-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathcpp_extension.html
1448 lines (1263 loc) · 107 KB
/
cpp_extension.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>torch.utils.cpp_extension — PyTorch 2.7 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/cpp_extension.html"/>
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="_static/sphinx-dropdown.css" type="text/css" />
<link rel="stylesheet" href="_static/panels-bootstrap.min.css" type="text/css" />
<link rel="stylesheet" href="_static/css/jit.css" type="text/css" />
<link rel="stylesheet" href="_static/css/custom.css" type="text/css" />
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="torch.utils.data" href="data.html" />
<link rel="prev" title="torch.utils.checkpoint" href="checkpoint.html" />
<!-- Google Tag Manager -->
<script>(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':
new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],
j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src=
'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);
})(window,document,'script','dataLayer','GTM-T8XT4PS');</script>
<!-- End Google Tag Manager -->
<script src="_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Learn
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/get-started">
<span class=dropdown-title>Get Started</span>
<p>Run PyTorch locally or get started quickly with one of the supported cloud platforms</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials">
<span class="dropdown-title">Tutorials</span>
<p>Whats new in PyTorch tutorials</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/basics/intro.html">
<span class="dropdown-title">Learn the Basics</span>
<p>Familiarize yourself with PyTorch concepts and modules</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/recipes/recipes_index.html">
<span class="dropdown-title">PyTorch Recipes</span>
<p>Bite-size, ready-to-deploy PyTorch code examples</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/introyt.html">
<span class="dropdown-title">Intro to PyTorch - YouTube Series</span>
<p>Master PyTorch basics with our engaging YouTube tutorial series</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Ecosystem
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem">
<span class="dropdown-title">Tools</span>
<p>Learn about the tools and frameworks in the PyTorch Ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class=dropdown-title>Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class=dropdown-title>Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class=dropdown-title>Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem/contributor-awards-2024">
<span class="dropdown-title">Contributor Awards - 2024</span>
<p>Award winners announced at this year's PyTorch Conference</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Edge
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/edge">
<span class="dropdown-title">About PyTorch Edge</span>
<p>Build innovative and privacy-aware AI experiences for edge devices</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch-overview">
<span class="dropdown-title">ExecuTorch</span>
<p>End-to-end solution for enabling on-device inference capabilities across mobile and edge devices</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch/stable/index.html">
<span class="dropdown-title">ExecuTorch Docs</span>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p>Explore the documentation for comprehensive guidance on how to use PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/pytorch-domains">
<span class="dropdown-title">PyTorch Domains</span>
<p>Read the PyTorch Domains documentation to learn more about domain-specific libraries</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Blogs & News
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/blog/">
<span class="dropdown-title">PyTorch Blog</span>
<p>Catch up on the latest technical news and happenings</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/community-blog">
<span class="dropdown-title">Community Blog</span>
<p>Stories from the PyTorch ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/videos">
<span class="dropdown-title">Videos</span>
<p>Learn about the latest PyTorch tutorials, new, and more </p>
<a class="nav-dropdown-item" href="https://pytorch.org/community-stories">
<span class="dropdown-title">Community Stories</span>
<p>Learn how our community solves real, everyday machine learning problems with PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/newsletter">
<span class="dropdown-title">Newsletter</span>
<p>Stay up-to-date with the latest updates</p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
About
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn more about the PyTorch Foundation</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/governing-board">
<span class="dropdown-title">Governing Board</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/credits">
<span class="dropdown-title">Cloud Credit Program</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tac">
<span class="dropdown-title">Technical Advisory Council</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/staff">
<span class="dropdown-title">Staff</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/contact-us">
<span class="dropdown-title">Contact Us</span>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div class="no-dropdown">
<a href="https://pytorch.org/join" data-cta="join">
Become a Member
</a>
</div>
</li>
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="github-icon">
</a>
</div>
</li>
<!--- TODO: This block adds the search icon to the nav bar. We will enable it later.
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="search-icon">
</a>
</div>
</li>
--->
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>2.7 ▼</a>
</div>
<div id="searchBox">
<div class="searchbox" id="googleSearchBox">
<script async src="https://cse.google.com/cse.js?cx=e65585f8c3ea1440e"></script>
<div class="gcse-search"></div>
</div>
<div id="sphinxSearchBox" style="display: none;">
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
</div>
<form id="searchForm">
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="google" checked>
Google Search
</label>
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="sphinx">
Classic Search
</label>
</form>
<script>
document.addEventListener('DOMContentLoaded', function() {
const searchForm = document.getElementById('searchForm');
const googleSearchBox = document.getElementById('googleSearchBox');
const sphinxSearchBox = document.getElementById('sphinxSearchBox');
// Function to toggle search box visibility
function toggleSearchBox(searchType) {
googleSearchBox.style.display = searchType === 'google' ? 'block' : 'none';
sphinxSearchBox.style.display = searchType === 'sphinx' ? 'block' : 'none';
}
// Determine the default search type
let defaultSearchType;
const currentUrl = window.location.href;
if (currentUrl.startsWith('https://pytorch.org/docs/stable')) {
// For the stable documentation, default to Google
defaultSearchType = localStorage.getItem('searchType') || 'google';
} else {
// For any other version, including docs-preview, default to Sphinx
defaultSearchType = 'sphinx';
}
// Set the default search type
document.querySelector(`input[name="searchType"][value="${defaultSearchType}"]`).checked = true;
toggleSearchBox(defaultSearchType);
// Event listener for changes in search type
searchForm.addEventListener('change', function(event) {
const selectedSearchType = event.target.value;
localStorage.setItem('searchType', selectedSearchType);
toggleSearchBox(selectedSearchType);
});
// Set placeholder text for Google search box
window.onload = function() {
var placeholderText = "Search Docs";
var googleSearchboxText = document.querySelector("#gsc-i-id1");
if (googleSearchboxText) {
googleSearchboxText.placeholder = placeholderText;
googleSearchboxText.style.fontFamily = 'FreightSans';
googleSearchboxText.style.fontSize = "1.2rem";
googleSearchboxText.style.color = '#262626';
}
};
});
</script>
</div>
<p class="caption" role="heading"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="community/build_ci_governance.html">PyTorch Governance | Build + CI</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/design.html">PyTorch Design Philosophy</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/governance.html">PyTorch Governance | Mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/persons_of_interest.html">PyTorch Governance | Maintainers</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Developer Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="notes/amp_examples.html">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/custom_operators.html">PyTorch Custom Operators Landing Page</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.func.html">Extending torch.func with autograd.Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/fsdp.html">FSDP Notes</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/get_start_xpu.html">Getting Started on Intel GPU</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/libtorch_stable_abi.html">LibTorch Stable ABI</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/mps.html">MPS backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/numerical_accuracy.html">Numerical accuracy</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
<li class="toctree-l1"><a class="reference internal" href="deploy.html">torch::deploy</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python API</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="amp.html">torch.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="library.html">torch.library</a></li>
<li class="toctree-l1"><a class="reference internal" href="accelerator.html">torch.accelerator</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpu.html">torch.cpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html">Understanding CUDA Memory Usage</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#generating-a-snapshot">Generating a Snapshot</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#using-the-visualizer">Using the visualizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#snapshot-api-reference">Snapshot API Reference</a></li>
<li class="toctree-l1"><a class="reference internal" href="mps.html">torch.mps</a></li>
<li class="toctree-l1"><a class="reference internal" href="xpu.html">torch.xpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="mtia.html">torch.mtia</a></li>
<li class="toctree-l1"><a class="reference internal" href="mtia.memory.html">torch.mtia.memory</a></li>
<li class="toctree-l1"><a class="reference internal" href="meta.html">Meta device</a></li>
<li class="toctree-l1"><a class="reference internal" href="backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="export.html">torch.export</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.tensor.html">torch.distributed.tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="fsdp.html">torch.distributed.fsdp</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.fsdp.fully_shard.html">torch.distributed.fsdp.fully_shard</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.tensor.parallel.html">torch.distributed.tensor.parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.pipelining.html">torch.distributed.pipelining</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.checkpoint.html">torch.distributed.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.compiler.html">torch.compiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="func.html">torch.func</a></li>
<li class="toctree-l1"><a class="reference internal" href="futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.experimental.html">torch.fx.experimental</a></li>
<li class="toctree-l1"><a class="reference internal" href="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="monitor.html">torch.monitor</a></li>
<li class="toctree-l1"><a class="reference internal" href="signal.html">torch.signal</a></li>
<li class="toctree-l1"><a class="reference internal" href="special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.attention.html">torch.nn.attention</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="masked.html">torch.masked</a></li>
<li class="toctree-l1"><a class="reference internal" href="nested.html">torch.nested</a></li>
<li class="toctree-l1"><a class="reference internal" href="size.html">torch.Size</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="utils.html">torch.utils</a></li>
<li class="toctree-l1"><a class="reference internal" href="benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="deterministic.html">torch.utils.deterministic</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit_utils.html">torch.utils.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="module_tracker.html">torch.utils.module_tracker</a></li>
<li class="toctree-l1"><a class="reference internal" href="type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="config_mod.html">torch.__config__</a></li>
<li class="toctree-l1"><a class="reference internal" href="future_mod.html">torch.__future__</a></li>
<li class="toctree-l1"><a class="reference internal" href="logging.html">torch._logging</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_environment_variables.html">Torch Environment Variables</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/data">TorchData</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/torchrec">TorchRec</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/ao">torchao</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="index.html">
Docs
</a> >
</li>
<li>torch.utils.cpp_extension</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/cpp_extension.rst.txt" rel="nofollow"><img src="_static/images/view-page-source-icon.svg"></a>
</li>
</ul>
</div>
</div>
<div class="pytorch-shortcuts-wrapper" id="pytorch-shortcuts-wrapper">
Shortcuts
</div>
</div>
<section data-toggle="wy-nav-shift" id="pytorch-content-wrap" class="pytorch-content-wrap">
<div class="pytorch-content-left">
<!-- Google Tag Manager (noscript) -->
<noscript><iframe src="https://www.googletagmanager.com/ns.html?id=GTM-T8XT4PS"
height="0" width="0" style="display:none;visibility:hidden"></iframe></noscript>
<!-- End Google Tag Manager (noscript) -->
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<div class="section" id="torch-utils-cpp-extension">
<h1>torch.utils.cpp_extension<a class="headerlink" href="#torch-utils-cpp-extension" title="Permalink to this heading">¶</a></h1>
<dl class="py function">
<dt class="sig sig-object py" id="torch.utils.cpp_extension.CppExtension">
<span class="sig-prename descclassname"><span class="pre">torch.utils.cpp_extension.</span></span><span class="sig-name descname"><span class="pre">CppExtension</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">name</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">sources</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/cpp_extension.html#CppExtension"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/utils/cpp_extension.py#L1089"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.cpp_extension.CppExtension" title="Permalink to this definition">¶</a></dt>
<dd><p>Create a <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.Extension</span></code> for C++.</p>
<p>Convenience method that creates a <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.Extension</span></code> with the
bare minimum (but often sufficient) arguments to build a C++ extension.</p>
<p>All arguments are forwarded to the <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.Extension</span></code>
constructor. Full list arguments can be found at
<a class="reference external" href="https://setuptools.pypa.io/en/latest/userguide/ext_modules.html#extension-api-reference">https://setuptools.pypa.io/en/latest/userguide/ext_modules.html#extension-api-reference</a></p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>The PyTorch python API (as provided in libtorch_python) cannot be built
with the flag <code class="docutils literal notranslate"><span class="pre">py_limited_api=True</span></code>. When this flag is passed, it is
the user’s responsibility in their library to not use APIs from
libtorch_python (in particular pytorch/python bindings) and to only use
APIs from libtorch (aten objects, operators and the dispatcher). For
example, to give access to custom ops from python, the library should
register the ops through the dispatcher.</p>
<p>Contrary to CPython setuptools, who does not define -DPy_LIMITED_API
as a compile flag when py_limited_api is specified as an option for
the “bdist_wheel” command in <code class="docutils literal notranslate"><span class="pre">setup</span></code>, PyTorch does! We will specify
-DPy_LIMITED_API=min_supported_cpython to best enforce consistency,
safety, and sanity in order to encourage best practices. To target a
different version, set min_supported_cpython to the hexcode of the
CPython version of choice.</p>
</div>
<p class="rubric">Example</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">setuptools</span> <span class="kn">import</span> <span class="n">setup</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.utils.cpp_extension</span> <span class="kn">import</span> <span class="n">BuildExtension</span><span class="p">,</span> <span class="n">CppExtension</span>
<span class="gp">>>> </span><span class="n">setup</span><span class="p">(</span>
<span class="gp">... </span> <span class="n">name</span><span class="o">=</span><span class="s1">'extension'</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">ext_modules</span><span class="o">=</span><span class="p">[</span>
<span class="gp">... </span> <span class="n">CppExtension</span><span class="p">(</span>
<span class="gp">... </span> <span class="n">name</span><span class="o">=</span><span class="s1">'extension'</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">sources</span><span class="o">=</span><span class="p">[</span><span class="s1">'extension.cpp'</span><span class="p">],</span>
<span class="gp">... </span> <span class="n">extra_compile_args</span><span class="o">=</span><span class="p">[</span><span class="s1">'-g'</span><span class="p">],</span>
<span class="gp">... </span> <span class="n">extra_link_args</span><span class="o">=</span><span class="p">[</span><span class="s1">'-Wl,--no-as-needed'</span><span class="p">,</span> <span class="s1">'-lm'</span><span class="p">])</span>
<span class="gp">... </span> <span class="p">],</span>
<span class="gp">... </span> <span class="n">cmdclass</span><span class="o">=</span><span class="p">{</span>
<span class="gp">... </span> <span class="s1">'build_ext'</span><span class="p">:</span> <span class="n">BuildExtension</span>
<span class="gp">... </span> <span class="p">})</span>
</pre></div>
</div>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.utils.cpp_extension.CUDAExtension">
<span class="sig-prename descclassname"><span class="pre">torch.utils.cpp_extension.</span></span><span class="sig-name descname"><span class="pre">CUDAExtension</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">name</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">sources</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/cpp_extension.html#CUDAExtension"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/utils/cpp_extension.py#L1159"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.cpp_extension.CUDAExtension" title="Permalink to this definition">¶</a></dt>
<dd><p>Create a <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.Extension</span></code> for CUDA/C++.</p>
<p>Convenience method that creates a <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.Extension</span></code> with the
bare minimum (but often sufficient) arguments to build a CUDA/C++
extension. This includes the CUDA include path, library path and runtime
library.</p>
<p>All arguments are forwarded to the <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.Extension</span></code>
constructor. Full list arguments can be found at
<a class="reference external" href="https://setuptools.pypa.io/en/latest/userguide/ext_modules.html#extension-api-reference">https://setuptools.pypa.io/en/latest/userguide/ext_modules.html#extension-api-reference</a></p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>The PyTorch python API (as provided in libtorch_python) cannot be built
with the flag <code class="docutils literal notranslate"><span class="pre">py_limited_api=True</span></code>. When this flag is passed, it is
the user’s responsibility in their library to not use APIs from
libtorch_python (in particular pytorch/python bindings) and to only use
APIs from libtorch (aten objects, operators and the dispatcher). For
example, to give access to custom ops from python, the library should
register the ops through the dispatcher.</p>
<p>Contrary to CPython setuptools, who does not define -DPy_LIMITED_API
as a compile flag when py_limited_api is specified as an option for
the “bdist_wheel” command in <code class="docutils literal notranslate"><span class="pre">setup</span></code>, PyTorch does! We will specify
-DPy_LIMITED_API=min_supported_cpython to best enforce consistency,
safety, and sanity in order to encourage best practices. To target a
different version, set min_supported_cpython to the hexcode of the
CPython version of choice.</p>
</div>
<p class="rubric">Example</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">setuptools</span> <span class="kn">import</span> <span class="n">setup</span>
<span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.utils.cpp_extension</span> <span class="kn">import</span> <span class="n">BuildExtension</span><span class="p">,</span> <span class="n">CUDAExtension</span>
<span class="gp">>>> </span><span class="n">setup</span><span class="p">(</span>
<span class="gp">... </span> <span class="n">name</span><span class="o">=</span><span class="s1">'cuda_extension'</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">ext_modules</span><span class="o">=</span><span class="p">[</span>
<span class="gp">... </span> <span class="n">CUDAExtension</span><span class="p">(</span>
<span class="gp">... </span> <span class="n">name</span><span class="o">=</span><span class="s1">'cuda_extension'</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">sources</span><span class="o">=</span><span class="p">[</span><span class="s1">'extension.cpp'</span><span class="p">,</span> <span class="s1">'extension_kernel.cu'</span><span class="p">],</span>
<span class="gp">... </span> <span class="n">extra_compile_args</span><span class="o">=</span><span class="p">{</span><span class="s1">'cxx'</span><span class="p">:</span> <span class="p">[</span><span class="s1">'-g'</span><span class="p">],</span>
<span class="gp">... </span> <span class="s1">'nvcc'</span><span class="p">:</span> <span class="p">[</span><span class="s1">'-O2'</span><span class="p">]},</span>
<span class="gp">... </span> <span class="n">extra_link_args</span><span class="o">=</span><span class="p">[</span><span class="s1">'-Wl,--no-as-needed'</span><span class="p">,</span> <span class="s1">'-lcuda'</span><span class="p">])</span>
<span class="gp">... </span> <span class="p">],</span>
<span class="gp">... </span> <span class="n">cmdclass</span><span class="o">=</span><span class="p">{</span>
<span class="gp">... </span> <span class="s1">'build_ext'</span><span class="p">:</span> <span class="n">BuildExtension</span>
<span class="gp">... </span> <span class="p">})</span>
</pre></div>
</div>
<p>Compute capabilities:</p>
<p>By default the extension will be compiled to run on all archs of the cards visible during the
building process of the extension, plus PTX. If down the road a new card is installed the
extension may need to be recompiled. If a visible card has a compute capability (CC) that’s
newer than the newest version for which your nvcc can build fully-compiled binaries, PyTorch
will make nvcc fall back to building kernels with the newest version of PTX your nvcc does
support (see below for details on PTX).</p>
<p>You can override the default behavior using <cite>TORCH_CUDA_ARCH_LIST</cite> to explicitly specify which
CCs you want the extension to support:</p>
<p><code class="docutils literal notranslate"><span class="pre">TORCH_CUDA_ARCH_LIST="6.1</span> <span class="pre">8.6"</span> <span class="pre">python</span> <span class="pre">build_my_extension.py</span></code>
<code class="docutils literal notranslate"><span class="pre">TORCH_CUDA_ARCH_LIST="5.2</span> <span class="pre">6.0</span> <span class="pre">6.1</span> <span class="pre">7.0</span> <span class="pre">7.5</span> <span class="pre">8.0</span> <span class="pre">8.6+PTX"</span> <span class="pre">python</span> <span class="pre">build_my_extension.py</span></code></p>
<p>The +PTX option causes extension kernel binaries to include PTX instructions for the specified
CC. PTX is an intermediate representation that allows kernels to runtime-compile for any CC >=
the specified CC (for example, 8.6+PTX generates PTX that can runtime-compile for any GPU with
CC >= 8.6). This improves your binary’s forward compatibility. However, relying on older PTX to
provide forward compat by runtime-compiling for newer CCs can modestly reduce performance on
those newer CCs. If you know exact CC(s) of the GPUs you want to target, you’re always better
off specifying them individually. For example, if you want your extension to run on 8.0 and 8.6,
“8.0+PTX” would work functionally because it includes PTX that can runtime-compile for 8.6, but
“8.0 8.6” would be better.</p>
<p>Note that while it’s possible to include all supported archs, the more archs get included the
slower the building process will be, as it will build a separate kernel image for each arch.</p>
<p>Note that CUDA-11.5 nvcc will hit internal compiler error while parsing torch/extension.h on Windows.
To workaround the issue, move python binding logic to pure C++ file.</p>
<dl class="simple">
<dt>Example use:</dt><dd><p>#include <ATen/ATen.h>
at::Tensor SigmoidAlphaBlendForwardCuda(….)</p>
</dd>
<dt>Instead of:</dt><dd><p>#include <torch/extension.h>
torch::Tensor SigmoidAlphaBlendForwardCuda(…)</p>
</dd>
</dl>
<p>Currently open issue for nvcc bug: <a class="reference external" href="https://github.com/pytorch/pytorch/issues/69460">https://github.com/pytorch/pytorch/issues/69460</a>
Complete workaround code example: <a class="reference external" href="https://github.com/facebookresearch/pytorch3d/commit/cb170ac024a949f1f9614ffe6af1c38d972f7d48">https://github.com/facebookresearch/pytorch3d/commit/cb170ac024a949f1f9614ffe6af1c38d972f7d48</a></p>
<p>Relocatable device code linking:</p>
<p>If you want to reference device symbols across compilation units (across object files),
the object files need to be built with <cite>relocatable device code</cite> (-rdc=true or -dc).
An exception to this rule is “dynamic parallelism” (nested kernel launches) which is not used a lot anymore.
<cite>Relocatable device code</cite> is less optimized so it needs to be used only on object files that need it.
Using <cite>-dlto</cite> (Device Link Time Optimization) at the device code compilation step and <cite>dlink</cite> step
helps reduce the protentional perf degradation of <cite>-rdc</cite>.
Note that it needs to be used at both steps to be useful.</p>
<p>If you have <cite>rdc</cite> objects you need to have an extra <cite>-dlink</cite> (device linking) step before the CPU symbol linking step.
There is also a case where <cite>-dlink</cite> is used without <cite>-rdc</cite>:
when an extension is linked against a static lib containing rdc-compiled objects
like the [NVSHMEM library](<a class="reference external" href="https://developer.nvidia.com/nvshmem">https://developer.nvidia.com/nvshmem</a>).</p>
<p>Note: Ninja is required to build a CUDA Extension with RDC linking.</p>
<p class="rubric">Example</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">CUDAExtension</span><span class="p">(</span>
<span class="gp">... </span> <span class="n">name</span><span class="o">=</span><span class="s1">'cuda_extension'</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">sources</span><span class="o">=</span><span class="p">[</span><span class="s1">'extension.cpp'</span><span class="p">,</span> <span class="s1">'extension_kernel.cu'</span><span class="p">],</span>
<span class="gp">... </span> <span class="n">dlink</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">dlink_libraries</span><span class="o">=</span><span class="p">[</span><span class="s2">"dlink_lib"</span><span class="p">],</span>
<span class="gp">... </span> <span class="n">extra_compile_args</span><span class="o">=</span><span class="p">{</span><span class="s1">'cxx'</span><span class="p">:</span> <span class="p">[</span><span class="s1">'-g'</span><span class="p">],</span>
<span class="gp">... </span> <span class="s1">'nvcc'</span><span class="p">:</span> <span class="p">[</span><span class="s1">'-O2'</span><span class="p">,</span> <span class="s1">'-rdc=true'</span><span class="p">]})</span>
</pre></div>
</div>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.utils.cpp_extension.SyclExtension">
<span class="sig-prename descclassname"><span class="pre">torch.utils.cpp_extension.</span></span><span class="sig-name descname"><span class="pre">SyclExtension</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">name</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">sources</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/cpp_extension.html#SyclExtension"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/utils/cpp_extension.py#L1350"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.cpp_extension.SyclExtension" title="Permalink to this definition">¶</a></dt>
<dd><p>Creates a <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.Extension</span></code> for SYCL/C++.</p>
<p>Convenience method that creates a <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.Extension</span></code> with the
bare minimum (but often sufficient) arguments to build a SYCL/C++
extension.</p>
<p>All arguments are forwarded to the <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.Extension</span></code>
constructor.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>The PyTorch python API (as provided in libtorch_python) cannot be built
with the flag <code class="docutils literal notranslate"><span class="pre">py_limited_api=True</span></code>. When this flag is passed, it is
the user’s responsibility in their library to not use APIs from
libtorch_python (in particular pytorch/python bindings) and to only use
APIs from libtorch (aten objects, operators and the dispatcher). For
example, to give access to custom ops from python, the library should
register the ops through the dispatcher.</p>
<p>Contrary to CPython setuptools, who does not define -DPy_LIMITED_API
as a compile flag when py_limited_api is specified as an option for
the “bdist_wheel” command in <code class="docutils literal notranslate"><span class="pre">setup</span></code>, PyTorch does! We will specify
-DPy_LIMITED_API=min_supported_cpython to best enforce consistency,
safety, and sanity in order to encourage best practices. To target a
different version, set min_supported_cpython to the hexcode of the
CPython version of choice.</p>
</div>
<p class="rubric">Example</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.utils.cpp_extension</span> <span class="kn">import</span> <span class="n">BuildExtension</span><span class="p">,</span> <span class="n">SyclExtension</span>
<span class="gp">>>> </span><span class="n">setup</span><span class="p">(</span>
<span class="gp">... </span> <span class="n">name</span><span class="o">=</span><span class="s1">'xpu_extension'</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">ext_modules</span><span class="o">=</span><span class="p">[</span>
<span class="gp">... </span> <span class="n">SyclExtension</span><span class="p">(</span>
<span class="gp">... </span> <span class="n">name</span><span class="o">=</span><span class="s1">'xpu_extension'</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">sources</span><span class="o">=</span><span class="p">[</span><span class="s1">'extension.cpp'</span><span class="p">,</span> <span class="s1">'extension_kernel.cpp'</span><span class="p">],</span>
<span class="gp">... </span> <span class="n">extra_compile_args</span><span class="o">=</span><span class="p">{</span><span class="s1">'cxx'</span><span class="p">:</span> <span class="p">[</span><span class="s1">'-g'</span><span class="p">,</span> <span class="s1">'-std=c++20'</span><span class="p">,</span> <span class="s1">'-fPIC'</span><span class="p">]})</span>
<span class="gp">... </span> <span class="p">],</span>
<span class="gp">... </span> <span class="n">cmdclass</span><span class="o">=</span><span class="p">{</span>
<span class="gp">... </span> <span class="s1">'build_ext'</span><span class="p">:</span> <span class="n">BuildExtension</span>
<span class="gp">... </span> <span class="p">})</span>
</pre></div>
</div>
<p>By default the extension will be compiled to run on all archs of the cards visible during the
building process of the extension. If down the road a new card is installed the
extension may need to be recompiled. You can override the default behavior using
<cite>TORCH_XPU_ARCH_LIST</cite> to explicitly specify which device architectures you want the extension
to support:</p>
<p><code class="docutils literal notranslate"><span class="pre">TORCH_XPU_ARCH_LIST="pvc,xe-lpg"</span> <span class="pre">python</span> <span class="pre">build_my_extension.py</span></code></p>
<p>Note that while it’s possible to include all supported archs, the more archs get included the
slower the building process will be, as it will build a separate kernel image for each arch.</p>
<p>Note: Ninja is required to build SyclExtension.</p>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.utils.cpp_extension.BuildExtension">
<span class="sig-prename descclassname"><span class="pre">torch.utils.cpp_extension.</span></span><span class="sig-name descname"><span class="pre">BuildExtension</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/cpp_extension.html#BuildExtension"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/utils/cpp_extension.py#L529"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.cpp_extension.BuildExtension" title="Permalink to this definition">¶</a></dt>
<dd><p>A custom <code class="xref py py-mod docutils literal notranslate"><span class="pre">setuptools</span></code> build extension .</p>
<p>This <code class="xref py py-class docutils literal notranslate"><span class="pre">setuptools.build_ext</span></code> subclass takes care of passing the
minimum required compiler flags (e.g. <code class="docutils literal notranslate"><span class="pre">-std=c++17</span></code>) as well as mixed
C++/CUDA/SYCL compilation (and support for CUDA/SYCL files in general).</p>
<p>When using <a class="reference internal" href="#torch.utils.cpp_extension.BuildExtension" title="torch.utils.cpp_extension.BuildExtension"><code class="xref py py-class docutils literal notranslate"><span class="pre">BuildExtension</span></code></a>, it is allowed to supply a dictionary
for <code class="docutils literal notranslate"><span class="pre">extra_compile_args</span></code> (rather than the usual list) that maps from
languages/compilers (the only expected values are <code class="docutils literal notranslate"><span class="pre">cxx</span></code>, <code class="docutils literal notranslate"><span class="pre">nvcc</span></code> or
<code class="docutils literal notranslate"><span class="pre">sycl</span></code>) to a list of additional compiler flags to supply to the compiler.
This makes it possible to supply different flags to the C++, CUDA and SYCL
compiler during mixed compilation.</p>
<p><code class="docutils literal notranslate"><span class="pre">use_ninja</span></code> (bool): If <code class="docutils literal notranslate"><span class="pre">use_ninja</span></code> is <code class="docutils literal notranslate"><span class="pre">True</span></code> (default), then we
attempt to build using the Ninja backend. Ninja greatly speeds up
compilation compared to the standard <code class="docutils literal notranslate"><span class="pre">setuptools.build_ext</span></code>.
Fallbacks to the standard distutils backend if Ninja is not available.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>By default, the Ninja backend uses #CPUS + 2 workers to build the
extension. This may use up too many resources on some systems. One
can control the number of workers by setting the <cite>MAX_JOBS</cite> environment
variable to a non-negative number.</p>
</div>
<dl class="field-list simple">
</dl>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.utils.cpp_extension.load">
<span class="sig-prename descclassname"><span class="pre">torch.utils.cpp_extension.</span></span><span class="sig-name descname"><span class="pre">load</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">name</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">sources</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_cflags</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_cuda_cflags</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_sycl_cflags</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_ldflags</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_include_paths</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">build_directory</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">verbose</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">with_cuda</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">with_sycl</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">is_python_module</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">is_standalone</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">keep_intermediates</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/cpp_extension.html#load"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/utils/cpp_extension.py#L1514"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.cpp_extension.load" title="Permalink to this definition">¶</a></dt>
<dd><p>Load a PyTorch C++ extension just-in-time (JIT).</p>
<p>To load an extension, a Ninja build file is emitted, which is used to
compile the given sources into a dynamic library. This library is
subsequently loaded into the current Python process as a module and
returned from this function, ready for use.</p>
<p>By default, the directory to which the build file is emitted and the
resulting library compiled to is <code class="docutils literal notranslate"><span class="pre"><tmp>/torch_extensions/<name></span></code>, where
<code class="docutils literal notranslate"><span class="pre"><tmp></span></code> is the temporary folder on the current platform and <code class="docutils literal notranslate"><span class="pre"><name></span></code>
the name of the extension. This location can be overridden in two ways.
First, if the <code class="docutils literal notranslate"><span class="pre">TORCH_EXTENSIONS_DIR</span></code> environment variable is set, it
replaces <code class="docutils literal notranslate"><span class="pre"><tmp>/torch_extensions</span></code> and all extensions will be compiled
into subfolders of this directory. Second, if the <code class="docutils literal notranslate"><span class="pre">build_directory</span></code>
argument to this function is supplied, it overrides the entire path, i.e.
the library will be compiled into that folder directly.</p>
<p>To compile the sources, the default system compiler (<code class="docutils literal notranslate"><span class="pre">c++</span></code>) is used,
which can be overridden by setting the <code class="docutils literal notranslate"><span class="pre">CXX</span></code> environment variable. To pass
additional arguments to the compilation process, <code class="docutils literal notranslate"><span class="pre">extra_cflags</span></code> or
<code class="docutils literal notranslate"><span class="pre">extra_ldflags</span></code> can be provided. For example, to compile your extension
with optimizations, pass <code class="docutils literal notranslate"><span class="pre">extra_cflags=['-O3']</span></code>. You can also use
<code class="docutils literal notranslate"><span class="pre">extra_cflags</span></code> to pass further include directories.</p>
<p>CUDA support with mixed compilation is provided. Simply pass CUDA source
files (<code class="docutils literal notranslate"><span class="pre">.cu</span></code> or <code class="docutils literal notranslate"><span class="pre">.cuh</span></code>) along with other sources. Such files will be
detected and compiled with nvcc rather than the C++ compiler. This includes
passing the CUDA lib64 directory as a library directory, and linking
<code class="docutils literal notranslate"><span class="pre">cudart</span></code>. You can pass additional flags to nvcc via
<code class="docutils literal notranslate"><span class="pre">extra_cuda_cflags</span></code>, just like with <code class="docutils literal notranslate"><span class="pre">extra_cflags</span></code> for C++. Various
heuristics for finding the CUDA install directory are used, which usually
work fine. If not, setting the <code class="docutils literal notranslate"><span class="pre">CUDA_HOME</span></code> environment variable is the
safest option.</p>
<p>SYCL support with mixed compilation is provided. Simply pass SYCL source
files (<code class="docutils literal notranslate"><span class="pre">.sycl</span></code>) along with other sources. Such files will be detected
and compiled with SYCL compiler (such as Intel DPC++ Compiler) rather
than the C++ compiler. You can pass additional flags to SYCL compiler
via <code class="docutils literal notranslate"><span class="pre">extra_sycl_cflags</span></code>, just like with <code class="docutils literal notranslate"><span class="pre">extra_cflags</span></code> for C++.
SYCL compiler is expected to be found via system PATH environment
variable.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>name</strong> – The name of the extension to build. This MUST be the same as the
name of the pybind11 module!</p></li>
<li><p><strong>sources</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Union" title="(in Python v3.13)"><em>Union</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.13)"><em>str</em></a><em>, </em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#list" title="(in Python v3.13)"><em>list</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.13)"><em>str</em></a><em>]</em><em>]</em>) – A list of relative or absolute paths to C++ source files.</p></li>
<li><p><strong>extra_cflags</strong> – optional list of compiler flags to forward to the build.</p></li>
<li><p><strong>extra_cuda_cflags</strong> – optional list of compiler flags to forward to nvcc
when building CUDA sources.</p></li>
<li><p><strong>extra_sycl_cflags</strong> – optional list of compiler flags to forward to SYCL
compiler when building SYCL sources.</p></li>
<li><p><strong>extra_ldflags</strong> – optional list of linker flags to forward to the build.</p></li>
<li><p><strong>extra_include_paths</strong> – optional list of include directories to forward
to the build.</p></li>
<li><p><strong>build_directory</strong> – optional path to use as build workspace.</p></li>
<li><p><strong>verbose</strong> – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, turns on verbose logging of load steps.</p></li>
<li><p><strong>with_cuda</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Optional" title="(in Python v3.13)"><em>Optional</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><em>bool</em></a><em>]</em>) – Determines whether CUDA headers and libraries are added to
the build. If set to <code class="docutils literal notranslate"><span class="pre">None</span></code> (default), this value is
automatically determined based on the existence of <code class="docutils literal notranslate"><span class="pre">.cu</span></code> or
<code class="docutils literal notranslate"><span class="pre">.cuh</span></code> in <code class="docutils literal notranslate"><span class="pre">sources</span></code>. Set it to <cite>True`</cite> to force CUDA headers
and libraries to be included.</p></li>
<li><p><strong>with_sycl</strong> (<a class="reference external" href="https://docs.python.org/3/library/typing.html#typing.Optional" title="(in Python v3.13)"><em>Optional</em></a><em>[</em><a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.13)"><em>bool</em></a><em>]</em>) – Determines whether SYCL headers and libraries are added to
the build. If set to <code class="docutils literal notranslate"><span class="pre">None</span></code> (default), this value is
automatically determined based on the existence of <code class="docutils literal notranslate"><span class="pre">.sycl</span></code> in
<code class="docutils literal notranslate"><span class="pre">sources</span></code>. Set it to <cite>True`</cite> to force SYCL headers and
libraries to be included.</p></li>
<li><p><strong>is_python_module</strong> – If <code class="docutils literal notranslate"><span class="pre">True</span></code> (default), imports the produced shared
library as a Python module. If <code class="docutils literal notranslate"><span class="pre">False</span></code>, behavior depends on
<code class="docutils literal notranslate"><span class="pre">is_standalone</span></code>.</p></li>
<li><p><strong>is_standalone</strong> – If <code class="docutils literal notranslate"><span class="pre">False</span></code> (default) loads the constructed extension
into the process as a plain dynamic library. If <code class="docutils literal notranslate"><span class="pre">True</span></code>, build a
standalone executable.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p><p>Returns the loaded PyTorch extension as a Python module.</p>
<dl class="simple">
<dt>If <code class="docutils literal notranslate"><span class="pre">is_python_module</span></code> is <code class="docutils literal notranslate"><span class="pre">False</span></code> and <code class="docutils literal notranslate"><span class="pre">is_standalone</span></code> is <code class="docutils literal notranslate"><span class="pre">False</span></code>:</dt><dd><p>Returns nothing. (The shared library is loaded into the process as
a side effect.)</p>
</dd>
<dt>If <code class="docutils literal notranslate"><span class="pre">is_standalone</span></code> is <code class="docutils literal notranslate"><span class="pre">True</span></code>.</dt><dd><p>Return the path to the executable. (On Windows, TORCH_LIB_PATH is
added to the PATH environment variable as a side effect.)</p>
</dd>
</dl>
</p>
</dd>
<dt class="field-odd">Return type</dt>
<dd class="field-odd"><p>If <code class="docutils literal notranslate"><span class="pre">is_python_module</span></code> is <code class="docutils literal notranslate"><span class="pre">True</span></code></p>
</dd>
</dl>
<p class="rubric">Example</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.utils.cpp_extension</span> <span class="kn">import</span> <span class="n">load</span>
<span class="gp">>>> </span><span class="n">module</span> <span class="o">=</span> <span class="n">load</span><span class="p">(</span>
<span class="gp">... </span> <span class="n">name</span><span class="o">=</span><span class="s1">'extension'</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">sources</span><span class="o">=</span><span class="p">[</span><span class="s1">'extension.cpp'</span><span class="p">,</span> <span class="s1">'extension_kernel.cu'</span><span class="p">],</span>
<span class="gp">... </span> <span class="n">extra_cflags</span><span class="o">=</span><span class="p">[</span><span class="s1">'-O2'</span><span class="p">],</span>
<span class="gp">... </span> <span class="n">verbose</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
</pre></div>
</div>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.utils.cpp_extension.load_inline">
<span class="sig-prename descclassname"><span class="pre">torch.utils.cpp_extension.</span></span><span class="sig-name descname"><span class="pre">load_inline</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">name</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">cpp_sources</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">cuda_sources</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">sycl_sources</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">functions</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_cflags</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_cuda_cflags</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_sycl_cflags</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_ldflags</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">extra_include_paths</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">build_directory</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">verbose</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">with_cuda</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">with_sycl</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">is_python_module</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">with_pytorch_error_handling</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">keep_intermediates</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">use_pch</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/cpp_extension.html#load_inline"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="reference external" href="https://github.com/pytorch/pytorch/blob/v2.7.0/torch/utils/cpp_extension.py#L1819"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.cpp_extension.load_inline" title="Permalink to this definition">¶</a></dt>
<dd><p>Load a PyTorch C++ extension just-in-time (JIT) from string sources.</p>
<p>This function behaves exactly like <a class="reference internal" href="#torch.utils.cpp_extension.load" title="torch.utils.cpp_extension.load"><code class="xref py py-func docutils literal notranslate"><span class="pre">load()</span></code></a>, but takes its sources as
strings rather than filenames. These strings are stored to files in the
build directory, after which the behavior of <a class="reference internal" href="#torch.utils.cpp_extension.load_inline" title="torch.utils.cpp_extension.load_inline"><code class="xref py py-func docutils literal notranslate"><span class="pre">load_inline()</span></code></a> is
identical to <a class="reference internal" href="#torch.utils.cpp_extension.load" title="torch.utils.cpp_extension.load"><code class="xref py py-func docutils literal notranslate"><span class="pre">load()</span></code></a>.</p>
<p>See <a class="reference external" href="https://github.com/pytorch/pytorch/blob/master/test/test_cpp_extensions_jit.py">the
tests</a>
for good examples of using this function.</p>
<p>Sources may omit two required parts of a typical non-inline C++ extension:
the necessary header includes, as well as the (pybind11) binding code. More
precisely, strings passed to <code class="docutils literal notranslate"><span class="pre">cpp_sources</span></code> are first concatenated into a
single <code class="docutils literal notranslate"><span class="pre">.cpp</span></code> file. This file is then prepended with <code class="docutils literal notranslate"><span class="pre">#include</span>
<span class="pre"><torch/extension.h></span></code>.</p>
<p>Furthermore, if the <code class="docutils literal notranslate"><span class="pre">functions</span></code> argument is supplied, bindings will be
automatically generated for each function specified. <code class="docutils literal notranslate"><span class="pre">functions</span></code> can
either be a list of function names, or a dictionary mapping from function
names to docstrings. If a list is given, the name of each function is used
as its docstring.</p>
<p>The sources in <code class="docutils literal notranslate"><span class="pre">cuda_sources</span></code> are concatenated into a separate <code class="docutils literal notranslate"><span class="pre">.cu</span></code>
file and prepended with <code class="docutils literal notranslate"><span class="pre">torch/types.h</span></code>, <code class="docutils literal notranslate"><span class="pre">cuda.h</span></code> and
<code class="docutils literal notranslate"><span class="pre">cuda_runtime.h</span></code> includes. The <code class="docutils literal notranslate"><span class="pre">.cpp</span></code> and <code class="docutils literal notranslate"><span class="pre">.cu</span></code> files are compiled
separately, but ultimately linked into a single library. Note that no
bindings are generated for functions in <code class="docutils literal notranslate"><span class="pre">cuda_sources</span></code> per se. To bind
to a CUDA kernel, you must create a C++ function that calls it, and either
declare or define this C++ function in one of the <code class="docutils literal notranslate"><span class="pre">cpp_sources</span></code> (and
include its name in <code class="docutils literal notranslate"><span class="pre">functions</span></code>).</p>
<p>The sources in <code class="docutils literal notranslate"><span class="pre">sycl_sources</span></code> are concatenated into a separate <code class="docutils literal notranslate"><span class="pre">.sycl</span></code>
file and prepended with <code class="docutils literal notranslate"><span class="pre">torch/types.h</span></code>, <code class="docutils literal notranslate"><span class="pre">sycl/sycl.hpp</span></code> includes.
The <code class="docutils literal notranslate"><span class="pre">.cpp</span></code> and <code class="docutils literal notranslate"><span class="pre">.sycl</span></code> files are compiled separately, but ultimately
linked into a single library. Note that no bindings are generated for
functions in <code class="docutils literal notranslate"><span class="pre">sycl_sources</span></code> per se. To bind to a SYCL kernel, you must
create a C++ function that calls it, and either declare or define this
C++ function in one of the <code class="docutils literal notranslate"><span class="pre">cpp_sources</span></code> (and include its name
in <code class="docutils literal notranslate"><span class="pre">functions</span></code>).</p>
<p>See <a class="reference internal" href="#torch.utils.cpp_extension.load" title="torch.utils.cpp_extension.load"><code class="xref py py-func docutils literal notranslate"><span class="pre">load()</span></code></a> for a description of arguments omitted below.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>cpp_sources</strong> – A string, or list of strings, containing C++ source code.</p></li>
<li><p><strong>cuda_sources</strong> – A string, or list of strings, containing CUDA source code.</p></li>
<li><p><strong>sycl_sources</strong> – A string, or list of strings, containing SYCL source code.</p></li>
<li><p><strong>functions</strong> – A list of function names for which to generate function
bindings. If a dictionary is given, it should map function names to
docstrings (which are otherwise just the function names).</p></li>
<li><p><strong>with_cuda</strong> – Determines whether CUDA headers and libraries are added to
the build. If set to <code class="docutils literal notranslate"><span class="pre">None</span></code> (default), this value is
automatically determined based on whether <code class="docutils literal notranslate"><span class="pre">cuda_sources</span></code> is
provided. Set it to <code class="docutils literal notranslate"><span class="pre">True</span></code> to force CUDA headers
and libraries to be included.</p></li>
<li><p><strong>with_sycl</strong> – Determines whether SYCL headers and libraries are added to
the build. If set to <code class="docutils literal notranslate"><span class="pre">None</span></code> (default), this value is
automatically determined based on whether <code class="docutils literal notranslate"><span class="pre">sycl_sources</span></code> is
provided. Set it to <code class="docutils literal notranslate"><span class="pre">True</span></code> to force SYCL headers
and libraries to be included.</p></li>
<li><p><strong>with_pytorch_error_handling</strong> – Determines whether pytorch error and
warning macros are handled by pytorch instead of pybind. To do
this, each function <code class="docutils literal notranslate"><span class="pre">foo</span></code> is called via an intermediary <code class="docutils literal notranslate"><span class="pre">_safe_foo</span></code>
function. This redirection might cause issues in obscure cases
of cpp. This flag should be set to <code class="docutils literal notranslate"><span class="pre">False</span></code> when this redirect
causes issues.</p></li>
</ul>
</dd>
</dl>
<p class="rubric">Example</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">torch.utils.cpp_extension</span> <span class="kn">import</span> <span class="n">load_inline</span>
<span class="gp">>>> </span><span class="n">source</span> <span class="o">=</span> <span class="s2">"""</span>
<span class="go">at::Tensor sin_add(at::Tensor x, at::Tensor y) {</span>
<span class="go"> return x.sin() + y.sin();</span>
<span class="go">}</span>
<span class="go">"""</span>
<span class="gp">>>> </span><span class="n">module</span> <span class="o">=</span> <span class="n">load_inline</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s1">'inline_extension'</span><span class="p">,</span>
<span class="gp">... </span> <span class="n">cpp_sources</span><span class="o">=</span><span class="p">[</span><span class="n">source</span><span class="p">],</span>
<span class="gp">... </span> <span class="n">functions</span><span class="o">=</span><span class="p">[</span><span class="s1">'sin_add'</span><span class="p">])</span>