-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathgraphs.html
1238 lines (967 loc) · 95.8 KB
/
graphs.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 name="robots" content="noindex">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>torch.cuda.graphs — PyTorch 2.0 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/_modules/torch/cuda/graphs.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" />
<!-- Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-117752657-2"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-117752657-2');
</script>
<!-- End Google Analytics -->
<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>
<a href="https://pytorch.org/mobile">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="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'>2.0 ▼</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">torch.compile</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../../../dynamo/index.html">TorchDynamo Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../dynamo/installation.html">Installing TorchDynamo</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../dynamo/get-started.html">Getting Started</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../dynamo/guards-overview.html">Guards Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../dynamo/custom-backends.html">Custom Backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../dynamo/deep-dive.html">TorchDynamo Deeper Dive</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../dynamo/troubleshooting.html">TorchDynamo Troubleshooting</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../dynamo/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../ir.html">IRs</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../../../cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../deploy.html">torch::deploy</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python API</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../../../torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../amp.html">torch.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../library.html">torch.library</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../cuda.html">torch.cuda</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"><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="../../../_dynamo.html">torch._dynamo</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="../../../onnx_diagnostics.html">torch.onnx diagnostics</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="../../../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="../../../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>
</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="../../index.html">Module code</a> ></li>
<li><a href="../../torch.html">torch</a> ></li>
<li><a href="../cuda.html">torch.cuda</a> ></li>
<li>torch.cuda.graphs</li>
<li class="pytorch-breadcrumbs-aside">
</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">
<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">
<h1>Source code for torch.cuda.graphs</h1><div class="highlight"><pre>
<span></span><span class="kn">import</span> <span class="nn">gc</span>
<span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">from</span> <span class="nn">._utils</span> <span class="kn">import</span> <span class="n">_dummy_type</span>
<span class="kn">from</span> <span class="nn">torch.utils._pytree</span> <span class="kn">import</span> <span class="n">tree_flatten</span> <span class="k">as</span> <span class="n">_tree_flatten</span>
<span class="kn">from</span> <span class="nn">torch.utils._pytree</span> <span class="kn">import</span> <span class="n">tree_unflatten</span> <span class="k">as</span> <span class="n">_tree_unflatten</span>
<span class="k">if</span> <span class="ow">not</span> <span class="nb">hasattr</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">_C</span><span class="p">,</span> <span class="s1">'_CudaStreamBase'</span><span class="p">):</span>
<span class="c1"># Define dummy base classes</span>
<span class="n">torch</span><span class="o">.</span><span class="n">_C</span><span class="o">.</span><span class="vm">__dict__</span><span class="p">[</span><span class="s1">'_CUDAGraph'</span><span class="p">]</span> <span class="o">=</span> <span class="n">_dummy_type</span><span class="p">(</span><span class="s1">'_CUDAGraph'</span><span class="p">)</span>
<span class="n">torch</span><span class="o">.</span><span class="n">_C</span><span class="o">.</span><span class="vm">__dict__</span><span class="p">[</span><span class="s1">'_graph_pool_handle'</span><span class="p">]</span> <span class="o">=</span> <span class="n">_dummy_type</span><span class="p">(</span><span class="s1">'_graph_pool_handle'</span><span class="p">)</span>
<span class="n">torch</span><span class="o">.</span><span class="n">_C</span><span class="o">.</span><span class="vm">__dict__</span><span class="p">[</span><span class="s1">'_cuda_isCurrentStreamCapturing'</span><span class="p">]</span> <span class="o">=</span> <span class="n">_dummy_type</span><span class="p">(</span><span class="s1">'_cuda_isCurrentStreamCapturing'</span><span class="p">)</span>
<span class="kn">from</span> <span class="nn">torch._C</span> <span class="kn">import</span> <span class="n">_CUDAGraph</span> <span class="c1"># noqa: F401</span>
<span class="kn">from</span> <span class="nn">torch._C</span> <span class="kn">import</span> <span class="n">_graph_pool_handle</span>
<span class="kn">from</span> <span class="nn">torch._C</span> <span class="kn">import</span> <span class="n">_cuda_isCurrentStreamCapturing</span>
<div class="viewcode-block" id="is_current_stream_capturing"><a class="viewcode-back" href="../../../generated/torch.cuda.is_current_stream_capturing.html#torch.cuda.is_current_stream_capturing">[docs]</a><span class="k">def</span> <span class="nf">is_current_stream_capturing</span><span class="p">():</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Returns True if CUDA graph capture is underway on the current CUDA stream, False otherwise.</span>
<span class="sd"> If a CUDA context does not exist on the current device, returns False without initializing the context.</span>
<span class="sd"> """</span>
<span class="k">return</span> <span class="n">_cuda_isCurrentStreamCapturing</span><span class="p">()</span></div>
<span class="c1"># Python shim helps Sphinx process docstrings more reliably.</span>
<div class="viewcode-block" id="graph_pool_handle"><a class="viewcode-back" href="../../../generated/torch.cuda.graph_pool_handle.html#torch.cuda.graph_pool_handle">[docs]</a><span class="k">def</span> <span class="nf">graph_pool_handle</span><span class="p">():</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Returns an opaque token representing the id of a graph memory pool.</span>
<span class="sd"> See :ref:`Graph memory management<graph-memory-management>`.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> This API is in beta and may change in future releases.</span>
<span class="sd"> """</span>
<span class="k">return</span> <span class="n">_graph_pool_handle</span><span class="p">()</span></div>
<span class="c1"># Python shim helps Sphinx process docstrings more reliably.</span>
<div class="viewcode-block" id="CUDAGraph"><a class="viewcode-back" href="../../../generated/torch.cuda.CUDAGraph.html#torch.cuda.CUDAGraph">[docs]</a><span class="k">class</span> <span class="nc">CUDAGraph</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">_C</span><span class="o">.</span><span class="n">_CUDAGraph</span><span class="p">):</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Wrapper around a CUDA graph.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> This API is in beta and may change in future releases.</span>
<span class="sd"> """</span>
<span class="k">def</span> <span class="fm">__new__</span><span class="p">(</span><span class="bp">cls</span><span class="p">):</span>
<span class="k">return</span> <span class="nb">super</span><span class="p">(</span><span class="n">CUDAGraph</span><span class="p">,</span> <span class="bp">cls</span><span class="p">)</span><span class="o">.</span><span class="fm">__new__</span><span class="p">(</span><span class="bp">cls</span><span class="p">)</span>
<div class="viewcode-block" id="CUDAGraph.capture_begin"><a class="viewcode-back" href="../../../generated/torch.cuda.CUDAGraph.html#torch.cuda.CUDAGraph.capture_begin">[docs]</a> <span class="k">def</span> <span class="nf">capture_begin</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">pool</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Begins capturing CUDA work on the current stream.</span>
<span class="sd"> Typically, you shouldn't call ``capture_begin`` yourself.</span>
<span class="sd"> Use :class:`~torch.cuda.graph` or :func:`~torch.cuda.make_graphed_callables`,</span>
<span class="sd"> which call ``capture_begin`` internally.</span>
<span class="sd"> Arguments:</span>
<span class="sd"> pool (optional): Token (returned by :func:`~torch.cuda.graph_pool_handle` or</span>
<span class="sd"> :meth:`other_Graph_instance.pool()<torch.cuda.CUDAGraph.pool>`) that hints this graph may share memory</span>
<span class="sd"> with the indicated pool. See :ref:`Graph memory management<graph-memory-management>`.</span>
<span class="sd"> """</span>
<span class="c1"># I'm not sure if pybind11 converts a None arg to the default defined on the C++ side,</span>
<span class="c1"># so I'm not taking any chances.</span>
<span class="k">if</span> <span class="n">pool</span> <span class="ow">is</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="n">capture_begin</span><span class="p">()</span>
<span class="k">else</span><span class="p">:</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="n">capture_begin</span><span class="p">(</span><span class="n">pool</span><span class="p">)</span></div>
<div class="viewcode-block" id="CUDAGraph.capture_end"><a class="viewcode-back" href="../../../generated/torch.cuda.CUDAGraph.html#torch.cuda.CUDAGraph.capture_end">[docs]</a> <span class="k">def</span> <span class="nf">capture_end</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Ends CUDA graph capture on the current stream.</span>
<span class="sd"> After ``capture_end``, ``replay`` may be called on this instance.</span>
<span class="sd"> Typically, you shouldn't call ``capture_end`` yourself.</span>
<span class="sd"> Use :class:`~torch.cuda.graph` or :func:`~torch.cuda.make_graphed_callables`,</span>
<span class="sd"> which call ``capture_end`` internally.</span>
<span class="sd"> """</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="n">capture_end</span><span class="p">()</span></div>
<div class="viewcode-block" id="CUDAGraph.replay"><a class="viewcode-back" href="../../../generated/torch.cuda.CUDAGraph.html#torch.cuda.CUDAGraph.replay">[docs]</a> <span class="k">def</span> <span class="nf">replay</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Replays the CUDA work captured by this graph.</span>
<span class="sd"> """</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="n">replay</span><span class="p">()</span></div>
<div class="viewcode-block" id="CUDAGraph.reset"><a class="viewcode-back" href="../../../generated/torch.cuda.CUDAGraph.html#torch.cuda.CUDAGraph.reset">[docs]</a> <span class="k">def</span> <span class="nf">reset</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Deletes the graph currently held by this instance.</span>
<span class="sd"> """</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="n">reset</span><span class="p">()</span></div>
<div class="viewcode-block" id="CUDAGraph.pool"><a class="viewcode-back" href="../../../generated/torch.cuda.CUDAGraph.html#torch.cuda.CUDAGraph.pool">[docs]</a> <span class="k">def</span> <span class="nf">pool</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Returns an opaque token representing the id of this graph's memory pool.</span>
<span class="sd"> This id can optionally be passed to another graph's ``capture_begin``,</span>
<span class="sd"> which hints the other graph may share the same memory pool.</span>
<span class="sd"> """</span>
<span class="k">return</span> <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="n">pool</span><span class="p">()</span></div>
<div class="viewcode-block" id="CUDAGraph.enable_debug_mode"><a class="viewcode-back" href="../../../generated/torch.cuda.CUDAGraph.html#torch.cuda.CUDAGraph.enable_debug_mode">[docs]</a> <span class="k">def</span> <span class="nf">enable_debug_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Enables debugging mode for CUDAGraph.debug_dump.</span>
<span class="sd"> """</span>
<span class="k">return</span> <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="n">enable_debug_mode</span><span class="p">()</span></div>
<div class="viewcode-block" id="CUDAGraph.debug_dump"><a class="viewcode-back" href="../../../generated/torch.cuda.CUDAGraph.html#torch.cuda.CUDAGraph.debug_dump">[docs]</a> <span class="k">def</span> <span class="nf">debug_dump</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">debug_path</span><span class="p">):</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Arguments:</span>
<span class="sd"> debug_path (required): Path to dump the graph to.</span>
<span class="sd"> Calls a debugging function to dump the graph if the debugging is</span>
<span class="sd"> enabled via CUDAGraph.enable_debug_mode()</span>
<span class="sd"> """</span>
<span class="k">return</span> <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="n">debug_dump</span><span class="p">(</span><span class="n">debug_path</span><span class="p">)</span></div></div>
<div class="viewcode-block" id="graph"><a class="viewcode-back" href="../../../generated/torch.cuda.graph.html#torch.cuda.graph">[docs]</a><span class="k">class</span> <span class="nc">graph</span><span class="p">:</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Context-manager that captures CUDA work into a :class:`torch.cuda.CUDAGraph`</span>
<span class="sd"> object for later replay.</span>
<span class="sd"> See :ref:`CUDA Graphs <cuda-graph-semantics>` for a general introduction,</span>
<span class="sd"> detailed use, and constraints.</span>
<span class="sd"> Arguments:</span>
<span class="sd"> cuda_graph (torch.cuda.CUDAGraph): Graph object used for capture.</span>
<span class="sd"> pool (optional): Opaque token (returned by a call to :func:`~torch.cuda.graph_pool_handle()` or</span>
<span class="sd"> :meth:`other_Graph_instance.pool()<torch.cuda.CUDAGraph.pool>`) hinting this graph's capture</span>
<span class="sd"> may share memory from the specified pool. See :ref:`Graph memory management<graph-memory-management>`.</span>
<span class="sd"> stream (torch.cuda.Stream, optional): If supplied, will be set as the current stream in the context.</span>
<span class="sd"> If not supplied, ``graph`` sets its own internal side stream as the current stream in the context.</span>
<span class="sd"> .. note::</span>
<span class="sd"> For effective memory sharing, if you pass a ``pool`` used by a previous capture and the previous capture</span>
<span class="sd"> used an explicit ``stream`` argument, you should pass the same ``stream`` argument to this capture.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> This API is in beta and may change in future releases.</span>
<span class="sd"> """</span>
<span class="n">default_capture_stream</span> <span class="o">=</span> <span class="kc">None</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="n">cuda_graph</span><span class="p">,</span>
<span class="n">pool</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">stream</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
<span class="c1"># Lazy-init of default_capture_stream helps avoid circular-import errors.</span>
<span class="c1"># Not thread safe, but graphs already have the general (explicitly documented)</span>
<span class="c1"># restriction that only one capture may be underway at a time in the process.</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="vm">__class__</span><span class="o">.</span><span class="n">default_capture_stream</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="bp">self</span><span class="o">.</span><span class="vm">__class__</span><span class="o">.</span><span class="n">default_capture_stream</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">Stream</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">pool</span> <span class="o">=</span> <span class="p">()</span> <span class="k">if</span> <span class="n">pool</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="p">(</span><span class="n">pool</span><span class="p">,)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">capture_stream</span> <span class="o">=</span> <span class="n">stream</span> <span class="k">if</span> <span class="n">stream</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="bp">self</span><span class="o">.</span><span class="vm">__class__</span><span class="o">.</span><span class="n">default_capture_stream</span>
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">capture_stream</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
<span class="bp">self</span><span class="o">.</span><span class="n">stream_ctx</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">stream</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">capture_stream</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">cuda_graph</span> <span class="o">=</span> <span class="n">cuda_graph</span>
<span class="k">def</span> <span class="fm">__enter__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="c1"># Free as much memory as we can for the graph</span>
<span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">synchronize</span><span class="p">()</span>
<span class="n">gc</span><span class="o">.</span><span class="n">collect</span><span class="p">()</span>
<span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">empty_cache</span><span class="p">()</span>
<span class="c1"># Stackoverflow seems comfortable with this pattern</span>
<span class="c1"># https://stackoverflow.com/questions/26635684/calling-enter-and-exit-manually#39172487</span>
<span class="bp">self</span><span class="o">.</span><span class="n">stream_ctx</span><span class="o">.</span><span class="fm">__enter__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">cuda_graph</span><span class="o">.</span><span class="n">capture_begin</span><span class="p">(</span><span class="o">*</span><span class="bp">self</span><span class="o">.</span><span class="n">pool</span><span class="p">)</span>
<span class="k">def</span> <span class="fm">__exit__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">exc_type</span><span class="p">,</span> <span class="n">exc_value</span><span class="p">,</span> <span class="n">traceback</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">cuda_graph</span><span class="o">.</span><span class="n">capture_end</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">stream_ctx</span><span class="o">.</span><span class="fm">__exit__</span><span class="p">(</span><span class="n">exc_type</span><span class="p">,</span> <span class="n">exc_value</span><span class="p">,</span> <span class="n">traceback</span><span class="p">)</span></div>
<span class="c1"># returning None should propagate exceptions from either capture_end or stream_ctx.__exit__()</span>
<div class="viewcode-block" id="make_graphed_callables"><a class="viewcode-back" href="../../../generated/torch.cuda.make_graphed_callables.html#torch.cuda.make_graphed_callables">[docs]</a><span class="k">def</span> <span class="nf">make_graphed_callables</span><span class="p">(</span><span class="n">callables</span><span class="p">,</span> <span class="n">sample_args</span><span class="p">,</span> <span class="n">num_warmup_iters</span><span class="o">=</span><span class="mi">3</span><span class="p">,</span> <span class="n">allow_unused_input</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
<span class="sa">r</span><span class="sd">"""</span>
<span class="sd"> Accepts callables (functions or :class:`nn.Module<torch.nn.Module>`\ s)</span>
<span class="sd"> and returns graphed versions.</span>
<span class="sd"> Each graphed callable's forward pass runs its source callable's</span>
<span class="sd"> forward CUDA work as a CUDA graph inside a single autograd node.</span>
<span class="sd"> The graphed callable's forward pass also appends</span>
<span class="sd"> a backward node to the autograd graph. During backward, this node runs the</span>
<span class="sd"> callable's backward work as a CUDA graph.</span>
<span class="sd"> Therefore, each graphed callable should be a drop-in replacement for its source callable</span>
<span class="sd"> in an autograd-enabled training loop.</span>
<span class="sd"> See :ref:`Partial-network capture<partial-network-capture>` for detailed use and constraints.</span>
<span class="sd"> If you pass a tuple of several callables, their captures will use the same memory pool.</span>
<span class="sd"> See :ref:`Graph memory management<graph-memory-management>` for when this is appropriate.</span>
<span class="sd"> Arguments:</span>
<span class="sd"> callables (torch.nn.Module or Python function, or tuple of these): Callable or callables to graph.</span>
<span class="sd"> See :ref:`Graph memory management<graph-memory-management>` for when passing a tuple of callables</span>
<span class="sd"> is appropriate. If you pass a tuple of callables, their order in the tuple must be the same order</span>
<span class="sd"> they'll run in the live workload.</span>
<span class="sd"> sample_args (tuple of Tensors, or tuple of tuples of Tensors): Samples args for each callable.</span>
<span class="sd"> If a single callable was passed, ``sample_args`` must be a single tuple of argument Tensors.</span>
<span class="sd"> If a tuple of callables was passed, ``sample_args`` must be tuple of tuples of argument Tensors.</span>
<span class="sd"> num_warmup_iters (int): The number of warmup iterations. Currently, ``DataDistributedParallel`` needs</span>
<span class="sd"> 11 iterations for warm up. Default: ``3``.</span>
<span class="sd"> allow_unused_input (bool): If False, specifying inputs that were not used when computing outputs</span>
<span class="sd"> (and therefore their grad is always zero) is an error. Defaults to False.</span>
<span class="sd"> .. note::</span>
<span class="sd"> The ``requires_grad`` state of each Tensor in ``sample_args`` must match the state</span>
<span class="sd"> that's expected for the corresponding real input in the training loop.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> This API is in beta and may change in future releases.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> ``sample_args`` for each callable must contain only Tensors. Other types are not allowed.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> Returned callables do not support higher order differentiation (e.g., double backward).</span>
<span class="sd"> .. warning::</span>
<span class="sd"> In any :class:`~torch.nn.Module` passed to :func:`~make_graphed_callables`, only parameters</span>
<span class="sd"> may be trainable. Buffers must have ``requires_grad=False``.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> After you pass a :class:`torch.nn.Module` through :func:`~make_graphed_callables`,</span>
<span class="sd"> you may not add or remove any of that Module's parameters or buffers.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> :class:`torch.nn.Module`\s passed to :func:`~torch.cuda.make_graphed_callables` must not have module hooks</span>
<span class="sd"> registered on them at the time they are passed. However, registering hooks on modules *after* passing them</span>
<span class="sd"> through :func:`~torch.cuda.make_graphed_callables` is allowed.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> When running a graphed callable, you must pass its arguments in the same order and format</span>
<span class="sd"> they appeared in that callable's ``sample_args``.</span>
<span class="sd"> .. warning::</span>
<span class="sd"> The automatic mixed precision is supported in :func:`~torch.cuda.make_graphed_callables` only with disabled</span>
<span class="sd"> caching. The context manager `torch.cuda.amp.autocast()` must have `cache_enabled=False`.</span>
<span class="sd"> """</span>
<span class="k">if</span> <span class="n">torch</span><span class="o">.</span><span class="n">is_autocast_enabled</span><span class="p">()</span> <span class="ow">and</span> <span class="n">torch</span><span class="o">.</span><span class="n">is_autocast_cache_enabled</span><span class="p">():</span>
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span><span class="s2">"make_graphed_callables does not support the autocast caching. Please set `cache_enabled=False`."</span><span class="p">)</span>
<span class="n">just_one_callable</span> <span class="o">=</span> <span class="kc">False</span>
<span class="k">if</span> <span class="ow">not</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">callables</span><span class="p">,</span> <span class="nb">tuple</span><span class="p">):</span>
<span class="n">just_one_callable</span> <span class="o">=</span> <span class="kc">True</span>
<span class="n">callables</span> <span class="o">=</span> <span class="p">(</span><span class="n">callables</span><span class="p">,)</span>
<span class="n">sample_args</span> <span class="o">=</span> <span class="p">(</span><span class="n">sample_args</span><span class="p">,)</span>
<span class="n">flatten_sample_args</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">c</span><span class="p">,</span> <span class="n">args</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">callables</span><span class="p">,</span> <span class="n">sample_args</span><span class="p">):</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">c</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">assert</span> <span class="nb">len</span><span class="p">(</span><span class="n">c</span><span class="o">.</span><span class="n">_backward_hooks</span><span class="p">)</span> <span class="o">==</span> <span class="mi">0</span> <span class="ow">and</span> <span class="nb">len</span><span class="p">(</span><span class="n">c</span><span class="o">.</span><span class="n">_forward_hooks</span><span class="p">)</span> <span class="o">==</span> <span class="mi">0</span> <span class="ow">and</span> <span class="nb">len</span><span class="p">(</span><span class="n">c</span><span class="o">.</span><span class="n">_forward_pre_hooks</span><span class="p">)</span> <span class="o">==</span> <span class="mi">0</span><span class="p">,</span> \
<span class="s2">"Modules must not have hooks registered at the time they are passed. However, registering hooks "</span> <span class="o">+</span> \
<span class="s2">"on modules after passing them through make_graphed_callables is allowed."</span>
<span class="k">assert</span> <span class="nb">all</span><span class="p">(</span><span class="n">b</span><span class="o">.</span><span class="n">requires_grad</span> <span class="ow">is</span> <span class="kc">False</span> <span class="k">for</span> <span class="n">b</span> <span class="ow">in</span> <span class="n">c</span><span class="o">.</span><span class="n">buffers</span><span class="p">()),</span> <span class="s2">"In any :class:`~torch.nn.Module` passed to "</span> <span class="o">+</span> \
<span class="s2">":func:`~make_graphed_callables`, only parameters may be trainable. All buffers must have "</span> <span class="o">+</span> \
<span class="s2">"``requires_grad=False``."</span>
<span class="n">flatten_arg</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="n">_tree_flatten</span><span class="p">(</span><span class="n">args</span><span class="p">)</span>
<span class="n">flatten_sample_args</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="nb">tuple</span><span class="p">(</span><span class="n">flatten_arg</span><span class="p">))</span>
<span class="k">assert</span> <span class="nb">all</span><span class="p">(</span><span class="nb">isinstance</span><span class="p">(</span><span class="n">arg</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="k">for</span> <span class="n">arg</span> <span class="ow">in</span> <span class="n">flatten_arg</span><span class="p">),</span> <span class="s2">"In the beta API, sample_args "</span> <span class="o">+</span> \
<span class="s2">"for each callable must contain only Tensors. Other types are not allowed."</span>
<span class="c1"># If a callable is an nn.Module, its graph's full input surface is the args the user explicitly</span>
<span class="c1"># passes to forward (ie, its sample_args) AND the module's parameter attributes.</span>
<span class="n">per_callable_len_user_args</span> <span class="o">=</span> <span class="p">[</span><span class="nb">len</span><span class="p">(</span><span class="n">args</span><span class="p">)</span> <span class="k">for</span> <span class="n">args</span> <span class="ow">in</span> <span class="n">flatten_sample_args</span><span class="p">]</span>
<span class="n">per_callable_module_params</span> <span class="o">=</span> <span class="p">[</span><span class="nb">tuple</span><span class="p">(</span><span class="n">c</span><span class="o">.</span><span class="n">parameters</span><span class="p">())</span> <span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">c</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">else</span> <span class="p">()</span>
<span class="k">for</span> <span class="n">c</span> <span class="ow">in</span> <span class="n">callables</span><span class="p">]</span>
<span class="n">per_callable_static_input_surfaces</span> <span class="o">=</span> <span class="p">[</span><span class="n">flatten_sample_args</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">+</span> <span class="n">per_callable_module_params</span><span class="p">[</span><span class="n">i</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="nb">len</span><span class="p">(</span><span class="n">callables</span><span class="p">))]</span>
<span class="n">fwd_graphs</span> <span class="o">=</span> <span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">CUDAGraph</span><span class="p">()</span> <span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="n">callables</span><span class="p">))]</span>
<span class="n">bwd_graphs</span> <span class="o">=</span> <span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">CUDAGraph</span><span class="p">()</span> <span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="n">callables</span><span class="p">))]</span>
<span class="n">mempool</span> <span class="o">=</span> <span class="n">graph_pool_handle</span><span class="p">()</span>
<span class="c1"># Warmup</span>
<span class="c1"># Hopefully prevents cudnn benchmarking and other lazy-initialization cuda work</span>
<span class="c1"># from ending up in any captures.</span>
<span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">synchronize</span><span class="p">()</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">stream</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">Stream</span><span class="p">()):</span>
<span class="k">for</span> <span class="n">func</span><span class="p">,</span> <span class="n">args</span><span class="p">,</span> <span class="n">static_input_surface</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">callables</span><span class="p">,</span>
<span class="n">sample_args</span><span class="p">,</span>
<span class="n">per_callable_static_input_surfaces</span><span class="p">):</span>
<span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">num_warmup_iters</span><span class="p">):</span>
<span class="n">outputs</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="n">_tree_flatten</span><span class="p">(</span><span class="n">func</span><span class="p">(</span><span class="o">*</span><span class="n">args</span><span class="p">))</span>
<span class="n">grad_inputs</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">autograd</span><span class="o">.</span><span class="n">grad</span><span class="p">(</span><span class="n">outputs</span><span class="o">=</span><span class="nb">tuple</span><span class="p">(</span><span class="n">o</span> <span class="k">for</span> <span class="n">o</span> <span class="ow">in</span> <span class="n">outputs</span> <span class="k">if</span> <span class="n">o</span><span class="o">.</span><span class="n">requires_grad</span><span class="p">),</span>
<span class="n">inputs</span><span class="o">=</span><span class="nb">tuple</span><span class="p">(</span><span class="n">i</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="n">static_input_surface</span> <span class="k">if</span> <span class="n">i</span><span class="o">.</span><span class="n">requires_grad</span><span class="p">),</span>
<span class="n">grad_outputs</span><span class="o">=</span><span class="nb">tuple</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">empty_like</span><span class="p">(</span><span class="n">o</span><span class="p">)</span> <span class="k">for</span> <span class="n">o</span> <span class="ow">in</span> <span class="n">outputs</span> <span class="k">if</span> <span class="n">o</span><span class="o">.</span><span class="n">requires_grad</span><span class="p">),</span>
<span class="n">only_inputs</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="n">allow_unused</span><span class="o">=</span><span class="n">allow_unused_input</span><span class="p">)</span>
<span class="k">del</span> <span class="n">outputs</span><span class="p">,</span> <span class="n">grad_inputs</span>
<span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">synchronize</span><span class="p">()</span>
<span class="c1"># All captures here share a mempool. To avoid replays corrupting each other's memory,</span>
<span class="c1"># the safest approach is to capture all passes in the same order they'll run:</span>
<span class="c1"># fwd 1, fwd 2, ... fwd N, then bwd N, bwd N-1, ... bwd 1.</span>
<span class="c1"># Capture forward graphs</span>
<span class="n">per_callable_static_outputs</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">per_callable_output_unflatten_spec</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">func</span><span class="p">,</span> <span class="n">args</span><span class="p">,</span> <span class="n">fwd_graph</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">callables</span><span class="p">,</span>
<span class="n">sample_args</span><span class="p">,</span>
<span class="n">fwd_graphs</span><span class="p">):</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">graph</span><span class="p">(</span><span class="n">fwd_graph</span><span class="p">,</span> <span class="n">pool</span><span class="o">=</span><span class="n">mempool</span><span class="p">):</span>
<span class="n">outputs</span> <span class="o">=</span> <span class="n">func</span><span class="p">(</span><span class="o">*</span><span class="n">args</span><span class="p">)</span>
<span class="n">flatten_outputs</span><span class="p">,</span> <span class="n">spec</span> <span class="o">=</span> <span class="n">_tree_flatten</span><span class="p">(</span><span class="n">outputs</span><span class="p">)</span>
<span class="n">per_callable_static_outputs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="nb">tuple</span><span class="p">(</span><span class="n">flatten_outputs</span><span class="p">))</span>
<span class="n">per_callable_output_unflatten_spec</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">spec</span><span class="p">)</span>
<span class="c1"># Capture backward graphs in reverse order</span>
<span class="n">per_callable_static_grad_outputs</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">per_callable_static_grad_inputs</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">static_input_surface</span><span class="p">,</span> <span class="n">static_outputs</span><span class="p">,</span> <span class="n">bwd_graph</span><span class="p">,</span> <span class="n">module_params</span> <span class="ow">in</span> \
<span class="nb">zip</span><span class="p">(</span><span class="nb">reversed</span><span class="p">(</span><span class="n">per_callable_static_input_surfaces</span><span class="p">),</span>
<span class="nb">reversed</span><span class="p">(</span><span class="n">per_callable_static_outputs</span><span class="p">),</span>
<span class="nb">reversed</span><span class="p">(</span><span class="n">bwd_graphs</span><span class="p">),</span>
<span class="nb">reversed</span><span class="p">(</span><span class="n">per_callable_module_params</span><span class="p">)):</span>
<span class="c1"># For now, assumes all static_outputs require grad</span>
<span class="c1"># assert all(o.requires_grad for o in static_outputs), "Outputs of graphed callables must require grad."</span>
<span class="n">static_grad_outputs</span> <span class="o">=</span> <span class="nb">tuple</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">empty_like</span><span class="p">(</span><span class="n">o</span><span class="p">)</span> <span class="k">if</span> <span class="n">o</span><span class="o">.</span><span class="n">requires_grad</span> <span class="k">else</span> <span class="kc">None</span> <span class="k">for</span> <span class="n">o</span> <span class="ow">in</span> <span class="n">static_outputs</span><span class="p">)</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">graph</span><span class="p">(</span><span class="n">bwd_graph</span><span class="p">,</span> <span class="n">pool</span><span class="o">=</span><span class="n">mempool</span><span class="p">):</span>
<span class="n">grad_inputs</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">autograd</span><span class="o">.</span><span class="n">grad</span><span class="p">(</span><span class="n">outputs</span><span class="o">=</span><span class="nb">tuple</span><span class="p">(</span><span class="n">o</span> <span class="k">for</span> <span class="n">o</span> <span class="ow">in</span> <span class="n">static_outputs</span> <span class="k">if</span> <span class="n">o</span><span class="o">.</span><span class="n">requires_grad</span><span class="p">),</span>
<span class="n">inputs</span><span class="o">=</span><span class="nb">tuple</span><span class="p">(</span><span class="n">i</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="n">static_input_surface</span> <span class="k">if</span> <span class="n">i</span><span class="o">.</span><span class="n">requires_grad</span><span class="p">),</span>
<span class="n">grad_outputs</span><span class="o">=</span><span class="nb">tuple</span><span class="p">(</span><span class="n">o</span> <span class="k">for</span> <span class="n">o</span> <span class="ow">in</span> <span class="n">static_grad_outputs</span> <span class="k">if</span> <span class="n">o</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">),</span>
<span class="n">only_inputs</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="n">allow_unused</span><span class="o">=</span><span class="n">allow_unused_input</span><span class="p">)</span>
<span class="c1"># Constructs a tuple suitable for returning from Graphed.backward:</span>
<span class="c1"># Pads out the actually-needed grads with Nones in gradient slots for inputs that don't require grad.</span>
<span class="c1"># I couldn't think of a slick one-liner for this pattern.</span>
<span class="n">static_grad_inputs</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">grad_idx</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">for</span> <span class="n">arg</span> <span class="ow">in</span> <span class="n">static_input_surface</span><span class="p">:</span>
<span class="k">if</span> <span class="n">arg</span><span class="o">.</span><span class="n">requires_grad</span><span class="p">:</span>
<span class="n">static_grad_inputs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">grad_inputs</span><span class="p">[</span><span class="n">grad_idx</span><span class="p">])</span>
<span class="n">grad_idx</span> <span class="o">+=</span> <span class="mi">1</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">static_grad_inputs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="kc">None</span><span class="p">)</span> <span class="c1"># type: ignore[arg-type]</span>
<span class="n">static_grad_inputs</span> <span class="o">=</span> <span class="nb">tuple</span><span class="p">(</span><span class="n">static_grad_inputs</span><span class="p">)</span> <span class="c1"># type: ignore[assignment]</span>
<span class="n">per_callable_static_grad_outputs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">static_grad_outputs</span><span class="p">)</span>
<span class="n">per_callable_static_grad_inputs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">static_grad_inputs</span><span class="p">)</span>
<span class="c1"># Reverses the most recent two lists</span>
<span class="n">per_callable_static_grad_outputs</span> <span class="o">=</span> <span class="nb">list</span><span class="p">(</span><span class="nb">reversed</span><span class="p">(</span><span class="n">per_callable_static_grad_outputs</span><span class="p">))</span>
<span class="n">per_callable_static_grad_inputs</span> <span class="o">=</span> <span class="nb">list</span><span class="p">(</span><span class="nb">reversed</span><span class="p">(</span><span class="n">per_callable_static_grad_inputs</span><span class="p">))</span>
<span class="c1"># Now for every per_callable list, per_callable_*[i] holds the stuff for the ith callable.</span>
<span class="k">def</span> <span class="nf">make_graphed_autograd_function</span><span class="p">(</span><span class="n">fwd_graph</span><span class="p">,</span>
<span class="n">bwd_graph</span><span class="p">,</span>
<span class="n">module_params</span><span class="p">,</span>
<span class="n">len_user_args</span><span class="p">,</span>
<span class="n">output_unflatten_spec</span><span class="p">,</span>
<span class="n">static_input_surface</span><span class="p">,</span>
<span class="n">static_outputs</span><span class="p">,</span>
<span class="n">static_grad_outputs</span><span class="p">,</span>
<span class="n">static_grad_inputs</span><span class="p">):</span>
<span class="k">class</span> <span class="nc">Graphed</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">autograd</span><span class="o">.</span><span class="n">Function</span><span class="p">):</span>
<span class="nd">@staticmethod</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="n">ctx</span><span class="p">,</span> <span class="o">*</span><span class="n">inputs</span><span class="p">):</span>
<span class="c1"># At this stage, only the user args may (potentially) be new tensors.</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">len_user_args</span><span class="p">):</span>
<span class="k">if</span> <span class="n">static_input_surface</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">.</span><span class="n">data_ptr</span><span class="p">()</span> <span class="o">!=</span> <span class="n">inputs</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">.</span><span class="n">data_ptr</span><span class="p">():</span>
<span class="n">static_input_surface</span><span class="p">[</span><span class="n">i</span><span class="p">]</span><span class="o">.</span><span class="n">copy_</span><span class="p">(</span><span class="n">inputs</span><span class="p">[</span><span class="n">i</span><span class="p">])</span>
<span class="n">fwd_graph</span><span class="o">.</span><span class="n">replay</span><span class="p">()</span>
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">static_outputs</span><span class="p">,</span> <span class="nb">tuple</span><span class="p">)</span>
<span class="k">return</span> <span class="nb">tuple</span><span class="p">(</span><span class="n">o</span><span class="o">.</span><span class="n">detach</span><span class="p">()</span> <span class="k">for</span> <span class="n">o</span> <span class="ow">in</span> <span class="n">static_outputs</span><span class="p">)</span>
<span class="nd">@staticmethod</span>
<span class="nd">@torch</span><span class="o">.</span><span class="n">autograd</span><span class="o">.</span><span class="n">function</span><span class="o">.</span><span class="n">once_differentiable</span>
<span class="k">def</span> <span class="nf">backward</span><span class="p">(</span><span class="n">ctx</span><span class="p">,</span> <span class="o">*</span><span class="n">grads</span><span class="p">):</span>
<span class="k">assert</span> <span class="nb">len</span><span class="p">(</span><span class="n">grads</span><span class="p">)</span> <span class="o">==</span> <span class="nb">len</span><span class="p">(</span><span class="n">static_grad_outputs</span><span class="p">)</span>
<span class="k">for</span> <span class="n">g</span><span class="p">,</span> <span class="n">grad</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">static_grad_outputs</span><span class="p">,</span> <span class="n">grads</span><span class="p">):</span>
<span class="k">if</span> <span class="n">g</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="c1"># don't copy if autograd gods have been kind and the</span>
<span class="c1"># incoming grad is already in the right place</span>
<span class="k">if</span> <span class="n">g</span><span class="o">.</span><span class="n">data_ptr</span><span class="p">()</span> <span class="o">!=</span> <span class="n">grad</span><span class="o">.</span><span class="n">data_ptr</span><span class="p">():</span>
<span class="n">g</span><span class="o">.</span><span class="n">copy_</span><span class="p">(</span><span class="n">grad</span><span class="p">)</span>
<span class="n">bwd_graph</span><span class="o">.</span><span class="n">replay</span><span class="p">()</span>
<span class="c1"># Input args that didn't require grad expect a None gradient.</span>
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">static_grad_inputs</span><span class="p">,</span> <span class="nb">tuple</span><span class="p">)</span>
<span class="k">return</span> <span class="nb">tuple</span><span class="p">(</span><span class="n">b</span><span class="o">.</span><span class="n">detach</span><span class="p">()</span> <span class="k">if</span> <span class="n">b</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">b</span> <span class="k">for</span> <span class="n">b</span> <span class="ow">in</span> <span class="n">static_grad_inputs</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">functionalized</span><span class="p">(</span><span class="o">*</span><span class="n">user_args</span><span class="p">):</span>
<span class="c1"># Runs the autograd function with inputs == all inputs to the graph that might require grad</span>
<span class="c1"># (explicit user args + module parameters)</span>
<span class="c1"># Assumes module params didn't change since capture.</span>
<span class="n">flatten_user_args</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="n">_tree_flatten</span><span class="p">(</span><span class="n">user_args</span><span class="p">)</span>
<span class="n">out</span> <span class="o">=</span> <span class="n">Graphed</span><span class="o">.</span><span class="n">apply</span><span class="p">(</span><span class="o">*</span><span class="p">(</span><span class="nb">tuple</span><span class="p">(</span><span class="n">flatten_user_args</span><span class="p">)</span> <span class="o">+</span> <span class="n">module_params</span><span class="p">))</span>
<span class="k">return</span> <span class="n">_tree_unflatten</span><span class="p">(</span><span class="n">out</span><span class="p">,</span> <span class="n">output_unflatten_spec</span><span class="p">)</span>
<span class="k">return</span> <span class="n">functionalized</span>
<span class="c1"># Put together the final graphed callables</span>
<span class="n">ret</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">i</span><span class="p">,</span> <span class="n">func</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">callables</span><span class="p">):</span>
<span class="n">graphed</span> <span class="o">=</span> <span class="n">make_graphed_autograd_function</span><span class="p">(</span><span class="n">fwd_graphs</span><span class="p">[</span><span class="n">i</span><span class="p">],</span>
<span class="n">bwd_graphs</span><span class="p">[</span><span class="n">i</span><span class="p">],</span>
<span class="n">per_callable_module_params</span><span class="p">[</span><span class="n">i</span><span class="p">],</span>
<span class="n">per_callable_len_user_args</span><span class="p">[</span><span class="n">i</span><span class="p">],</span>
<span class="n">per_callable_output_unflatten_spec</span><span class="p">[</span><span class="n">i</span><span class="p">],</span>
<span class="n">per_callable_static_input_surfaces</span><span class="p">[</span><span class="n">i</span><span class="p">],</span>
<span class="n">per_callable_static_outputs</span><span class="p">[</span><span class="n">i</span><span class="p">],</span>
<span class="n">per_callable_static_grad_outputs</span><span class="p">[</span><span class="n">i</span><span class="p">],</span>
<span class="n">per_callable_static_grad_inputs</span><span class="p">[</span><span class="n">i</span><span class="p">])</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">func</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">make_graphed_forward</span><span class="p">(</span><span class="n">func</span><span class="p">,</span> <span class="n">graph_training_state</span><span class="p">,</span> <span class="n">graphed</span><span class="p">,</span> <span class="n">orig_fwd</span><span class="p">):</span>
<span class="k">def</span> <span class="nf">new_fwd</span><span class="p">(</span><span class="o">*</span><span class="n">user_args</span><span class="p">):</span>
<span class="c1"># If the module's training-or-eval state matches what we graphed,</span>
<span class="c1"># run the graph, otherwise run the original forward method</span>
<span class="k">if</span> <span class="n">func</span><span class="o">.</span><span class="n">training</span> <span class="o">==</span> <span class="n">graph_training_state</span><span class="p">:</span>
<span class="k">return</span> <span class="n">graphed</span><span class="p">(</span><span class="o">*</span><span class="n">user_args</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="n">orig_fwd</span><span class="p">(</span><span class="o">*</span><span class="n">user_args</span><span class="p">)</span>
<span class="k">return</span> <span class="n">new_fwd</span>
<span class="n">func</span><span class="o">.</span><span class="n">forward</span> <span class="o">=</span> <span class="n">make_graphed_forward</span><span class="p">(</span><span class="n">func</span><span class="p">,</span> <span class="n">func</span><span class="o">.</span><span class="n">training</span><span class="p">,</span> <span class="n">graphed</span><span class="p">,</span> <span class="n">func</span><span class="o">.</span><span class="n">forward</span><span class="p">)</span> <span class="c1"># type: ignore[assignment]</span>
<span class="n">ret</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">func</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">ret</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">graphed</span><span class="p">)</span>
<span class="k">if</span> <span class="n">just_one_callable</span><span class="p">:</span>
<span class="k">return</span> <span class="n">ret</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
<span class="k">return</span> <span class="nb">tuple</span><span class="p">(</span><span class="n">ret</span><span class="p">)</span></div>
</pre></div>
</article>
</div>
<footer>
<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">
</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/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>