-
Notifications
You must be signed in to change notification settings - Fork 302
/
Copy pathindex.html
1097 lines (930 loc) · 53 KB
/
index.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>
<html lang="en">
<head>
<!-- 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 -->
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<meta http-equiv="X-UA-Compatible" content="ie=edge">
<link rel="shortcut icon" type="image/x-icon" href="/favicon.ico?">
<title>
Accelerated CPU Inference with PyTorch Inductor using torch.compile | PyTorch
</title>
<meta name="robots" content="index, follow" />
<meta name="description" content="Story at a Glance
" />
<meta property="og:image" content="https://pytorch.org/assets/images/social-share.jpg" />
<meta name="twitter:image" content="https://pytorch.org/assets/images/social-share.jpg" />
<meta property="og:locale" content="en_US" />
<meta property="og:type" content="website" />
<meta property="og:title" content="Accelerated CPU Inference with PyTorch Inductor using torch.compile" />
<meta property="og:description" content="Story at a Glance
" />
<meta property="og:site_name" content="PyTorch" />
<meta name="twitter:card" content="summary_large_image" />
<meta name="twitter:title" content="Accelerated CPU Inference with PyTorch Inductor using torch.compile" />
<meta name="twitter:description" content="Story at a Glance
" />
<link rel="stylesheet" href="/assets/main.css">
<script src="/assets/vendor/jquery.min.js"></script>
<script src="/assets/vendor/popper.min.js"></script>
<script src="/assets/vendor/bootstrap.min.js"></script>
<script src="/assets/vendor/anchor.min.js"></script>
<script type="text/x-mathjax-config">
MathJax.Hub.Config({
tex2jax: {
skipTags: ['script', 'noscript', 'style', 'textarea', 'pre'],
inlineMath: [['$','$']]
}
});
</script>
<script src="https://cdn.mathjax.org/mathjax/latest/MathJax.js?config=TeX-AMS-MML_HTMLorMML" type="text/javascript"></script>
<script type="text/javascript" src="https://cdn.jsdelivr.net/npm/docsearch.js@2/dist/cdn/docsearch.min.js"></script>
<script>
!function(f,b,e,v,n,t,s)
{if(f.fbq)return;n=f.fbq=function(){n.callMethod?
n.callMethod.apply(n,arguments):n.queue.push(arguments)};
if(!f._fbq)f._fbq=n;n.push=n;n.loaded=!0;n.version='2.0';
n.queue=[];t=b.createElement(e);t.async=!0;
t.src=v;s=b.getElementsByTagName(e)[0];
s.parentNode.insertBefore(t,s)}(window,document,'script',
'https://connect.facebook.net/en_US/fbevents.js');
fbq('init', '243028289693773');
fbq('track', 'PageView');
</script>
<noscript>
<img height="1" width="1"
src="https://www.facebook.com/tr?id=243028289693773&ev=PageView
&noscript=1"/>
</noscript>
<!-- Twitter universal website tag code -->
<img height="1" width="1" style="display:none;" alt="" src="https://analytics.twitter.com/i/adsct?p_id=Twitter&p_user_id=0&txn_id=o2gi1&events=%5B%5B%22pageview%22%2Cnull%5D%5D&tw_sale_amount=0&tw_order_quantity=0 (https://urldefense.proofpoint.com/v2/url?u=https-3A__analytics.twitter.com_i_adsct-3Fp-5Fid-3DTwitter-26p-5Fuser-5Fid-3D0-26txn-5Fid-3Do2gi1-26events-3D-255B-255B-2522pageview-2522-252Cnull-255D-255D-26tw-5Fsale-5Famount-3D0-26tw-5Forder-5Fquantity-3D0&d=DwMGaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=GMr8XYCDyeQQZuD3noL91A&m=dAJyokk16UvYy-vMrGn_JwYiGfp_eEgo25B9iGDCG-A&s=o6i4D0V0088WH2RnzIoqiF-vj45PL-2sTrsxQ0SNO3A&e=)" />
<img height="1" width="1" style="display:none;" alt="" src="//t.co/i/adsct?p_id=Twitter&p_user_id=0&txn_id=o2gi1&events=%5B%5B%22pageview%22%2Cnull%5D%5D&tw_sale_amount=0&tw_order_quantity=0 (https://urldefense.proofpoint.com/v2/url?u=https-3A__linkprotect.cudasvc.com_url-3Fa-3Dhttp-253a-252f-252ft.co-252fi-252fadsct-253fp-5Fid-253dTwitter-2526p-5Fuser-5Fid-253d0-2526txn-5Fid-253do2gi1-2526events-253d-25255B-25255B-252522pageview-252522-25252Cnull-25255D-25255D-2526tw-5Fsale-5Famount-253d0-2526tw-5Forder-5Fquantity-253d0-26c-3DE-2C1-2CC33dLwIhtuEcl5FhdztSnUwsioeej5k-2DWy0RYREBAq51kGji32A2Cw94YU9vQBpY5tPN3AukEw3C-5F-2DlbtndnLoR7-5FA-5FLoH0Rr7zLtP1ykptN-26typo-3D1&d=DwMGaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=GMr8XYCDyeQQZuD3noL91A&m=dAJyokk16UvYy-vMrGn_JwYiGfp_eEgo25B9iGDCG-A&s=Abgc3XBkhESv8XBYtLchdDZyISGsK6v_BB6cLMJGyCw&e=)" />
<!-- End Twitter universal website tag code -->
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/docsearch.js@2/dist/cdn/docsearch.min.css" />
<link href="/feed.xml" type="application/atom+xml" rel="alternate" title="Pythorch Blog Posts" />
</head>
<body class="blog">
<!-- 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="main-background blog-background blog-detail-background"></div>
<div class="hello-bar">
<div class="container">
Join us at PyTorch Conference in San Francisco, October 22-23. CFP open now! <a target="_blank" href="https://events.linuxfoundation.org/pytorch-conference/">Learn more</a>.
</div>
</div>
<div class="container-fluid header-holder blog-detail-header">
<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="dropdownMenuButton" 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="/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>
<a class="nav-dropdown-item" href="/new">
<span class="dropdown-title">New to PyTorch Foundation</span>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div id="dropdownMenuButton" 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://landscape.pytorch.org/" target="_blank">
<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="/join-ecosystem">
<span class="dropdown-title">Join the Ecosystem</span>
</a>
<a class="nav-dropdown-item" href="/#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="/resources">
<span class=dropdown-title>Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="/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 class="main-menu-item">
<div id="dropdownMenuButton" 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="/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="/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" target="_blank" href="https://pytorch.org/executorch/stable/index.html">
<span class="dropdown-title">ExecuTorch Documentation</span>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div id="docsDropdownButton" 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">
<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="/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 class="main-menu-item">
<div id="dropdownMenuButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Blog & News
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="/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="/community-blog">
<span class="dropdown-title">Community Blog</span>
<p>Stories from the PyTorch ecosystem</p>
</a>
<a class="nav-dropdown-item" href="/videos">
<span class="dropdown-title">Videos</span>
<p>Learn about the latest PyTorch tutorials, new, and more </p>
</a>
<a class="nav-dropdown-item" href="/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="/events">
<span class=dropdown-title>Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
<a class="nav-dropdown-item" href="/newsletter">
<span class=dropdown-title>Newsletter</span>
<p>Stay up-to-date with the latest updates</p>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<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="/foundation">
<span class=dropdown-title>PyTorch Foundation</span>
<p>Learn more about the PyTorch Foundation.</p>
</a>
<a class="nav-dropdown-item" href="/governing-board">
<span class=dropdown-title>Governing Board</span>
</a>
<a class="nav-dropdown-item" href="/credits">
<span class=dropdown-title>Cloud Credit Program</span>
</a>
<a class="nav-dropdown-item" href="/tac">
<span class=dropdown-title>Technical Advisory Council</span>
</a>
<a class="nav-dropdown-item" href="/staff">
<span class=dropdown-title>Staff</span>
</a>
<a class="nav-dropdown-item" href="/contact-us">
<span class=dropdown-title>Contact Us</span>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<a href="/join" data-cta="join">
Become a Member
</a>
</li>
<li class="main-menu-item" id="github-main-menu-link">
<a href="https://github.com/pytorch/pytorch" title="Go to PyTorch GitHub">
<div id="topnav-gh-icon"></div>
</a>
</li>
<li class="navSearchWrapper reactNavSearchWrapper" key="search">
<div class="search-border">
<div id="search-icon"></div>
<input
id="search-input"
type="text"
title="Search"
/>
<div id="close-search">X</div>
</div>
</li>
</ul>
</div>
<script src="/assets/main-menu-dropdown.js"></script>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<div class="jumbotron jumbotron-fluid blog-detail-jumbotron">
<div class="container blog-detail-container">
<p class="featured-post">September 13, 2023</p>
<h1>
<a class="blog-title">Accelerated CPU Inference with PyTorch Inductor using torch.compile</a>
</h1>
</div>
</div>
<div class="main-content-wrapper blog-detail-wrapper">
<div class="main-content blog-detail-content">
<div class="container">
<img src="/assets/images/logo-icon.svg" class="img-fluid author-icon">
<article class="pytorch-article">
<p class="author">
by
Intel
</p>
<h2 id="story-at-a-glance">Story at a Glance</h2>
<ul>
<li><em>Although the PyTorch* Inductor C++/OpenMP* backend has enabled users to take advantage of modern CPU architectures and parallel processing, it has lacked optimizations, resulting in the backend performing worse than eager mode in terms of end-to-end performance.</em></li>
<li><em>Intel optimized the Inductor backend using a hybrid strategy that classified operations into two categories: Conv/GEMM and non-Conv/GEMM element-wise and reduction ops.</em></li>
<li><em>For popular deep learning models, this hybrid strategy demonstrates promising performance improvements compared to eager mode and improves the C++/OpenMP backend’s efficiency and reliability for PyTorch models.</em></li>
</ul>
<hr />
<h2 id="inductor-backend-challenges">Inductor Backend Challenges</h2>
<p>The PyTorch Inductor C++/OpenMP backend enables users to take advantage of modern CPU architectures and parallel processing to accelerate computations.</p>
<p>However, during the early stages of its development, the backend lacked some optimizations, which prevented it from fully utilizing the CPU computation capabilities. As a result, for most models the C++/OpenMP backend performed worse than eager mode in terms of end-to-end performance, with 45% of TorchBench, 100% of Hugging Face, and 75% of TIMM models performing worse than eager mode.</p>
<p>In this post, we highlight Intel’s optimizations to the Inductor CPU backend, including the technologies and results.</p>
<p>We optimized the backend by using a hybrid strategy that classified operations into two categories: Conv/GEMM and non-Conv/GEMM element-wise and reduction ops. Post-op fusion and weight prepacking using the oneDNN performance library were utilized to optimize the former, while explicit vectorization in C++ codegen was used to optimize the latter.</p>
<p>This hybrid strategy demonstrated promising performance improvements compared to eager mode, particularly on popular deep learning models such as Inductor Hugging Face, Inductor TorchBench and Inductor TIMM. Overall, Intel’s optimizations improve the C++/OpenMP backend’s efficiency and reliability for PyTorch models.</p>
<p><img src="/assets/images/accelerated-cpu-inference/f1-pytorch-inference-speedup-ratio-trend-multi.png.rendition.intel.web.1648.927.png" alt="Figure 1. Performance Speedup Ratio Trend" style="width:100%;" /></p>
<p><small style="line-height: 1.1"><em><strong>Figure 1</strong>: Performance Speedup Ratio Trend</em></small></p>
<h3 id="performance-status-of-intel-hybrid-optimizations">Performance Status of Intel Hybrid Optimizations</h3>
<p>Compared to eager mode with the hybrid optimizations, the C++/OpenMP backend shows promising performance improvements. We measured the performance of the three Inductor benchmark suites—TorchBench, Hugging Face, and TIMM—and the results are as follows. (<em>Note: we publish our performance data twice per week on <a href="http://github.com/pytorch/pytorch/issues/93531">GitHub</a>.</em>)</p>
<p>Overall, these optimizations help to ensure that the C++/OpenMP backend provides efficient and reliable support for PyTorch models.</p>
<h3 id="passrate">Passrate</h3>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>+----------+------------+-------------+-------------+
| Compiler | torchbench | huggingface | timm_models |
+----------+------------+-------------+-------------+
| inductor | 93%, 56/60 | 96%, 44/46 | 100%, 61/61 |
+----------+------------+-------------+-------------+
</code></pre></div></div>
<h3 id="geometric-mean-speedup-single-socket-multi-threads">Geometric mean speedup (Single-Socket Multi-threads)</h3>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>+----------+------------+-------------+-------------+
| Compiler | torchbench | huggingface | timm_models |
+----------+------------+-------------+-------------+
| inductor | 1.39x | 1.20x | 1.73x |
+----------+------------+-------------+-------------+
</code></pre></div></div>
<h3 id="individual-model-performance">Individual Model Performance</h3>
<p><img src="/assets/images/accelerated-cpu-inference/f2-torchbench-fp32-performance-multithread.png.rendition.intel.web.1648.927.png" alt="Figure 2. TorchBench FP32 Performance (Single-Socket Multi-threads)" style="width:100%;" /></p>
<p><small style="line-height: 1.1"><em><strong>Figure 2</strong>: TorchBench FP32 Performance (Single-Socket Multi-threads)</em></small></p>
<p><img src="/assets/images/accelerated-cpu-inference/f3-huggingface-fp32-performance-multithread.png.rendition.intel.web.1648.927.png" alt="Figure 3. Hugging Face FP32 Performance (Single-Socket Multi-thread)" style="width:100%;margin-top: 3em;" /></p>
<p><small style="line-height: 1.1"><em><strong>Figure 3</strong>: Hugging Face FP32 Performance (Single-Socket Multi-thread)</em></small></p>
<p><img src="/assets/images/accelerated-cpu-inference/f4-timm-fp32-performance-multithread.png.rendition.intel.web.1648.927.png" alt="Figure 4. TIMM FP32 Performance (Single-Socket Multi-threads)" style="width:100%;margin-top: 3em;" /></p>
<p><small style="line-height: 1.1"><em><strong>Figure 4</strong>: TIMM FP32 Performance (Single-Socket Multi-threads)</em></small></p>
<h3 id="geometric-mean-speedup-single-core-single-thread">Geometric mean speedup (Single-core Single-thread)</h3>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>+----------+------------+-------------+-------------+
| Compiler | torchbench | huggingface | timm_models |
+----------+------------+-------------+-------------+
| inductor | 1.29x | 1.15x | 1.37x |
+----------+------------+-------------+-------------+
</code></pre></div></div>
<p><img src="/assets/images/accelerated-cpu-inference/f5-torchbench-fp32-performance-single-thread.png.rendition.intel.web.1648.927.png" alt="Figure 5. TorchBench FP32 Performance (Single-Socket Single-thread)" style="width:100%;" /></p>
<p><small style="line-height: 1.1"><em><strong>Figure 5</strong>: TorchBench FP32 Performance (Single-Socket Single-thread)</em></small></p>
<p><img src="/assets/images/accelerated-cpu-inference/f6-huggingface-fp32-performance-single-thread.png.rendition.intel.web.1648.927.png" alt="Figure 6. Hugging Face FP32 Performance (Single-Socket Single Thread)" style="width:100%;margin-top: 3em;" /></p>
<p><small style="line-height: 1.1"><em><strong>Figure 6</strong>: Hugging Face FP32 Performance (Single-Socket Single Thread)</em></small></p>
<p><img src="/assets/images/accelerated-cpu-inference/f7-timm-fp32-performance-single-thread.png.rendition.intel.web.1648.927.png" alt="Figure 7. TIMM FP32 Performance (Single-Socket Single-thread)" style="width:100%;margin-top: 3em;" /></p>
<p><small style="line-height: 1.1"><em><strong>Figure 7</strong>: TIMM FP32 Performance (Single-Socket Single-thread)</em></small></p>
<h2 id="technical-deep-dive">Technical Deep Dive</h2>
<p>Now, let’s take a closer look at the two primary optimizations used in the Inductor C++/OpenMP backend:</p>
<ol>
<li>weight prepacking and post-operation fusion via oneDNN library</li>
<li>explicit vectorization in Inductor C++ codegen</li>
</ol>
<h3 id="weight-prepackaging--post-op-fusion-via-onednn">Weight Prepackaging & Post-op Fusion via oneDNN</h3>
<p>Shorthand for Intel® oneAPI Deep Neural Network Library, oneDNN library provides a range of post-op fusions (i.e., fuse convolution and matmal with its consecutive operation) that can benefit popular models. The <a href="https://github.com/intel/intel-extension-for-pytorch">Intel® Extension for PyTorch</a> has implemented most of these fusions and has achieved significant performance improvements. As a result, we have upstreamed all of these fusions that have been applied in Intel’s PyTorch extension to Inductor, enabling a wider range of models to benefit from these optimizations. We have defined these fusions as operators under the mkldnn namespace. This allows the Python module to invoke these mkldnn operations directly.</p>
<p>Currently, the defined fused operations are as follows. You can find these defined fused operations at <a href="https://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/aten/src/ATen/native/mkldnn/RegisterMkldnnOpContextClass.cpp#L35-#L48">RegisterMkldnnOpContextClass.cpp</a>.</p>
<ul>
<li><code class="language-plaintext highlighter-rouge">_linear_pointwise</code>: Fuses Linear and its post-unary element-wise operations</li>
<li><code class="language-plaintext highlighter-rouge">_linear_pointwise.binary</code>: Fuses Linear and its post-binary element-wise operations</li>
<li><code class="language-plaintext highlighter-rouge">_convolution_pointwise</code>: Fuses Convolution and its post-unary element-wise operations</li>
<li><code class="language-plaintext highlighter-rouge">_convolution_pointwise.binary</code>: Fuses Convolution and its post-binary element-wise operations</li>
</ul>
<p>The detailed fusion patterns are defined in the <a href="https://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/torch/_inductor/mkldnn.py#L774-#L818">mkldnn.py</a> file: <code class="language-plaintext highlighter-rouge">convolution/linear + sigmoid/hardsigmoid/tanh/hardtanh/hardswish/leaky_relu/gelu/relu/relu6/siluconvolution/linear + add/add_/iadd/sub/sub_</code></p>
<p>On the Inductor side, we apply these fusions on the FX graph that has been lowered. We have defined <a href="https://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/torch/_inductor/mkldnn.py#L491">mkldnn_fuse_fx</a> as the entry point to apply all the fusions. The code snippet for this is as follows:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>def mkldnn_fuse_fx(gm: torch.fx.GraphModule, example_inputs):
...
gm = fuse_unary(gm)
gm = fuse_binary(gm)
...
if config.cpp.weight_prepack:
gm = pack_module(gm)
return gm
</code></pre></div></div>
<p>In the <code class="language-plaintext highlighter-rouge">mkldnn_fuse_fx</code> function, we apply fusion on the FX graph that hasn’t been lowered yet. To fuse convolution/linear and its consecutive elementwise operations, we invoke <code class="language-plaintext highlighter-rouge">fuse_unary</code> and <code class="language-plaintext highlighter-rouge">fuse_binary</code> as follows:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code> gm = fuse_unary(gm)
gm = fuse_binary(gm)
</code></pre></div></div>
<p>In addition to the post-op fusion, we apply weight prepacking to improve the Conv/GEMM performance further:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code> gm = pack_module(gm)
</code></pre></div></div>
<p>Weight prepacking involves rearranging the weight tensor in a blocked layout, which:</p>
<ul>
<li>can improve vectorization and cache reuse compared to plain formats like NCHW or NHWC and;</li>
<li>can help avoid weight reordering at runtime, which can reduce overhead and improve performance and;</li>
<li>increases memory usage as the tradeoff.</li>
</ul>
<p>For these reasons, we provide <code class="language-plaintext highlighter-rouge">config.cpp.weight_prepack</code> flag in Inductor to provide users with more control over this optimization, allowing them to enable it based on their specific needs.</p>
<h3 id="explicit-vectorization-in-inductor-c-codegen">Explicit Vectorization in Inductor C++ Codegen</h3>
<p>Vectorization is a key optimization technique that can significantly improve the performance of numerical computations. By utilizing SIMD (Single Instruction, Multiple Data) instructions, vectorization enables multiple computations to be performed simultaneously on a single processor core, which can lead to significant performance improvements.</p>
<p>In the Inductor C++/OpenMP backend, we use <a href="https://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/torch/_inductor/codecache.py#L372">Intel® AVX2</a> and <a href="https://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/torch/_inductor/codecache.py#L359">Intel® AVX-512</a> ISA (Instruction Set Architecture) options for vectorization by leveraging the aten vectorization library to facilitate the implementation. Aten vectorization supports multiple platforms, including x86 and Arm, as well as multiple data types. It can be extended to support other ISAs easily by adding more <a href="https://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/torch/_inductor/codecache.py#L275">VecISA</a> sub-classes. This allows Inductor to easily support other platforms and data types in the future.</p>
<p>Due to differences in platforms, the C++/OpenMP backend of Inductor starts by detecting the CPU features to determine the vectorization bit width at the beginning of code generation. By default, if the machine supports both AVX-512 and AVX2, the backend will choose 512-bit vectorization.</p>
<p>If the hardware supports vectorization, the C++/OpenMP backend first detects if the loop body can be vectorized or not. There are primarily three scenarios that we are not able to generate kernel with vectorization:</p>
<ol>
<li>Loop body lacks vector intrinsics support, e.g., <code class="language-plaintext highlighter-rouge">rand</code> and <code class="language-plaintext highlighter-rouge">atomic_add</code>.</li>
<li>Loop body lacks efficient vector intrinsics support, e.g., non-contiguous <code class="language-plaintext highlighter-rouge">load/store</code>.</li>
<li>Data types with vectorization not yet supported but work in progress, e.g., integer, double, half, and bfloat16.</li>
</ol>
<p>To address this issue, the C++/OpenMP backend uses <a href="https://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/torch/_inductor/codegen/cpp.py#L1396">CppVecKernelChecker</a> to detect whether all operations in a particular loop body can be vectorized or not. In general, we classified the operations into two categories by identifying if they depend on the context.</p>
<p>For most elementwise operations such as <code class="language-plaintext highlighter-rouge">add</code>, <code class="language-plaintext highlighter-rouge">sub</code>, <code class="language-plaintext highlighter-rouge">relu</code>, vectorization is straightforward, and their execution does not depend on context.</p>
<p>However, for certain other operations, their semantics are more complex and their execution depends on context through static analysis.</p>
<p>For example, let’s consider the where operation that takes in mask, <code class="language-plaintext highlighter-rouge">true_value</code>, and <code class="language-plaintext highlighter-rouge">false_value</code> while the mask value is loaded from a <code class="language-plaintext highlighter-rouge">uint8</code> tensor. The fx graph could be as follows:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>graph():
%ops : [#users=9] = placeholder[target=ops]
%get_index : [#users=1] = call_module[target=get_index](args = (index0,), kwargs = {})
%load : [#users=1] = call_method[target=load](args = (%ops, arg1_1, %get_index), kwargs = {})
%to_dtype : [#users=1] = call_method[target=to_dtype](args = (%ops, %load, torch.bool), kwargs = {})
...
%where : [#users=1] = call_method[target=where](args = (%ops, %to_dtype, %to_dtype_2, %to_dtype_3), kwargs = {})
</code></pre></div></div>
<p>Regarding <code class="language-plaintext highlighter-rouge">uint8</code>, it is a general data type and could be used for computation but is not limited to being used as Boolean for mask. Hence, we need to analyze its context statically. In particular, the <a href="https://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/torch/_inductor/codegen/cpp.py#L1396">CppVecKernelChecker</a> will check whether a uint8 tensor is only used by <code class="language-plaintext highlighter-rouge">to_dtype</code> and <code class="language-plaintext highlighter-rouge">to_dtype</code> is only used by where. If yes, it could be vectorized. Otherwise, it will fall back to the scalar version. The generated code could be as follows:</p>
<p>Scalar Version</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>auto tmp0 = in_ptr0[i1 + (17*i0)];
auto tmp3 = in_ptr1[i1 + (17*i0)];
auto tmp1 = static_cast<bool>(tmp0);
auto tmp2 = static_cast<float>(-33.0);
auto tmp4 = tmp1 ? tmp2 : tmp3;
tmp5 = std::max(tmp5, tmp4);
</code></pre></div></div>
<p>Vectorization Version</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>float g_tmp_buffer_in_ptr0[16] = {0};
// Convert the flag to float for vectorization.
flag_to_float(in_ptr0 + (16*i1) + (17*i0), g_tmp_buffer_in_ptr0, 16);
auto tmp0 = at::vec::Vectorized<float>::loadu(g_tmp_buffer_in_ptr0);
auto tmp3 = at::vec::Vectorized<float>::loadu(in_ptr1 + (16*i1) + (17*i0));
auto tmp1 = (tmp0);
auto tmp2 = at::vec::Vectorized<float>(static_cast<float>(-33.0));
auto tmp4 = decltype(tmp2)::blendv(tmp3, tmp2, tmp1);
</code></pre></div></div>
<p>In addition to context analysis, the C++/OpenMP backend also incorporates several other vectorization-related optimizations. These include:</p>
<ul>
<li>Tiled kernel implementation for supporting transpose load - <a href="http://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/torch/_inductor/codegen/cpp.py#L1211">cpp.py</a></li>
<li>Data type demotion based on value range - <a href="http://github.com/pytorch/pytorch/blob/fe05266fda4f908130dea7cbac37e9264c0429a2/torch/_inductor/codegen/cpp.py#L1647-#L1672">cpp.py</a></li>
<li>Replacement of <a href="http://github.com/shibatch/sleef/tree/e0a003ee838b75d11763aa9c3ef17bf71a725bff">sleef</a> implementation with oneDNN/oneMKL implementation for optimizing aten vectorization - <a href="http://github.com/pytorch/pytorch/pull/94577">#94577</a>, <a href="http://github.com/pytorch/pytorch/pull/92289">#92289</a>, <a href="http://github.com/pytorch/pytorch/pull/91613">#91613</a></li>
</ul>
<p>In summary, we examined vectorization optimization in Inductor C++ backend for FP32 training and inference of 150 benchmark models with 90% of inference kernels and 71% of training kernels being vectorized.</p>
<p>In terms of inference, a total of 28,185 CPP kernels were generated, with 25,579 (90%) of them being vectorized, while the remaining 10% were scalar. As for training, 103,084 kernels were generated, with 73,909 (71%) being vectorized and 29% not vectorized.</p>
<p>The results indicate that <strong>the vectorization of inference kernels is quite impressive</strong> (there is still some work to be done in training kernels since we just started to work on the training). The remaining non-vectorized kernels are analyzed in different categories, highlighting the next steps to improve vectorization coverage: index-related operations, int64 support, vertical reduction, vectorization with fallback, and more.</p>
<p>In addition, we also optimized the C++/OpenMP backend with other optimizations like buffer-reuse and CppWrapper.</p>
<h4 id="future-work">Future Work</h4>
<p>The next step, we will continue optimizing the C++/OpenMP backend and extend it to support more data types as the next step. This includes:</p>
<ol>
<li>Improve vectorization coverage</li>
<li>Support and optimize low precision kernel including BF16, FP16, Quantization</li>
<li>Training optimization</li>
<li>Loop tiling</li>
<li>Autotune</li>
<li>Further fusion optimization of Conv/GEMM kernels.</li>
<li>Explore alternative codegen paths: clang/llvm/triton</li>
</ol>
<h2 id="summary">Summary</h2>
<p>Inductor C++/OpenMP backend is a flexible and efficient backend for the CPU. This blog describes the optimizations used in the C++/OpenMP backend of Inductor for inference and training of three benchmark suites – TorchBench, Hugging</p>
<p>Face and TIMM. The primary optimizations include weight prepacking and post-operation fusion via the oneDNN library, as well as explicit vectorization in Inductor C++ codegen using AVX2 and AVX-512 instructions.</p>
<p>The results show that 90% of inference kernels and 71% of training kernels are vectorized, indicating impressive vectorization for inference and room for improvement in training. In addition, we also applied other optimizations like buffer-reuse and CppWrapper. And we will continuously focus on the future work mentioned above to further improve the performance.</p>
<h3 id="acknowledgements">Acknowledgements</h3>
<p>The results presented in this blog post are the culmination of a collaborative effort between the Intel PyTorch team and Meta. We would like to express our sincere gratitude to <a href="http://dev-discuss.pytorch.org/u/jansel">@jansel</a>, <a href="http://dev-discuss.pytorch.org/u/desertfire">@desertfire</a>, and <a href="http://dev-discuss.pytorch.org/u/chillee">@Chillee</a> for their invaluable contributions and unwavering support throughout the development process. Their expertise and dedication have been instrumental in achieving the optimizations and performance improvements discussed here.</p>
<h3 id="configuration-details">Configuration Details</h3>
<h4 id="hardware-details">Hardware Details</h4>
<table>
<tr>
<td>
<strong>Item </strong>
</td>
<td>
<strong>Value </strong>
</td>
</tr>
<tr>
<td>
Manufacturer
</td>
<td>
Amazon EC2
</td>
</tr>
<tr>
<td>
Product Name
</td>
<td>
c6i.16xlarge
</td>
</tr>
<tr>
<td>
CPU Model
</td>
<td>
Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz
</td>
</tr>
<tr>
<td>
Installed Memory
</td>
<td>
128GB (1x128GB DDR4 3200 MT/s [Unknown])
</td>
</tr>
<tr>
<td>
OS
</td>
<td>
Ubuntu 22.04.2 LTS
</td>
</tr>
<tr>
<td>
Kernel
</td>
<td>
5.19.0-1022-aws
</td>
</tr>
<tr>
<td>
Microcode
</td>
<td>
0xd000389
</td>
</tr>
<tr>
<td>
GCC
</td>
<td>
gcc (Ubuntu 11.3.0-1ubuntu1~22.04) 11.3.0
</td>
</tr>
<tr>
<td>
GLIBC
</td>
<td>
ldd (Ubuntu GLIBC 2.35-0ubuntu3.1) 2.35
</td>
</tr>
<tr>
<td>
Binutils
</td>
<td>
GNU ld (GNU Binutils for Ubuntu) 2.38
</td>
</tr>
<tr>
<td>
Python
</td>
<td>
Python 3.10.6
</td>
</tr>
<tr>
<td>
OpenSSL
</td>
<td>
OpenSSL 3.0.2 15 Mar 2022 (Library: OpenSSL 3.0.2 15 Mar 2022)
</td>
</tr>
</table>
<h4 id="software-details">Software Details</h4>
<table>
<tr>
<td>
<strong>SW</strong>
</td>
<td>
<strong>Nightly commit</strong>
</td>
<td>
<strong>Main commit</strong>
</td>
</tr>
<tr>
<td>
Pytorch
</td>
<td>
a977a12
</td>
<td>
0b1b063
</td>
</tr>
<tr>
<td>
Torchbench
</td>
<td>
/
</td>
<td>
a0848e19
</td>
</tr>
<tr>
<td>
torchaudio
</td>
<td>
0a652f5
</td>
<td>
d5b2996
</td>
</tr>
<tr>
<td>
torchtext
</td>
<td>
c4ad5dd
</td>
<td>
79100a6
</td>
</tr>
<tr>
<td>
torchvision
</td>
<td>
f2009ab
</td>
<td>
b78d98b
</td>
</tr>
<tr>
<td>
torchdata
</td>
<td>
5cb3e6d
</td>
<td>
f2bfd3d
</td>
</tr>
<tr>
<td>
dynamo_benchmarks
</td>
<td>
fea73cb
</td>
<td>
/
</td>
</tr>
</table>
<h4 id="configuration">Configuration</h4>
<ul>
<li>Intel OpenMP</li>
<li>Jemalloc - oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:-1,muzzy_decay_ms:-1</li>
<li><strong>Single-Socket Multi-threads:</strong> #of Instances: 1; Cores/Instance: 32</li>
<li><strong>Single-Core Single-thread:</strong> #of Instances: 1; Cores/Instance: 1</li>
</ul>
</article>
</div>
</div>
</div>
<!--
-->
<div class="container-fluid 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="/docs">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="/resources">View Resources</a>
</div>
</div>
</div>
</div>
<footer class="site-footer">
<div class="container footer-container">
<div class="newsletter" id="newsletter">
<p
class="newsletter__title is-style-max-width-800"><strong>Stay in touch</strong> for updates, event info, and the latest news</p>
<script charset="utf-8" type="text/javascript" src="//js.hsforms.net/forms/embed/v2.js"></script>
<script>
hbspt.forms.create({
region: "na1",
portalId: "8112310",
formId: "2fb2231c-000b-4ec5-88a0-1ab242549c9e"
});
</script>
<p
class="newsletter__privacy">By submitting this form, I consent to receive marketing emails from the LF and its projects regarding their events, training, research, developments, and related announcements. I understand that I can unsubscribe at any time using the links in the footers of the emails I receive. <a href="https://www.linuxfoundation.org/privacy/">Privacy Policy</a>.</p>
</div>
<div class="lf-grid">
<div class="footer-logo-wrapper">
<a href="https://pytorch.org" class="footer-logo">
<img src="/assets/images/logo-icon.svg" alt="PyTorch logo" width="40">
</a>
</div>
<ul class="social-links">
<li><a href="https://www.facebook.com/pytorch" target="_blank" title="PyTorch on Facebook">
<svg xmlns="http://www.w3.org/2000/svg" viewbox="-0.51 -0.26 26.45 26.45" aria-label="Facebook"><path fill="currentColor" d="M25.497 13.075c0-2.45-.698-4.848-2.011-6.911a12.765 12.765 0 0 0-5.398-4.73A12.671 12.671 0 0 0 11.008.38a12.705 12.705 0 0 0-6.529 2.95A12.827 12.827 0 0 0 .563 9.358a12.896 12.896 0 0 0-.07 7.201 12.831 12.831 0 0 0 3.801 6.103 12.709 12.709 0 0 0 6.471 3.078v-8.957H7.53v-3.708h3.235v-2.824c0-3.213 1.903-4.988 4.813-4.988.956.014 1.909.097 2.852.25V8.67h-1.607a1.83 1.83 0 0 0-1.518.497 1.854 1.854 0 0 0-.561 1.505v2.404h3.535l-.563 3.708h-2.97v8.957a12.725 12.725 0 0 0 7.697-4.337 12.87 12.87 0 0 0 3.054-8.328z"/></svg>
</a></li>
<li><a href="https://twitter.com/pytorch" target="_blank" title="PyTorch on X">
<svg xmlns="http://www.w3.org/2000/svg" viewbox="0 0 300 300" aria-label="X"><path fill="currentColor" d="M178.57 127.15 290.27 0h-26.46l-97.03 110.38L89.34 0H0l117.13 166.93L0 300.25h26.46l102.4-116.59 81.8 116.59h89.34M36.01 19.54H76.66l187.13 262.13h-40.66"/></svg>
</a></li>
<li><a href="https://www.youtube.com/pytorch" target="_blank" title="PyTorch on YouTube">
<svg xmlns="http://www.w3.org/2000/svg" viewbox="0.21 0.27 34.45 25.07" aria-label="YouTube"><path fill="currentColor" d="M33.729 6.084s-.327-2.33-1.317-3.356a4.691 4.691 0 0 0-3.32-1.432c-4.634-.34-11.589-.34-11.589-.34h-.014s-6.954 0-11.59.342a4.692 4.692 0 0 0-3.32 1.432c-.993 1.025-1.315 3.354-1.315 3.354a52.189 52.189 0 0 0-.331 5.473v2.566c.014 1.829.125 3.656.331 5.472 0 0 .322 2.33 1.316 3.36 1.26 1.345 2.916 1.3 3.653 1.445 2.65.26 11.263.34 11.263.34s6.96-.01 11.597-.353a4.691 4.691 0 0 0 3.32-1.432c.993-1.026 1.316-3.356 1.316-3.356.206-1.817.316-3.644.33-5.473v-2.57a52.26 52.26 0 0 0-.33-5.472zM14.076 17.232V7.729l8.951 4.768-8.95 4.735z"/></svg>
</a></li>
<li><a href="https://www.linkedin.com/company/pytorch" target="_blank" title="PyTorch on LinkedIn">
<svg xmlns="http://www.w3.org/2000/svg" viewbox="-10.23 -10.23 531.96 531.96" aria-label="LinkedIn"><rect width="512" height="512" rx="0" fill="currentColor"/><circle fill="#000" cx="142" cy="138" r="37"/><path stroke="#000" stroke-width="66" d="M244 194v198M142 194v198"/><path fill="#000" d="M276 282c0-20 13-40 36-40 24 0 33 18 33 45v105h66V279c0-61-32-89-76-89-34 0-51 19-59 32"/></svg>
</a></li>
<li><a href="https://join.slack.com/t/pytorch/shared_invite/zt-2j2la612p-miUinTTaxXczKOJw48poHA" target="_blank" title="PyTorch Slack">
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0.16 -0.03 21.19 21.19" aria-label="Slack"><path fill="currentColor" d="M4.896 13.27a2.147 2.147 0 0 1-2.141 2.142A2.147 2.147 0 0 1 .613 13.27c0-1.178.963-2.141 2.142-2.141h2.141v2.141zm1.08 0c0-1.178.962-2.141 2.141-2.141s2.142.963 2.142 2.141v5.363a2.147 2.147 0 0 1-2.142 2.141 2.147 2.147 0 0 1-2.141-2.142V13.27zm2.141-8.6a2.147 2.147 0 0 1-2.141-2.14c0-1.18.962-2.142 2.141-2.142s2.142.963 2.142 2.141v2.142H8.117zm0 1.08c1.179 0 2.141.962 2.141 2.141a2.147 2.147 0 0 1-2.141 2.142H2.755A2.147 2.147 0 0 1 .613 7.89c0-1.179.963-2.141 2.142-2.141h5.362zm8.599 2.141c0-1.179.963-2.141 2.141-2.141 1.179 0 2.143.962 2.143 2.14a2.147 2.147 0 0 1-2.142 2.142h-2.141V7.89zm-1.08 0a2.147 2.147 0 0 1-2.141 2.142 2.147 2.147 0 0 1-2.141-2.142V2.53c0-1.178.962-2.141 2.141-2.141s2.142.963 2.142 2.141v5.362zm-2.141 8.6c1.179 0 2.142.962 2.142 2.14a2.147 2.147 0 0 1-2.142 2.142 2.147 2.147 0 0 1-2.141-2.141V16.49h2.141zm0-1.08a2.147 2.147 0 0 1-2.141-2.141c0-1.179.962-2.142 2.141-2.142h5.362c1.179 0 2.142.963 2.142 2.142a2.147 2.147 0 0 1-2.142 2.142h-5.362z"></path></svg>
</a></li>
<li><a href="/wechat" title="PyTorch on WeChat">
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0.14 -0.17 38.02 33.02" aria-label="WeChat"><path fill="currentColor" d="M26.289 10.976a12.972 12.972 0 0 0-8.742 3.53 10.386 10.386 0 0 0-3.224 8.795c-1.326-.164-2.535-.345-3.75-.448a2.332 2.332 0 0 0-1.273.216c-1.18.666-2.311 1.418-3.652 2.255.246-1.112.405-2.087.687-3.024a1.15 1.15 0 0 0-.523-1.52C1.737 17.902.02 13.601 1.307 9.165c1.189-4.1 4.11-6.587 8.077-7.884A13.54 13.54 0 0 1 24.18 5.617a10.135 10.135 0 0 1 2.109 5.359zM10.668 9.594a1.564 1.564 0 0 0-2.095-1.472 1.52 1.52 0 0 0-.895 1.964 1.502 1.502 0 0 0 1.391.966 1.545 1.545 0 0 0 1.598-1.46v.002zm8.15-1.566a1.567 1.567 0 0 0-1.528 1.543 1.528 1.528 0 0 0 1.571 1.492 1.52 1.52 0 0 0 1.375-2.117 1.518 1.518 0 0 0-1.415-.919l-.003.001z"></path><path fill="currentColor" d="M33.914 32.137c-1.075-.478-2.062-1.196-3.11-1.306-1.049-.11-2.145.494-3.24.605a10.821 10.821 0 0 1-8.781-2.864c-4.682-4.33-4.013-10.97 1.403-14.518 4.811-3.154 11.874-2.102 15.268 2.273a8.671 8.671 0 0 1-1.002 12.095c-1.046.929-1.422 1.693-.751 2.917.102.257.174.525.213.798zM21.68 20.292a1.264 1.264 0 1 0 .01-2.528 1.264 1.264 0 0 0-.01 2.528zm7.887-2.526a1.266 1.266 0 0 0-1.256 1.21 1.247 1.247 0 1 0 1.256-1.21z"></path></svg>
</a></li>
</ul>
</div>
<div class="privacy-policy">
<div class="copyright">
<p>© Copyright The Linux Foundation. The PyTorch Foundation is a project of The Linux Foundation.
For web site terms of use, trademark policy and other policies applicable to The PyTorch Foundation please see
<a href="https://www.linuxfoundation.org/legal/policies/">Linux Foundation Policies</a>. The PyTorch Foundation supports the PyTorch open source
project, which has been established as PyTorch Project a Series of LF Projects, LLC. For policies applicable to the PyTorch Project a Series of LF Projects, LLC,
please see <a href="https://www.lfprojects.org/policies/">LF Projects, LLC Policies</a>. <a href="https://www.linuxfoundation.org/privacy">Privacy Policy</a> and <a href="https://www.linuxfoundation.org/terms">Terms of Use</a>.</p>
</div>
</div>
</div>
</footer>
<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 class="navSearchWrapper reactNavSearchWrapper tabletSearchWrapper" key="search">
<div class="mobile-search-border">
<input
id="mobile-search-input"
type="text"
title="Search"
/>
<div id="mobile-search-icon"></div>
</div>
</li>
<li class="resources-mobile-menu-title">
<a>Learn</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li>
<a href="https://pytorch.org/tutorials/beginner/basics/intro.html">Learn the Basics</a>
</li>
<li>
<a href="https://pytorch.org/tutorials/recipes/recipes_index.html">PyTorch Recipes</a>
</li>
<li>
<a href="https://pytorch.org/tutorials/beginner/introyt.html">Introduction to PyTorch - YouTube Series</a>
</li>
<li>
<a href="/new">New to PyTorch Foundation</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a>Ecosystem</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://landscape.pytorch.org/">Tools</a>
</li>
<li>
<a href="/join-ecosystem">Join the Ecosystem</a>
</li>
<li>
<a href="/#community-module">Community</a>
</li>
<li>
<a href="https://discuss.pytorch.org">Forums</a>
</li>
<li>
<a href="/resources">Developer Resources</a>
</li>
<li>
<a href="/ecosystem/contributor-awards-2024">Contributor Awards - 2024</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a>Edge</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="/edge">About PyTorch Edge</a>
</li>
<li>
<a href="/executorch-overview">ExecuTorch</a>
</li>
<li>
<a href="https://pytorch.org/executorch/stable/index.html">ExecuTorch Documentation</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a>Docs</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/docs">PyTorch</a>
</li>
<li>
<a href="/pytorch-domains">PyTorch Domains</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a>Blog & News</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="/blog">PyTorch Blog</a>
</li>
<li>
<a href="/community-blog">Community Blog</a>
</li>
<li>
<a href="/videos">Videos</a>
</li>
<li>
<a href="/community-stories">Community Stories</a>
</li>
<li>