-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathcustom-backends.html
1079 lines (871 loc) · 73.3 KB
/
custom-backends.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta name="robots" content="noindex">
<meta charset="utf-8">
<meta name="generator" content="Docutils 0.18.1: http://docutils.sourceforge.net/" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Custom Backends — PyTorch 2.0 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/torch.compiler_custom_backends.html"/>
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="../_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="../_static/sphinx-dropdown.css" type="text/css" />
<link rel="stylesheet" href="../_static/panels-bootstrap.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/jit.css" type="text/css" />
<link rel="index" title="Index" href="../genindex.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="TorchDynamo Deeper Dive" href="deep-dive.html" />
<link rel="prev" title="Guards Overview" href="guards-overview.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/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<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 class="current">
<li class="toctree-l1 current"><a class="reference internal" href="index.html">TorchDynamo Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="installation.html">Installing TorchDynamo</a></li>
<li class="toctree-l1"><a class="reference internal" href="get-started.html">Getting Started</a></li>
<li class="toctree-l1"><a class="reference internal" href="guards-overview.html">Guards Overview</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Custom Backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="deep-dive.html">TorchDynamo Deeper Dive</a></li>
<li class="toctree-l1"><a class="reference internal" href="troubleshooting.html">TorchDynamo Troubleshooting</a></li>
<li class="toctree-l1"><a class="reference internal" href="faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../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">TorchDynamo Overview</a> ></li>
<li>Custom Backends</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/dynamo/custom-backends.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">
<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">
<section id="custom-backends">
<h1>Custom Backends<a class="headerlink" href="#custom-backends" title="Permalink to this heading">¶</a></h1>
<section id="overview">
<h2>Overview<a class="headerlink" href="#overview" title="Permalink to this heading">¶</a></h2>
<p><code class="docutils literal notranslate"><span class="pre">torch.compile</span></code> provides a straightforward method to enable users
to define custom backends.</p>
<p>A backend function has the contract
<code class="docutils literal notranslate"><span class="pre">(gm:</span> <span class="pre">torch.fx.GraphModule,</span> <span class="pre">example_inputs:</span> <span class="pre">List[torch.Tensor])</span> <span class="pre">-></span> <span class="pre">Callable</span></code>.</p>
<p>Backend functions can be called by TorchDynamo, the graph tracing component of <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code>,
after tracing an FX graph and are
expected to return a compiled function that is equivalent to the traced FX graph.
The returned callable should have the same contract as the <code class="docutils literal notranslate"><span class="pre">forward</span></code> function of the original <code class="docutils literal notranslate"><span class="pre">torch.fx.GraphModule</span></code>
passed into the backend:
<code class="docutils literal notranslate"><span class="pre">(*args:</span> <span class="pre">torch.Tensor)</span> <span class="pre">-></span> <span class="pre">List[torch.Tensor]</span></code>.</p>
<p>In order for TorchDynamo to call your backend, pass your backend function as the <code class="docutils literal notranslate"><span class="pre">backend</span></code> kwarg in
<code class="docutils literal notranslate"><span class="pre">torch.compile</span></code>. For example,</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="k">def</span> <span class="nf">my_custom_backend</span><span class="p">(</span><span class="n">gm</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">):</span>
<span class="k">return</span> <span class="n">gm</span><span class="o">.</span><span class="n">forward</span>
<span class="k">def</span> <span class="nf">f</span><span class="p">(</span><span class="o">...</span><span class="p">):</span>
<span class="o">...</span>
<span class="n">f_opt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">f</span><span class="p">,</span> <span class="n">backend</span><span class="o">=</span><span class="n">my_custom_backend</span><span class="p">)</span>
<span class="nd">@torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">backend</span><span class="o">=</span><span class="n">my_custom_backend</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">g</span><span class="p">(</span><span class="o">...</span><span class="p">):</span>
<span class="o">...</span>
</pre></div>
</div>
<p>See below for more examples.</p>
</section>
<section id="registering-custom-backends">
<h2>Registering Custom Backends<a class="headerlink" href="#registering-custom-backends" title="Permalink to this heading">¶</a></h2>
<p>You can register your backend using the <code class="docutils literal notranslate"><span class="pre">register_backend</span></code> decorator, for example,</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">torch._dynamo.optimizations</span> <span class="kn">import</span> <span class="n">register_backend</span>
<span class="nd">@register_backend</span>
<span class="k">def</span> <span class="nf">my_compiler</span><span class="p">(</span><span class="n">gm</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">):</span>
<span class="o">...</span>
</pre></div>
</div>
<p>Besides the <code class="docutils literal notranslate"><span class="pre">register_backend</span></code> decorator, if your backend is in another python package, you could also register your
backend through entry points of python package, which provides a way for a package to register a plugin for another one.</p>
<div class="admonition hint">
<p class="admonition-title">Hint</p>
<p>You can learn more about <code class="docutils literal notranslate"><span class="pre">entry_points</span></code> in the
<a class="reference external" href="https://setuptools.pypa.io/en/latest/userguide/entry_point.html">python packaging documentation</a>.</p>
</div>
<p>To register your backend through <code class="docutils literal notranslate"><span class="pre">entry_points</span></code>, you could add your backend function to the <code class="docutils literal notranslate"><span class="pre">torch_dynamo_backends</span></code> entry point group in the
<code class="docutils literal notranslate"><span class="pre">setup.py</span></code> file of your package like:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="o">...</span>
<span class="n">setup</span><span class="p">(</span>
<span class="o">...</span>
<span class="s1">'torch_dynamo_backends'</span><span class="p">:</span> <span class="p">[</span>
<span class="s1">'my_compiler = your_module.submodule:my_compiler'</span><span class="p">,</span>
<span class="p">]</span>
<span class="o">...</span>
<span class="p">)</span>
</pre></div>
</div>
<p>Please replace the <code class="docutils literal notranslate"><span class="pre">my_compiler</span></code> before <code class="docutils literal notranslate"><span class="pre">=</span></code> to the name of your backend’s name and replace the part after <code class="docutils literal notranslate"><span class="pre">=</span></code> to
the module and function name of your backend function.
The entry point will be added to your python environment after the installation of the package.
When you call <code class="docutils literal notranslate"><span class="pre">torch.compile(model,</span> <span class="pre">backend="my_compiler")</span></code>, PyTorch would first search the backend named <code class="docutils literal notranslate"><span class="pre">my_compiler</span></code>
that has been registered with <code class="docutils literal notranslate"><span class="pre">register_backend</span></code>. If not found, it will continue to search in all backends registered
via <code class="docutils literal notranslate"><span class="pre">entry_points</span></code>.</p>
<p>Registration serves two purposes:</p>
<ul class="simple">
<li><p>You can pass a string containing your backend function’s name to <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code> instead of the function itself,
for example, <code class="docutils literal notranslate"><span class="pre">torch.compile(model,</span> <span class="pre">backend="my_compiler")</span></code>.</p></li>
<li><p>It is required for use with the <a class="reference external" href="https://pytorch.org/docs/stable/torch.compiler_troubleshooting.html">minifier</a>. Any generated
code from the minifier must call your code that registers your backend function, typically through an <code class="docutils literal notranslate"><span class="pre">import</span></code> statement.</p></li>
</ul>
</section>
<section id="custom-backends-after-aotautograd">
<h2>Custom Backends after AOTAutograd<a class="headerlink" href="#custom-backends-after-aotautograd" title="Permalink to this heading">¶</a></h2>
<p>It is possible to define custom backends that are called by AOTAutograd rather than TorchDynamo.
This is useful for 2 main reasons:</p>
<ul class="simple">
<li><p>Users can define backends that support model training, as AOTAutograd can generate the backward graph for compilation.</p></li>
<li><p>AOTAutograd produces FX graphs consisting of <a class="reference external" href="https://pytorch.org/docs/master/ir.html#canonical-aten-ir">canonical Aten ops</a>. As a result,
custom backends only need to support the canonical Aten opset, which is a significantly smaller opset than the entire torch/Aten opset.</p></li>
</ul>
<p>Wrap your backend with
<code class="docutils literal notranslate"><span class="pre">torch._dynamo.optimizations.training.aot_autograd</span></code> and use <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code> with the <code class="docutils literal notranslate"><span class="pre">backend</span></code> kwarg as before.
Backend functions wrapped by <code class="docutils literal notranslate"><span class="pre">aot_autograd</span></code> should have the same contract as before.</p>
<p>Backend functions are passed to <code class="docutils literal notranslate"><span class="pre">aot_autograd</span></code> through the <code class="docutils literal notranslate"><span class="pre">fw_compiler</span></code> (forward compiler)
or <code class="docutils literal notranslate"><span class="pre">bw_compiler</span></code> (backward compiler) kwargs. If <code class="docutils literal notranslate"><span class="pre">bw_compiler</span></code> is not specified, the backward compile function
defaults to the forward compile function.</p>
<p>One caveat is that AOTAutograd requires compiled functions returned by backends to be “boxed”. This can be done by wrapping
the compiled function with <code class="docutils literal notranslate"><span class="pre">functorch.compile.make_boxed_func</span></code>.</p>
<p>For example,</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">torch._dynamo.optimizations.training</span> <span class="kn">import</span> <span class="n">aot_autograd</span>
<span class="kn">from</span> <span class="nn">functorch.compile</span> <span class="kn">import</span> <span class="n">make_boxed_func</span>
<span class="k">def</span> <span class="nf">my_compiler</span><span class="p">(</span><span class="n">gm</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">):</span>
<span class="k">return</span> <span class="n">make_boxed_func</span><span class="p">(</span><span class="n">gm</span><span class="o">.</span><span class="n">forward</span><span class="p">)</span>
<span class="n">my_backend</span> <span class="o">=</span> <span class="n">aot_autograd</span><span class="p">(</span><span class="n">fw_compiler</span><span class="o">=</span><span class="n">my_compiler</span><span class="p">)</span> <span class="c1"># bw_compiler=my_compiler</span>
<span class="n">model_opt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">backend</span><span class="o">=</span><span class="n">my_backend</span><span class="p">)</span>
</pre></div>
</div>
</section>
<section id="examples">
<h2>Examples<a class="headerlink" href="#examples" title="Permalink to this heading">¶</a></h2>
<section id="debugging-backend">
<h3>Debugging Backend<a class="headerlink" href="#debugging-backend" title="Permalink to this heading">¶</a></h3>
<p>If you want to better understand what is going on during a
compilation, you can create a custom compiler, which is referred to as
backend in this section, that will print pretty print the fx
<code class="docutils literal notranslate"><span class="pre">GraphModule</span></code> extracted from Dynamo’s bytecode analysis
and return a <code class="docutils literal notranslate"><span class="pre">forward()</span></code> callable.</p>
<p>For example:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">List</span>
<span class="kn">import</span> <span class="nn">torch</span>
<span class="k">def</span> <span class="nf">my_compiler</span><span class="p">(</span><span class="n">gm</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">:</span> <span class="n">List</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="nb">print</span><span class="p">(</span><span class="s2">"my_compiler() called with FX graph:"</span><span class="p">)</span>
<span class="n">gm</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">print_tabular</span><span class="p">()</span>
<span class="k">return</span> <span class="n">gm</span><span class="o">.</span><span class="n">forward</span> <span class="c1"># return a python callable</span>
<span class="nd">@torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">backend</span><span class="o">=</span><span class="n">my_compiler</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">fn</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">):</span>
<span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cos</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">sin</span><span class="p">(</span><span class="n">y</span><span class="p">)</span>
<span class="k">return</span> <span class="n">a</span> <span class="o">+</span> <span class="n">b</span>
<span class="n">fn</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">10</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">10</span><span class="p">))</span>
</pre></div>
</div>
<p>Running the above example produces the following output:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">my_compiler</span><span class="p">()</span> <span class="n">called</span> <span class="k">with</span> <span class="n">FX</span> <span class="n">graph</span><span class="p">:</span>
<span class="n">opcode</span> <span class="n">name</span> <span class="n">target</span> <span class="n">args</span> <span class="n">kwargs</span>
<span class="o">-------------</span> <span class="o">------</span> <span class="o">------------------------------------------------------</span> <span class="o">----------</span> <span class="o">--------</span>
<span class="n">placeholder</span> <span class="n">x</span> <span class="n">x</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">placeholder</span> <span class="n">y</span> <span class="n">y</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">cos</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">method</span> <span class="n">cos</span> <span class="n">of</span> <span class="nb">type</span> <span class="nb">object</span> <span class="n">at</span> <span class="mh">0x7f1a894649a8</span><span class="o">></span> <span class="p">(</span><span class="n">x</span><span class="p">,)</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">sin</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">method</span> <span class="n">sin</span> <span class="n">of</span> <span class="nb">type</span> <span class="nb">object</span> <span class="n">at</span> <span class="mh">0x7f1a894649a8</span><span class="o">></span> <span class="p">(</span><span class="n">y</span><span class="p">,)</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">add</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">function</span> <span class="n">add</span><span class="o">></span> <span class="p">(</span><span class="n">cos</span><span class="p">,</span> <span class="n">sin</span><span class="p">)</span> <span class="p">{}</span>
<span class="n">output</span> <span class="n">output</span> <span class="n">output</span> <span class="p">((</span><span class="n">add</span><span class="p">,),)</span> <span class="p">{}</span>
</pre></div>
</div>
<p>This works for <code class="docutils literal notranslate"><span class="pre">torch.nn.Module</span></code> as well as shown below:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">List</span>
<span class="kn">import</span> <span class="nn">torch</span>
<span class="k">def</span> <span class="nf">my_compiler</span><span class="p">(</span><span class="n">gm</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">:</span> <span class="n">List</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="nb">print</span><span class="p">(</span><span class="s2">"my_compiler() called with FX graph:"</span><span class="p">)</span>
<span class="n">gm</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">print_tabular</span><span class="p">()</span>
<span class="k">return</span> <span class="n">gm</span><span class="o">.</span><span class="n">forward</span> <span class="c1"># return a python callable</span>
<span class="k">class</span> <span class="nc">MockModule</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">relu</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">cos</span><span class="p">(</span><span class="n">x</span><span class="p">))</span>
<span class="n">mod</span> <span class="o">=</span> <span class="n">MockModule</span><span class="p">()</span>
<span class="n">optimized_mod</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">mod</span><span class="p">,</span> <span class="n">backend</span><span class="o">=</span><span class="n">my_compiler</span><span class="p">)</span>
<span class="n">optimized_mod</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">10</span><span class="p">))</span>
</pre></div>
</div>
<p>Let’s take a look at one more example with control flow:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">List</span>
<span class="kn">import</span> <span class="nn">torch</span>
<span class="k">def</span> <span class="nf">my_compiler</span><span class="p">(</span><span class="n">gm</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">:</span> <span class="n">List</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="nb">print</span><span class="p">(</span><span class="s2">"my_compiler() called with FX graph:"</span><span class="p">)</span>
<span class="n">gm</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">print_tabular</span><span class="p">()</span>
<span class="k">return</span> <span class="n">gm</span><span class="o">.</span><span class="n">forward</span> <span class="c1"># return a python callable</span>
<span class="nd">@torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">backend</span><span class="o">=</span><span class="n">my_compiler</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">toy_example</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">a</span> <span class="o">/</span> <span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">abs</span><span class="p">(</span><span class="n">a</span><span class="p">)</span> <span class="o">+</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">if</span> <span class="n">b</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o"><</span> <span class="mi">0</span><span class="p">:</span>
<span class="n">b</span> <span class="o">=</span> <span class="n">b</span> <span class="o">*</span> <span class="o">-</span><span class="mi">1</span>
<span class="k">return</span> <span class="n">x</span> <span class="o">*</span> <span class="n">b</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="mi">100</span><span class="p">):</span>
<span class="n">toy_example</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">10</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">10</span><span class="p">))</span>
</pre></div>
</div>
<p>Running this example produces the following output:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">my_compiler</span><span class="p">()</span> <span class="n">called</span> <span class="k">with</span> <span class="n">FX</span> <span class="n">graph</span><span class="p">:</span>
<span class="n">opcode</span> <span class="n">name</span> <span class="n">target</span> <span class="n">args</span> <span class="n">kwargs</span>
<span class="o">-------------</span> <span class="o">-------</span> <span class="o">------------------------------------------------------</span> <span class="o">----------------</span> <span class="o">--------</span>
<span class="n">placeholder</span> <span class="n">a</span> <span class="n">a</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">placeholder</span> <span class="n">b</span> <span class="n">b</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">abs_1</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">method</span> <span class="nb">abs</span> <span class="n">of</span> <span class="nb">type</span> <span class="nb">object</span> <span class="n">at</span> <span class="mh">0x7f8d259298a0</span><span class="o">></span> <span class="p">(</span><span class="n">a</span><span class="p">,)</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">add</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">function</span> <span class="n">add</span><span class="o">></span> <span class="p">(</span><span class="n">abs_1</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">truediv</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">function</span> <span class="n">truediv</span><span class="o">></span> <span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">add</span><span class="p">)</span> <span class="p">{}</span>
<span class="n">call_method</span> <span class="n">sum_1</span> <span class="nb">sum</span> <span class="p">(</span><span class="n">b</span><span class="p">,)</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">lt</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">function</span> <span class="n">lt</span><span class="o">></span> <span class="p">(</span><span class="n">sum_1</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span> <span class="p">{}</span>
<span class="n">output</span> <span class="n">output</span> <span class="n">output</span> <span class="p">((</span><span class="n">truediv</span><span class="p">,</span> <span class="n">lt</span><span class="p">),)</span> <span class="p">{}</span>
<span class="n">my_compiler</span><span class="p">()</span> <span class="n">called</span> <span class="k">with</span> <span class="n">FX</span> <span class="n">graph</span><span class="p">:</span>
<span class="n">opcode</span> <span class="n">name</span> <span class="n">target</span> <span class="n">args</span> <span class="n">kwargs</span>
<span class="o">-------------</span> <span class="o">------</span> <span class="o">-----------------------</span> <span class="o">-----------</span> <span class="o">--------</span>
<span class="n">placeholder</span> <span class="n">b</span> <span class="n">b</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">placeholder</span> <span class="n">x</span> <span class="n">x</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">mul</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">function</span> <span class="n">mul</span><span class="o">></span> <span class="p">(</span><span class="n">b</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">)</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">mul_1</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">function</span> <span class="n">mul</span><span class="o">></span> <span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">mul</span><span class="p">)</span> <span class="p">{}</span>
<span class="n">output</span> <span class="n">output</span> <span class="n">output</span> <span class="p">((</span><span class="n">mul_1</span><span class="p">,),)</span> <span class="p">{}</span>
<span class="n">my_compiler</span><span class="p">()</span> <span class="n">called</span> <span class="k">with</span> <span class="n">FX</span> <span class="n">graph</span><span class="p">:</span>
<span class="n">opcode</span> <span class="n">name</span> <span class="n">target</span> <span class="n">args</span> <span class="n">kwargs</span>
<span class="o">-------------</span> <span class="o">------</span> <span class="o">-----------------------</span> <span class="o">---------</span> <span class="o">--------</span>
<span class="n">placeholder</span> <span class="n">b</span> <span class="n">b</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">placeholder</span> <span class="n">x</span> <span class="n">x</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">mul</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">function</span> <span class="n">mul</span><span class="o">></span> <span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">b</span><span class="p">)</span> <span class="p">{}</span>
<span class="n">output</span> <span class="n">output</span> <span class="n">output</span> <span class="p">((</span><span class="n">mul</span><span class="p">,),)</span> <span class="p">{}</span>
</pre></div>
</div>
<p>The order of the last two graphs is nondeterministic depending
on which one is encountered first by the just-in-time compiler.</p>
</section>
<section id="speedy-backend">
<h3>Speedy Backend<a class="headerlink" href="#speedy-backend" title="Permalink to this heading">¶</a></h3>
<p>Integrating a custom backend that offers superior performance is also
easy and we’ll integrate a real one
with <a class="reference external" href="https://pytorch.org/docs/stable/generated/torch.jit.optimize_for_inference.html">optimize_for_inference</a>:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">optimize_for_inference_compiler</span><span class="p">(</span><span class="n">gm</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">]):</span>
<span class="n">scripted</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">script</span><span class="p">(</span><span class="n">gm</span><span class="p">)</span>
<span class="k">return</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">optimize_for_inference</span><span class="p">(</span><span class="n">scripted</span><span class="p">)</span>
</pre></div>
</div>
<p>And then you should be able to optimize any existing code with:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="nd">@torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">backend</span><span class="o">=</span><span class="n">optimize_for_inference_compiler</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">code_to_accelerate</span><span class="p">():</span>
<span class="o">...</span>
</pre></div>
</div>
</section>
<section id="composable-backends">
<h3>Composable Backends<a class="headerlink" href="#composable-backends" title="Permalink to this heading">¶</a></h3>
<p>TorchDynamo includes many backends, which can be found in
<a class="reference external" href="https://github.com/pytorch/pytorch/blob/master/torch/_dynamo/optimizations/backends.py">backends.py</a>
or <code class="docutils literal notranslate"><span class="pre">torch._dynamo.list_backends()</span></code>. You can combine these backends
together with the following code:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">torch._dynamo.optimizations</span> <span class="kn">import</span> <span class="n">BACKENDS</span>
<span class="k">def</span> <span class="nf">my_compiler</span><span class="p">(</span><span class="n">gm</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">:</span> <span class="n">List</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">try</span><span class="p">:</span>
<span class="n">trt_compiled</span> <span class="o">=</span> <span class="n">BACKENDS</span><span class="p">[</span><span class="s2">"tensorrt"</span><span class="p">](</span><span class="n">gm</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">)</span>
<span class="k">if</span> <span class="n">trt_compiled</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">return</span> <span class="n">trt_compiled</span>
<span class="k">except</span> <span class="ne">Exception</span><span class="p">:</span>
<span class="k">pass</span>
<span class="c1"># first backend failed, try something else...</span>
<span class="k">try</span><span class="p">:</span>
<span class="n">inductor_compiled</span> <span class="o">=</span> <span class="n">BACKENDS</span><span class="p">[</span><span class="s2">"inductor"</span><span class="p">](</span><span class="n">gm</span><span class="p">,</span> <span class="n">example_inputs</span><span class="p">)</span>
<span class="k">if</span> <span class="n">inductor_compiled</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">return</span> <span class="n">inductor_compiled</span>
<span class="k">except</span> <span class="ne">Exception</span><span class="p">:</span>
<span class="k">pass</span>
<span class="k">return</span> <span class="n">gm</span><span class="o">.</span><span class="n">forward</span>
</pre></div>
</div>
</section>
</section>
</section>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="deep-dive.html" class="btn btn-neutral float-right" title="TorchDynamo Deeper Dive" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="guards-overview.html" class="btn btn-neutral" title="Guards Overview" 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="#">Custom Backends</a><ul>
<li><a class="reference internal" href="#overview">Overview</a></li>
<li><a class="reference internal" href="#registering-custom-backends">Registering Custom Backends</a></li>
<li><a class="reference internal" href="#custom-backends-after-aotautograd">Custom Backends after AOTAutograd</a></li>
<li><a class="reference internal" href="#examples">Examples</a><ul>
<li><a class="reference internal" href="#debugging-backend">Debugging Backend</a></li>
<li><a class="reference internal" href="#speedy-backend">Speedy Backend</a></li>
<li><a class="reference internal" href="#composable-backends">Composable Backends</a></li>
</ul>
</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/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="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="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="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="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>