-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtorch.compiler_fake_tensor.html
970 lines (775 loc) · 58.1 KB
/
torch.compiler_fake_tensor.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
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Fake tensor — PyTorch 2.2 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/torch.compiler_fake_tensor.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/katex@0.10.0-beta/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.13.11/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="Distributed communication package - torch.distributed" href="distributed.html" />
<link rel="prev" title="Dynamic shapes" href="torch.compiler_dynamic_shapes.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/katex@0.10.0/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/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="reference internal" href="export.html">torch.export</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.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 current"><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><a href="export.html">torch.export</a> ></li>
<li>Fake tensor</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/torch.compiler_fake_tensor.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="fake-tensor">
<h1>Fake tensor<a class="headerlink" href="#fake-tensor" title="Permalink to this heading">¶</a></h1>
<p>Code: <a class="reference external" href="https://github.com/pytorch/pytorch/blob/db4572dbf18f1cf50cf662547e272d3117063747/torch/_subclasses/fake_tensor.py">fake_tensor.py</a></p>
<div class="section" id="motivation">
<h2>Motivation<a class="headerlink" href="#motivation" title="Permalink to this heading">¶</a></h2>
<p>When doing Dynamo symbolic evaluation and compiler passes, we often want to be able to run tensor operations to understand what output sizes/dtypes/devices are, without actually running those operations (or trashing preexisting tensors), which would be slower (if you’re doing a lot of compute) and take a lot of memory (it’s bad if your compiler needs to use GPU memory while you are compiling the program). A fake tensor is like a real tensor in all respects, except that it doesn’t actually have any data. For example, when we do Dynamo tracing, we need to trace through user Tensor code and answer questions about intermediates (e.g., if a user does a conditional on an intermediate tensor). Without fake tensor, we would not have accurate information for these queries.</p>
<p>Similarly, suppose you want to store metadata for a tensor, e.g., on an FX IR node (meta[‘val’]). You can instead store a fake tensor directly on the node, which will give you all the metadata you need for the tensor, including subtle stuff that you probably wouldn’t have handled (e.g., aliasing relationships).</p>
</div>
<div class="section" id="related-work">
<h2>Related work<a class="headerlink" href="#related-work" title="Permalink to this heading">¶</a></h2>
<ul class="simple">
<li><p>A meta tensor is a tensor with device=’meta’. This is actually a lot of what you want for fake tensor, but meta tensors don’t model devices, and sometimes stride behavior varies depending on your device, so fake tensors really can get a lot more accurate info this way. Also, meta tensors are “global” (they exist on their own, similar to how a CPU/CUDA tensor exist on their own), whereas fake tensors are scoped to a FakeTensorMode.</p></li>
<li><p>A tensor subclass lets you subclass torch.Tensor and customize their behavior. Fake tensors are implemented as a tensor subclass; that means almost all of its implementation lives in Python! For more simple examples of tensor subclasses check out <a class="reference external" href="https://github.com/albanD/subclass_zoo/">subclass_zoo</a>.</p></li>
<li><p>Dynamic shapes allow you to create tensors with symbolic sizes rather than only concrete sizes, and propagate these sizes symbolically through operations. Dynamic shapes maintain state in a ShapeEnv, which is always associated with a FakeTensorMode (so fake tensors also are responsible for managing symbolic sizes.) In general, whenever we compile a subgraph with PT2, there is a tracing context associated with this compilation, which contains, among other things, a FakeTensorMode and (possibly) a ShapeEnv.</p></li>
</ul>
</div>
<div class="section" id="overall-architecture">
<h2>Overall architecture<a class="headerlink" href="#overall-architecture" title="Permalink to this heading">¶</a></h2>
<p>All fake tensors are associated with a FakeTensorMode. Because fake tensor’s primary use case is to do analysis on real tensors, the general workflow is you have a bunch of real tensors, you allocate a FakeTensorMode, and then you use from_real_tensor to convert all those real tensors into fake tensors, and then you do things to the fake tensors. In particular, the FakeTensorMode maintains a memo table persistently mapping tensors (and storages) to the same storages. If you fakeify the same tensor multiple times, you will get the same fake tensor; if you fakeify two tensors which alias each other, you will get two fake tensors which alias the same fake storage. FakeTensors are tensor subclasses, so if you do operations on them, you’ll automatically get a fake tensor, but in general you will want to do operations on fake tensors (e.g., if you’re running an FX pass) with the FakeTensorMode active; what a tensor operation will do is automatically turn on the fake tensor mode and try again.</p>
<p>A fake tensor is represented as a __torch_dispatch__ tensor subclass of a meta tensor. This means under the hood, fake tensors are meta device tensors; they then use extra extensibility hooks, specifically dispatch_device, to lie about what the actual device of the tensor is. This was one of the more error-prone parts of fake tensors in the early days: sometimes, fake tensors were too good at lying about being CPU/CUDA whatever, and you’d end up with a CPU kernel getting called with a fake tensor trying to dereference the data pointer, which obviously won’t work. If you are segfaulting in fake tensor code, this is the first thing you should check: is the C++ backtrace in a CPU kernel (unexpected!) or a meta kernel (expected!) A meta kernel is like a real kernel, but all it does is allocate the outputs, it doesn’t do any data compute.</p>
<p>A tensor subclass has to define how to implement various operations. Here is the general fake tensor recipe:</p>
<ul class="simple">
<li><p>Run the meta kernel on the input fake tensors, reinterpreting them as meta tensors. This is done via a magic context manager in_kernel_invocation_manager which instructs all of PyTorch to view fake tensors as their underlying meta tensors, rather than “unwrapping” fake tensors into meta tensors (a fake tensor is a meta tensor). Fake tensors are represented this way to avoid having to keep two sets of metadata in sync (the meta tensor’s metadata, and the fake tensor’s metadata); the “is a” relationship ensures there is only one canonical copy of metadata.</p></li>
<li><p>If you’re a factory function, you’ll instead call the underlying factory function with device=’meta’.</p></li>
<li><p>Convert the resulting meta tensor into a fake tensor, computing what the output device of the tensor should be (this is usually trivial, but sometimes it is not, e.g., cpu scalar promotion, or device-converting operations.)</p></li>
</ul>
</div>
<div class="section" id="api-the-important-bits">
<h2>API: the important bits<a class="headerlink" href="#api-the-important-bits" title="Permalink to this heading">¶</a></h2>
<p>Non-PT2 usage (check out test/test_fake_tensor.py for more examples):</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Create a fake mode</span>
<span class="kn">from</span> <span class="nn">torch._subclasses.fake_tensor</span> <span class="kn">import</span> <span class="n">FakeTensorMode</span>
<span class="n">fake_mode</span> <span class="o">=</span> <span class="n">FakeTensorMode</span><span class="p">()</span>
<span class="c1"># Fakeify some real tensors</span>
<span class="n">fake_x</span> <span class="o">=</span> <span class="n">fake_mode</span><span class="o">.</span><span class="n">from_real_tensor</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">with</span> <span class="n">fake_mode</span><span class="p">:</span>
<span class="c1"># Do some operations on the fake tensors</span>
<span class="n">fake_y</span> <span class="o">=</span> <span class="n">fake_x</span> <span class="o">*</span> <span class="mi">2</span>
<span class="c1"># Factory operations automatically get fakeified in the context manager</span>
<span class="n">fake_z</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">empty</span><span class="p">(</span><span class="mi">20</span><span class="p">)</span>
</pre></div>
</div>
<p>Q: Why do you have real tensors as inputs?</p>
<p>A: In a PT2 context, this is because you typically are compiling just-in-time, so for all the inputs to a graph you’re compiling, you already have the “real” inputs, because you’re compiling while you’re executing the program.</p>
<p>PT2 pre-AOTAutograd usage (this is unusual, you probably don’t want to do this):</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Fake mode is not enabled!</span>
<span class="kn">from</span> <span class="nn">torch._guards</span> <span class="kn">import</span> <span class="n">detect_fake_mode</span>
<span class="n">fake_mode</span> <span class="o">=</span> <span class="n">detect_fake_mode</span><span class="p">(</span><span class="n">args</span><span class="p">)</span>
<span class="n">fake_args</span> <span class="o">=</span> <span class="p">[</span><span class="n">fake_mode</span><span class="o">.</span><span class="n">from_real_tensor</span><span class="p">(</span><span class="n">arg</span><span class="p">)</span> <span class="k">for</span> <span class="n">arg</span> <span class="ow">in</span> <span class="n">args</span><span class="p">]</span>
<span class="k">with</span> <span class="n">fake_mode</span><span class="p">:</span>
<span class="o">...</span> <span class="n">do</span> <span class="n">stuff</span> <span class="k">with</span> <span class="n">the</span> <span class="n">fake</span> <span class="n">args</span><span class="p">,</span> <span class="k">if</span> <span class="n">needed</span> <span class="o">...</span>
</pre></div>
</div>
<p>detect_fake_mode will search a number of locations to try to find “the” fake tensor mode associated with the lifecycle. Typically it will be pulled off of the tracing context.</p>
<p>PT2 post-AOTAutograd usage:</p>
<p># Fake mode is enabled! example_inputs is typically fake already
# TODO: we probably want to change this
# Still do this to access fake mode
fake_mode = detect_fake_mode(example_inputs)
# But in general you don’t have to turn it on</p>
<p>Other useful stuff:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">torch.fx.experimental.proxy_tensor</span> <span class="kn">import</span> <span class="n">maybe_disable_fake_tensor_mode</span>
<span class="k">with</span> <span class="n">maybe_disable_fake_tensor_mode</span><span class="p">():</span>
<span class="c1"># fake mode is disabled here, you can do real tensor compute</span>
</pre></div>
</div>
<p>When might you want to disable fake tensor mode? Usually you don’t want to do this. One niche case where we’ve found it useful is to implement constant propagation on fake tensors: in this case, we need to do some actual tensor computation even though we’re in a fake tensor mode.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">FakeTensorProp</span>
<span class="kn">from</span> <span class="nn">torch.fx.passes.fake_tensor_prop</span>
<span class="n">gm</span><span class="p">:</span> <span class="n">GraphModule</span>
<span class="n">real_inputs</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">Tensor</span><span class="p">]</span>
<span class="n">FakeTensorProp</span><span class="p">(</span><span class="n">gm</span><span class="p">)</span><span class="o">.</span><span class="n">propagate</span><span class="p">(</span><span class="o">*</span><span class="n">real_inputs</span><span class="p">)</span>
<span class="c1"># This will populate meta['val'] on all the FX nodes with a fake tensor</span>
<span class="c1"># or if you have a preexisting fake mode, you should use it</span>
<span class="n">FakeTensorProp</span><span class="p">(</span><span class="n">gm</span><span class="p">,</span> <span class="n">mode</span><span class="o">=</span><span class="n">fake_mode</span><span class="p">)</span><span class="o">.</span><span class="n">propagate</span><span class="p">(</span><span class="o">*</span><span class="n">real_inputs</span><span class="p">)</span>
<span class="c1"># There is also propagate_dont_convert_inputs if your inputs are already fake</span>
<span class="n">fake_inputs</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">FakeTensor</span><span class="p">]</span>
<span class="n">FakeTensorProp</span><span class="p">(</span><span class="n">gm</span><span class="p">,</span> <span class="n">mode</span><span class="o">=</span><span class="n">fake_mode</span><span class="p">)</span><span class="o">.</span><span class="n">propagate_dont_convert_inputs</span><span class="p">(</span><span class="o">*</span><span class="n">fake_inputs</span><span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="details">
<h2>Details<a class="headerlink" href="#details" title="Permalink to this heading">¶</a></h2>
<p>Auto-convert or not?
Originally, FakeTensorMode would not automatically fakeify real tensors if you tried to do compute on them inside a FakeTensorMode region. The motivation behind this was to prevent the following footgun:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">with</span> <span class="n">FakeTensorMode</span><span class="p">():</span>
<span class="n">real_tensor</span><span class="o">.</span><span class="n">t_</span><span class="p">()</span>
</pre></div>
</div>
<p>What should this code do? It would be surprising if we actually modified the metadata on the real tensor. But at the same time, there isn’t any obvious opportunity to create a FakeTensor. So we conservatively decided to make this raise an error: “Invoking operators with non-Fake Tensor inputs in FakeTensorMode is not yet supported. Please convert all Tensors to FakeTensors first.”</p>
<p>This error is pretty annoying in practice. For example, suppose you have a real nn.Module and you want to feed fake tensors through it. You need to somehow fakeify the nn.Module. This motivated FakeCopyMode.</p>
<p>Eventually, we gave up and added automatic fakeification. However, this is still not yet enabled by default in many uses of FakeTensorMode.</p>
<p>Metadata mutation on fake tensor
If you have a fake tensor, and you t_() it, the metadata on the fake tensor changes. This is reasonable on its face, but sometimes you want to also store fake tensors as metadata on FX nodes; mutating a fake tensor is bad because this will invalidate old metadata!</p>
<p>In fact, there is a fundamental tension here, which is that fake tensors maintain extremely accurate metadata about tensors, up to and including object identity. If object metadata changes over time in an FX graph, there is not actually any way to represent this change over time. Most of the time, our serious FX analyses are done on functionalized graphs, which don’t have this, but occasionally you need to do an analysis on a non-functionalized graph. Maybe it was a mistake to put fake tensor in meta[‘val’]</p>
</div>
<div class="section" id="about-the-tensor-subclass">
<h2>About the tensor subclass<a class="headerlink" href="#about-the-tensor-subclass" title="Permalink to this heading">¶</a></h2>
<p>Fake tensor uses both a subclass and a mode tensor subclass pattern, where FakeTensor.__torch_dispatch__ enables the FakeTensorMode associated with the fake tensor, and then redispatches (relying on FakeTensorMode to do the heavy lifting). If fake tensor operations get a subclass argument it doesn’t recognize, it will return NotImplemented, giving the other subclass a chance to run first (hopefully desugaring into plain tensor operations), before it tries again. This can cause infinite loops.</p>
</div>
<div class="section" id="how-is-each-individual-operator-implemented">
<h2>How is each individual operator implemented?<a class="headerlink" href="#how-is-each-individual-operator-implemented" title="Permalink to this heading">¶</a></h2>
<p>Unfortunately, there is a pretty complicated set of places where any given operator may be implemented. Some important cases to know about:</p>
<ul class="simple">
<li><p>Tensor subclasses support limited constant propagation if the number of elements is very small (this helps deal with some cases where we immediately call item() on such tensors.)</p></li>
<li><p>We have some fastpath implementations for certain operators, which are done entirely in fake tensor, for performance reasons.</p></li>
<li><p>If you use @custom_op to generate a custom tensor, these will register impl_abstract directly to fake tensor.</p></li>
<li><p>Fake tensor itself has some hardcoded special cases for device-converting operations.</p></li>
<li><p>If there is no meta implementation nor any decomposition, we will generate real zero-filled tensors and attempt to run the operator directly to find out what the results will be. This can cause segfaults if the operator attempts to do indexing with data, so we don’t turn this on by default for custom ops.</p></li>
</ul>
</div>
<div class="section" id="how-does-the-converter-work">
<h2>How does the converter work?<a class="headerlink" href="#how-does-the-converter-work" title="Permalink to this heading">¶</a></h2>
<p>Because fake tensors are used in situations that are very sensitive to the exact properties of a tensor, fake tensors do conversion very carefully, preserving leaf-ness, requires_grad’ness, aliasing, and a whole host of other properties. The bulk of the heavy lifting is in MetaConverter.</p>
</div>
<div class="section" id="performance-characteristics">
<h2>Performance characteristics<a class="headerlink" href="#performance-characteristics" title="Permalink to this heading">¶</a></h2>
<p>You would think fake tensors are fast because they don’t do any tensor compute. But at small tensor sizes we are actually entirely overhead bound, and, well, fake tensor is in Python, and we often do a LOT of work to do a single tensor operation (because they are implemented as decompositions). So fake tensors are actually pretty slow in practice, especially when symbolic shapes are involved. There are two important fastpaths we currently have in fake tensor that make a big difference in practice:</p>
<ul class="simple">
<li><p>Pointwise ops don’t go through PrimTorch decomps, instead we’ve hand-coded their propagation rule.</p></li>
<li><p>If possible, we should.</p></li>
</ul>
</div>
<div class="section" id="fake-tensor-of-fake-tensor">
<h2>Fake tensor of fake tensor?<a class="headerlink" href="#fake-tensor-of-fake-tensor" title="Permalink to this heading">¶</a></h2>
<p>There is interest in sending fake tensors as user inputs into the PT2 stack, which would imply we would need to be able to create a fake tensor of a fake tensor. This isn’t really supported right now, but maybe it would not be too difficult to do.</p>
</div>
<div class="section" id="interaction-with-dynamic-shapes">
<h2>Interaction with dynamic shapes<a class="headerlink" href="#interaction-with-dynamic-shapes" title="Permalink to this heading">¶</a></h2>
<p>Every FakeTensorMode contains a ShapeEnv, which tracks all symbolic shapes information. Their lifetimes are typically tied: they live and die together.</p>
<p>Because FakeTensorMode has a ShapeEnv (but meta implementations do not), meta functions that are data-dependent and require allocating an unbacked SymInt live in fake tensor. Fake tensor also takes care of memoizing unbacked SymInts, so that, e.g., if you call nonzero() on the same fake tensor twice, you get the same symbolic size.</p>
</div>
<div class="section" id="other-resources">
<h2>Other resources<a class="headerlink" href="#other-resources" title="Permalink to this heading">¶</a></h2>
<p><a class="reference external" href="https://colab.research.google.com/drive/1zjAisRrc8R6uixKsrs1DRm3lwz5MWN68">Colab Tutorial On Using FakeTensor To Determine Max Batch Size</a></p>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="distributed.html" class="btn btn-neutral float-right" title="Distributed communication package - torch.distributed" accesskey="n" rel="next">Next <img src="_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="torch.compiler_dynamic_shapes.html" class="btn btn-neutral" title="Dynamic shapes" accesskey="p" rel="prev"><img src="_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright 2023, PyTorch Contributors.
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
<script>
var match = window.location.href.match(/\/_[a-zA-Z0-9_]*.html|_dynamo/gi);
var url = window.location.href.lastIndexOf(match[match.length-1]);
if (url)
{
var div = '<div class="admonition note"><p class="admonition-title">Note</p><p><i class="fa fa-exclamation-circle" aria-hidden="true"> </i> This page describes an internal API which is not intended to be used outside of the PyTorch codebase and can be modified or removed without notice.</p></div>'
document.getElementById("pytorch-article").insertAdjacentHTML('afterBegin', div)
}
</script>
</div>
<div class="pytorch-content-right" id="pytorch-content-right">
<div class="pytorch-right-menu" id="pytorch-right-menu">
<div class="pytorch-side-scroll" id="pytorch-side-scroll-right">
<ul>
<li><a class="reference internal" href="#">Fake tensor</a><ul>
<li><a class="reference internal" href="#motivation">Motivation</a></li>
<li><a class="reference internal" href="#related-work">Related work</a></li>
<li><a class="reference internal" href="#overall-architecture">Overall architecture</a></li>
<li><a class="reference internal" href="#api-the-important-bits">API: the important bits</a></li>
<li><a class="reference internal" href="#details">Details</a></li>
<li><a class="reference internal" href="#about-the-tensor-subclass">About the tensor subclass</a></li>
<li><a class="reference internal" href="#how-is-each-individual-operator-implemented">How is each individual operator implemented?</a></li>
<li><a class="reference internal" href="#how-does-the-converter-work">How does the converter work?</a></li>
<li><a class="reference internal" href="#performance-characteristics">Performance characteristics</a></li>
<li><a class="reference internal" href="#fake-tensor-of-fake-tensor">Fake tensor of fake tensor?</a></li>
<li><a class="reference internal" href="#interaction-with-dynamic-shapes">Interaction with dynamic shapes</a></li>
<li><a class="reference internal" href="#other-resources">Other resources</a></li>
</ul>
</li>
</ul>
</div>
</div>
</div>
</section>
</div>
<script type="text/javascript" id="documentation_options" data-url_root="./" src="_static/documentation_options.js"></script>
<script data-url_root="./" id="documentation_options" src="_static/documentation_options.js"></script>
<script src="_static/jquery.js"></script>
<script src="_static/underscore.js"></script>
<script src="_static/_sphinx_javascript_frameworks_compat.js"></script>
<script src="_static/doctools.js"></script>
<script src="_static/sphinx_highlight.js"></script>
<script src="_static/clipboard.min.js"></script>
<script src="_static/copybutton.js"></script>
<script type="text/javascript" src="_static/js/vendor/popper.min.js"></script>
<script type="text/javascript" src="_static/js/vendor/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/list.js/1.5.0/list.min.js"></script>
<script type="text/javascript" src="_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<script script type="text/javascript">
var collapsedSections = ['Developer Notes', 'Language Bindings', 'Libraries', 'Community'];
</script>
<img height="1" width="1" style="border-style:none;" alt="" src="https://www.googleadservices.com/pagead/conversion/795629140/?label=txkmCPmdtosBENSssfsC&guid=ON&script=0"/>
<!-- Begin Footer -->
<div class="container-fluid docs-tutorials-resources" id="docs-tutorials-resources">
<div class="container">
<div class="row">
<div class="col-md-4 text-center">
<h2>Docs</h2>
<p>Access comprehensive developer documentation for PyTorch</p>
<a class="with-right-arrow" href="https://pytorch.org/docs/stable/index.html">View Docs</a>
</div>
<div class="col-md-4 text-center">
<h2>Tutorials</h2>
<p>Get in-depth tutorials for beginners and advanced developers</p>
<a class="with-right-arrow" href="https://pytorch.org/tutorials">View Tutorials</a>
</div>
<div class="col-md-4 text-center">
<h2>Resources</h2>
<p>Find development resources and get your questions answered</p>
<a class="with-right-arrow" href="https://pytorch.org/resources">View Resources</a>
</div>
</div>
</div>
</div>
<footer class="site-footer">
<div class="container footer-container">
<div class="footer-logo-wrapper">
<a href="https://pytorch.org/" class="footer-logo"></a>
</div>
<div class="footer-links-wrapper">
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/">PyTorch</a></li>
<li><a href="https://pytorch.org/get-started">Get Started</a></li>
<li><a href="https://pytorch.org/features">Features</a></li>
<li><a href="https://pytorch.org/ecosystem">Ecosystem</a></li>
<li><a href="https://pytorch.org/blog/">Blog</a></li>
<li><a href="https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md">Contributing</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/resources">Resources</a></li>
<li><a href="https://pytorch.org/tutorials">Tutorials</a></li>
<li><a href="https://pytorch.org/docs/stable/index.html">Docs</a></li>
<li><a href="https://discuss.pytorch.org" target="_blank">Discuss</a></li>
<li><a href="https://github.com/pytorch/pytorch/issues" target="_blank">Github Issues</a></li>
<li><a href="https://pytorch.org/assets/brand-guidelines/PyTorch-Brand-Guidelines.pdf" target="_blank">Brand Guidelines</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title">Stay up to date</li>
<li><a href="https://www.facebook.com/pytorch" target="_blank">Facebook</a></li>
<li><a href="https://twitter.com/pytorch" target="_blank">Twitter</a></li>
<li><a href="https://www.youtube.com/pytorch" target="_blank">YouTube</a></li>
<li><a href="https://www.linkedin.com/company/pytorch" target="_blank">LinkedIn</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title">PyTorch Podcasts</li>
<li><a href="https://open.spotify.com/show/6UzHKeiy368jKfQMKKvJY5" target="_blank">Spotify</a></li>
<li><a href="https://podcasts.apple.com/us/podcast/pytorch-developer-podcast/id1566080008" target="_blank">Apple</a></li>
<li><a href="https://www.google.com/podcasts?feed=aHR0cHM6Ly9mZWVkcy5zaW1wbGVjYXN0LmNvbS9PQjVGa0lsOA%3D%3D" target="_blank">Google</a></li>
<li><a href="https://music.amazon.com/podcasts/7a4e6f0e-26c2-49e9-a478-41bd244197d0/PyTorch-Developer-Podcast?" target="_blank">Amazon</a></li>
</ul>
</div>
</div>
<div class="privacy-policy">
<ul>
<li class="privacy-policy-links"><a href="https://www.linuxfoundation.org/terms/" target="_blank">Terms</a></li>
<li class="privacy-policy-links">|</li>
<li class="privacy-policy-links"><a href="https://www.linuxfoundation.org/privacy-policy/" target="_blank">Privacy</a></li>
</ul>
</div>
<div class="copyright">
<p>© Copyright The Linux Foundation. The PyTorch Foundation is a project of The Linux Foundation.
For web site terms of use, trademark policy and other policies applicable to The PyTorch Foundation please see
<a href="https://www.linuxfoundation.org/policies/">www.linuxfoundation.org/policies/</a>. The PyTorch Foundation supports the PyTorch open source
project, which has been established as PyTorch Project a Series of LF Projects, LLC. For policies applicable to the PyTorch Project a Series of LF Projects, LLC,
please see <a href="https://www.lfprojects.org/policies/">www.lfprojects.org/policies/</a>.</p>
</div>
</div>
</footer>
<div class="cookie-banner-wrapper">
<div class="container">
<p class="gdpr-notice">To analyze traffic and optimize your experience, we serve cookies on this site. By clicking or navigating, you agree to allow our usage of cookies. As the current maintainers of this site, Facebook’s Cookies Policy applies. Learn more, including about available controls: <a href="https://www.facebook.com/policies/cookies/">Cookies Policy</a>.</p>
<img class="close-button" src="_static/images/pytorch-x.svg">
</div>
</div>
<!-- End Footer -->
<!-- Begin Mobile Menu -->
<div class="mobile-main-menu">
<div class="container-fluid">
<div class="container">
<div class="mobile-main-menu-header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<a class="main-menu-close-button" href="#" data-behavior="close-mobile-menu"></a>
</div>
</div>
</div>
<div class="mobile-main-menu-links-container">
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<a href="">Mobile</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="resources-mobile-menu-title" class="active">
Docs
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/docs/stable/index.html">PyTorch</a>
</li>
<li>
<a href="https://pytorch.org/audio/stable/index.html">torchaudio</a>
</li>
<li>
<a href="https://pytorch.org/text/stable/index.html">torchtext</a>
</li>
<li>
<a href="https://pytorch.org/vision/stable/index.html">torchvision</a>
</li>
<li>
<a href="https://pytorch.org/torcharrow">torcharrow</a>
</li>
<li>
<a href="https://pytorch.org/data">TorchData</a>
</li>
<li>
<a href="https://pytorch.org/torchrec">TorchRec</a>
</li>
<li>
<a href="https://pytorch.org/serve/">TorchServe</a>
</li>
<li>
<a href="https://pytorch.org/torchx/">TorchX</a>
</li>
<li>
<a href="https://pytorch.org/xla">PyTorch on XLA Devices</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
Resources
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/features">About</a>
</li>
<li>
<a href="https://pytorch.org/foundation">PyTorch Foundation</a>
</li>
<li>
<a href="https://pytorch.org/#community-module">Community</a>
</li>
<li>
<a href="https://pytorch.org/community-stories">Community Stories</a>
</li>
<li>
<a href="https://pytorch.org/resources">Developer Resources</a>
</li>
<li>
<a href="https://pytorch.org/events">Events</a>
</li>
<li>
<a href="https://discuss.pytorch.org/">Forums</a>
</li>
<li>
<a href="https://pytorch.org/hub">Models (Beta)</a>
</li>
</ul>
<li>
<a href="https://github.com/pytorch/pytorch">Github</a>
</li>
</ul>
</div>
</div>
</div>
<!-- End Mobile Menu -->
<script type="text/javascript" src="_static/js/vendor/anchor.min.js"></script>
<script type="text/javascript">
$(document).ready(function() {
mobileMenu.bind();
mobileTOC.bind();
pytorchAnchors.bind();
sideMenus.bind();
scrollToAnchor.bind();
highlightNavigation.bind();
mainMenuDropdown.bind();
filterTags.bind();
// Add class to links that have code blocks, since we cannot create links in code blocks
$("article.pytorch-article a span.pre").each(function(e) {
$(this).closest("a").addClass("has-code");
});
})
</script>
</body>
</html>