span></code> argument. That mechanism is now <strong>deprecated</strong> and will
not be supported in the future.)</p>
</div>
<div class="section" id="serialization">
<h3>Serialization<a class="headerlink" href="#serialization" title="Permalink to this heading">¶</a></h3>
<p>To save the <code class="docutils literal notranslate"><span class="pre">ExportedProgram</span></code>, users can use the <a class="reference internal" href="#torch.export.save" title="torch.export.save"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.save()</span></code></a> and
<a class="reference internal" href="#torch.export.load" title="torch.export.load"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.load()</span></code></a> APIs. A convention is to save the <code class="docutils literal notranslate"><span class="pre">ExportedProgram</span></code>
using a <code class="docutils literal notranslate"><span class="pre">.pt2</span></code> file extension.</p>
<p>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">import</span> <span class="nn">io</span>
<span class="k">class</span> <span class="nc">MyModule</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="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">return</span> <span class="n">x</span> <span class="o">+</span> <span class="mi">10</span>
<span class="n">exported_program</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</span><span class="p">(</span><span class="n">MyModule</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">5</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">save</span><span class="p">(</span><span class="n">exported_program</span><span class="p">,</span> <span class="s1">'exported_program.pt2'</span><span class="p">)</span>
<span class="n">saved_exported_program</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">load</span><span class="p">(</span><span class="s1">'exported_program.pt2'</span><span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="specialization">
<h3>Specialization<a class="headerlink" href="#specialization" title="Permalink to this heading">¶</a></h3>
<div class="section" id="input-shapes">
<h4>Input shapes<a class="headerlink" href="#input-shapes" title="Permalink to this heading">¶</a></h4>
<p>As mentioned before, by default, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> will trace the program
specializing on the input tensors’ shapes, unless a dimension is specified as
dynamic via the <a class="reference internal" href="#torch.export.dynamic_dim" title="torch.export.dynamic_dim"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.dynamic_dim()</span></code></a> API. This means that if there
exists shape-dependent control flow, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> will specialize on the
branch that is being taken with the given sample inputs. For 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="k">def</span> <span class="nf">fn</span><span class="p">(</span><span class="n">x</span><span class="p">):</span>
<span class="k">if</span> <span class="n">x</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">></span> <span class="mi">5</span><span class="p">:</span>
<span class="k">return</span> <span class="n">x</span> <span class="o">+</span> <span class="mi">1</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="n">x</span> <span class="o">-</span> <span class="mi">1</span>
<span class="n">example_inputs</span> <span class="o">=</span> <span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">2</span><span class="p">),)</span>
<span class="n">exported_program</span> <span class="o">=</span> <span class="n">export</span><span class="p">(</span><span class="n">fn</span><span class="p">,</span> <span class="n">example_inputs</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">arg0_1</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">2</span><span class="p">]):</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">2</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">arg0_1</span><span class="p">,</span> <span class="mi">1</span><span class="p">);</span>
<span class="k">return</span> <span class="p">(</span><span class="n">add</span><span class="p">,)</span>
</pre></div>
</div>
<p>The conditional of (<code class="docutils literal notranslate"><span class="pre">x.shape[0]</span> <span class="pre">></span> <span class="pre">5</span></code>) does not appear in the
<code class="docutils literal notranslate"><span class="pre">ExportedProgram</span></code> because the example inputs have the static
shape of (10, 2). Since <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> specializes on the inputs’ static
shapes, the else branch (<code class="docutils literal notranslate"><span class="pre">x</span> <span class="pre">-</span> <span class="pre">1</span></code>) will never be reached. To preserve the dynamic
branching behavior based on the shape of a tensor in the traced graph,
<a class="reference internal" href="#torch.export.dynamic_dim" title="torch.export.dynamic_dim"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.export.dynamic_dim()</span></code></a> will need to be used to specify the dimension
of the input tensor (<code class="docutils literal notranslate"><span class="pre">x.shape[0]</span></code>) to be dynamic, and the source code will
need to be <a class="reference internal" href="#data-shape-dependent-control-flow"><span class="std std-ref">rewritten</span></a>.</p>
</div>
<div class="section" id="non-tensor-inputs">
<h4>Non-tensor inputs<a class="headerlink" href="#non-tensor-inputs" title="Permalink to this heading">¶</a></h4>
<p><code class="docutils literal notranslate"><span class="pre">torch.export</span></code> also specializes the traced graph based on the values of inputs
that are not <code class="docutils literal notranslate"><span class="pre">torch.Tensor</span></code>, such as <code class="docutils literal notranslate"><span class="pre">int</span></code>, <code class="docutils literal notranslate"><span class="pre">float</span></code>, <code class="docutils literal notranslate"><span class="pre">bool</span></code>, and <code class="docutils literal notranslate"><span class="pre">str</span></code>.
However, we will likely change this in the near future to not specialize on
inputs of primitive types.</p>
<p>For 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="k">def</span> <span class="nf">fn</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">const</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span> <span class="n">times</span><span class="p">:</span> <span class="nb">int</span><span class="p">):</span>
<span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">times</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">x</span> <span class="o">+</span> <span class="n">const</span>
<span class="k">return</span> <span class="n">x</span>
<span class="n">example_inputs</span> <span class="o">=</span> <span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="n">exported_program</span> <span class="o">=</span> <span class="n">export</span><span class="p">(</span><span class="n">fn</span><span class="p">,</span> <span class="n">example_inputs</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">arg0_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">2</span><span class="p">,</span> <span class="mi">2</span><span class="p">],</span> <span class="n">arg1_1</span><span class="p">,</span> <span class="n">arg2_1</span><span class="p">):</span>
<span class="n">add</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">2</span><span class="p">,</span> <span class="mi">2</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">arg0_1</span><span class="p">,</span> <span class="mi">1</span><span class="p">);</span>
<span class="n">add_1</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">2</span><span class="p">,</span> <span class="mi">2</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">add</span><span class="p">,</span> <span class="mi">1</span><span class="p">);</span>
<span class="n">add_2</span><span class="p">:</span> <span class="n">f32</span><span class="p">[</span><span class="mi">2</span><span class="p">,</span> <span class="mi">2</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">add_1</span><span class="p">,</span> <span class="mi">1</span><span class="p">);</span>
<span class="k">return</span> <span class="p">(</span><span class="n">add_2</span><span class="p">,)</span>
</pre></div>
</div>
<p>Because integers are specialized, the <code class="docutils literal notranslate"><span class="pre">torch.ops.aten.add.Tensor</span></code> operations
are all computed with the inlined constant <code class="docutils literal notranslate"><span class="pre">1</span></code>, rather than <code class="docutils literal notranslate"><span class="pre">arg1_1</span></code>.
Additionally, the <code class="docutils literal notranslate"><span class="pre">times</span></code> iterator used in the <code class="docutils literal notranslate"><span class="pre">for</span></code> loop is also “inlined”
in the graph through the 3 repeated <code class="docutils literal notranslate"><span class="pre">torch.ops.aten.add.Tensor</span></code> calls, and the
input <code class="docutils literal notranslate"><span class="pre">arg2_1</span></code> is never used.</p>
</div>
</div>
</div>
<div class="section" id="limitations-of-torch-export">
<h2>Limitations of torch.export<a class="headerlink" href="#limitations-of-torch-export" title="Permalink to this heading">¶</a></h2>
<div class="section" id="graph-breaks">
<h3>Graph Breaks<a class="headerlink" href="#graph-breaks" title="Permalink to this heading">¶</a></h3>
<p>As <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> is a one-shot process for capturing a computation graph from
a PyTorch program, it might ultimately run into untraceable parts of programs as
it is nearly impossible to support tracing all PyTorch and Python features. In
the case of <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code>, an unsupported operation will cause a “graph
break” and the unsupported operation will be run with default Python evaluation.
In contrast, <code class="docutils literal notranslate"><span class="pre">torch.export</span></code> will require users to provide additional
information or rewrite parts of their code to make it traceable. As the
tracing is based on TorchDynamo, which evaluates at the Python
bytecode level, there will be significantly fewer rewrites required compared to
previous tracing frameworks.</p>
<p>When a graph break is encountered, <a class="reference internal" href="generated/exportdb/index.html#torch-export-db"><span class="std std-ref">ExportDB</span></a> is a great
resource for learning about the kinds of programs that are supported and
unsupported, along with ways to rewrite programs to make them traceable.</p>
</div>
<div class="section" id="data-shape-dependent-control-flow">
<span id="id2"></span><h3>Data/Shape-Dependent Control Flow<a class="headerlink" href="#data-shape-dependent-control-flow" title="Permalink to this heading">¶</a></h3>
<p>Graph breaks can also be encountered on data-dependent control flow (<code class="docutils literal notranslate"><span class="pre">if</span>
<span class="pre">x.shape[0]</span> <span class="pre">></span> <span class="pre">2</span></code>) when shapes are not being specialized, as a tracing compiler cannot
possibly deal with without generating code for a combinatorially exploding
number of paths. In such cases, users will need to rewrite their code using
special control flow operators. Currently, we support <a class="reference internal" href="cond.html#cond"><span class="std std-ref">torch.cond</span></a>
to express if-else like control flow (more coming soon!).</p>
</div>
<div class="section" id="missing-meta-kernels-for-operators">
<h3>Missing Meta Kernels for Operators<a class="headerlink" href="#missing-meta-kernels-for-operators" title="Permalink to this heading">¶</a></h3>
<p>When tracing, a META implementation (or “meta kernel”) is required for all
operators. This is used to reason about the input/output shapes for this
operator.</p>
<p>To register a meta kernel for a C++ Custom Operator, please refer to
<a class="reference external" href="https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU/edit#heading=h.ahugy69p2jmz">this documentation</a>.</p>
<p>The official API for registering custom meta kernels for custom ops implemented
in python is currently undergoing development. While the final API is being
refined, you can refer to the documentation
<a class="reference external" href="https://docs.google.com/document/d/1GgvOe7C8_NVOMLOCwDaYV1mXXyHMXY7ExoewHqooxrs/edit#heading=h.64r4npvq0w0">here</a>.</p>
<p>In the unfortunate case where your model uses an ATen operator that is does not
have a meta kernel implementation yet, please file an issue.</p>
</div>
</div>
<div class="section" id="read-more">
<h2>Read More<a class="headerlink" href="#read-more" title="Permalink to this heading">¶</a></h2>
<div class="toctree-wrapper compound">
<p class="caption" role="heading"><span class="caption-text">Additional Links for Export Users</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="export.ir_spec.html">torch.export IR Specification</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.compiler_transformations.html">Writing Graph Transformations on ATen IR</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.compiler_ir.html">IRs</a></li>
<li class="toctree-l1"><a class="reference internal" href="generated/exportdb/index.html">ExportDB</a></li>
<li class="toctree-l1"><a class="reference internal" href="cond.html">Control Flow - Cond</a></li>
</ul>
</div>
<div class="toctree-wrapper compound">
<p class="caption" role="heading"><span class="caption-text">Deep Dive for PyTorch Developers</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torch.compiler_deepdive.html">TorchDynamo Deep Dive</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.compiler_dynamic_shapes.html">Dynamic shapes</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.compiler_fake_tensor.html">Fake tensor</a></li>
</ul>
</div>
</div>
<div class="section" id="module-torch.export">
<span id="api-reference"></span><h2>API Reference<a class="headerlink" href="#module-torch.export" title="Permalink to this heading">¶</a></h2>
<dl class="py function">
<dt class="sig sig-object py" id="torch.export.export">
<span class="sig-prename descclassname"><span class="pre">torch.export.</span></span><span class="sig-name descname"><span class="pre">export</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">f</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">kwargs</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">constraints</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">dynamic_shapes</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">strict</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">preserve_module_call_signature</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">()</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/export.html#export"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.export.export" title="Permalink to this definition">¶</a></dt>
<dd><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">export()</span></code></a> takes an arbitrary Python callable (an nn.Module, a function or
a method) along with example inputs, 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 inputs or serialized. The
traced graph (1) produces normalized operators in the functional ATen operator set
(as well as any user-specified custom operators), (2) has eliminated all Python control
flow and data structures (with certain exceptions), and (3) records the set of
shape constraints needed to show that this normalization and control-flow elimination
is sound for future inputs.</p>
<p><strong>Soundness Guarantee</strong></p>
<p>While tracing, <a class="reference internal" href="#torch.export.export" title="torch.export.export"><code class="xref py py-func docutils literal notranslate"><span class="pre">export()</span></code></a> takes note of shape-related assumptions
made by the user program and the underlying PyTorch operator kernels.
The output <a class="reference internal" href="#torch.export.ExportedProgram" title="torch.export.ExportedProgram"><code class="xref py py-class docutils literal notranslate"><span class="pre">ExportedProgram</span></code></a> is considered valid only when these
assumptions hold true.</p>
<p>Tracing makes assumptions on the shapes (not values) of input tensors.
Such assumptions must be validated at graph capture time for <a class="reference internal" href="#torch.export.export" title="torch.export.export"><code class="xref py py-func docutils literal notranslate"><span class="pre">export()</span></code></a>
to succeed. Specifically:</p>
<ul class="simple">
<li><p>Assumptions on static shapes of input tensors are automatically validated without additional effort.</p></li>
<li><p>Assumptions on dynamic shape of input tensors require explicit specification
by using the <a class="reference internal" href="#torch.export.Dim" title="torch.export.Dim"><code class="xref py py-func docutils literal notranslate"><span class="pre">Dim()</span></code></a> API to construct dynamic dimensions and by associating
them with example inputs through the <code class="docutils literal notranslate"><span class="pre">dynamic_shapes</span></code> argument.</p></li>
</ul>
<p>If any assumption can not be validated, a fatal error will be raised. When that happens,
the error message will include suggested fixes to the specification that are needed
to validate the assumptions. For example <a class="reference internal" href="#torch.export.export" title="torch.export.export"><code class="xref py py-func docutils literal notranslate"><span class="pre">export()</span></code></a> might suggest the
following fix to the definition of a dynamic dimension <code class="docutils literal notranslate"><span class="pre">dim0_x</span></code>, say appearing in the
shape associated with input <code class="docutils literal notranslate"><span class="pre">x</span></code>, that was previously defined as <code class="docutils literal notranslate"><span class="pre">Dim("dim0_x")</span></code>:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">dim</span> <span class="o">=</span> <span class="n">Dim</span><span class="p">(</span><span class="s2">"dim0_x"</span><span class="p">,</span> <span class="nb">max</span><span class="o">=</span><span class="mi">5</span><span class="p">)</span>
</pre></div>
</div>
<p>This example means the generated code requires dimension 0 of input <code class="docutils literal notranslate"><span class="pre">x</span></code> to be less
than or equal to 5 to be valid. You can inspect the suggested fixes to dynamic dimension
definitions and then copy them verbatim into your code without needing to change the
<code class="docutils literal notranslate"><span class="pre">dynamic_shapes</span></code> argument to your <a class="reference internal" href="#torch.export.export" title="torch.export.export"><code class="xref py py-func docutils literal notranslate"><span class="pre">export()</span></code></a> call.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>f</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Callable" title="(in Python v3.12)"><em>Callable</em></a>) – The callable to trace.</p></li>
<li><p><strong>args</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Tuple" title="(in Python v3.12)"><em>Tuple</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Any" title="(in Python v3.12)"><em>Any</em></a><em>, </em><em>...</em><em>]</em>) – Example positional inputs.</p></li>
<li><p><strong>kwargs</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Optional" title="(in Python v3.12)"><em>Optional</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Dict" title="(in Python v3.12)"><em>Dict</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.12)"><em>str</em></a><em>, </em><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Any" title="(in Python v3.12)"><em>Any</em></a><em>]</em><em>]</em>) – Optional example keyword inputs.</p></li>
<li><p><strong>constraints</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Optional" title="(in Python v3.12)"><em>Optional</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.List" title="(in Python v3.12)"><em>List</em></a><em>[</em><a class="reference internal" href="#torch.export.Constraint" title="torch.export.Constraint"><em>Constraint</em></a><em>]</em><em>]</em>) – [DEPRECATED: use <code class="docutils literal notranslate"><span class="pre">dynamic_shapes</span></code> instead, see below]
An optional list of constraints on the dynamic arguments
that specify their possible range of shapes. By default, shapes of