-
Notifications
You must be signed in to change notification settings - Fork 302
/
Copy pathindex.html
1047 lines (854 loc) · 63.7 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 PyTorch inference with torch.compile on AWS Graviton processors | PyTorch
</title>
<meta name="robots" content="index, follow" />
<meta name="description" content="Summary
" />
<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 PyTorch inference with torch.compile on AWS Graviton processors" />
<meta property="og:description" content="Summary
" />
<meta property="og:site_name" content="PyTorch" />
<meta name="twitter:card" content="summary_large_image" />
<meta name="twitter:title" content="Accelerated PyTorch inference with torch.compile on AWS Graviton processors" />
<meta name="twitter:description" content="Summary
" />
<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">July 09, 2024</p>
<h1>
<a class="blog-title">Accelerated PyTorch inference with torch.compile on AWS Graviton processors</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
Sunita Nadampalli
</p>
<h2 id="summary">Summary</h2>
<p>Originally PyTorch, used an eager mode where each PyTorch operation that forms the model is run independently as soon as it’s reached. PyTorch 2.0 introduced <a href="https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html">torch.compile</a> to speed up PyTorch code over the default eager mode. In contrast to eager mode, the torch.compile pre-compiles the entire model into a single graph in a manner that’s optimal for running on a given hardware platform. AWS optimized the PyTorch torch.compile feature for <a href="https://aws.amazon.com/about-aws/whats-new/2022/05/amazon-ec2-c7g-instances-powered-aws-graviton3-processors/">AWS Graviton3 processors</a>. This optimization results in up to 2x better performance for <a href="https://huggingface.co/models">Hugging Face</a> model inference (based on geomean of performance improvement for 33 models) and up to 1.35x better performance for <a href="https://github.com/pytorch/benchmark">TorchBench</a> model inference (geomean of performance improvement for 45 models) compared to the default eager mode inference across several natural language processing (NLP), computer vision (CV), and recommendation models on AWS Graviton3-based Amazon EC2 instances. Starting with PyTorch 2.3.1, the optimizations are available in torch Python <a href="https://pypi.org/project/torch/2.3.1/">wheels</a> and AWS Graviton PyTorch <a href="https://github.com/aws/deep-learning-containers/blob/master/available_images.md#ec2-framework-graviton-containers-ec2-ecs-and-eks-support-only">deep learning container (DLC)</a>.</p>
<p>In this blog post, we show how we optimized torch.compile performance on AWS Graviton3-based EC2 instances, how to use the optimizations to improve inference performance, and the resulting speedups.</p>
<h2 id="why-torchcompile-and-whats-the-goal">Why torch.compile and what’s the goal?</h2>
<p>In eager mode, operators in a model are run immediately as they are encountered. It’s easier to use, more suitable for machine learning (ML) researchers, and hence is the default mode. However, eager mode incurs runtime overhead because of redundant kernel launch and memory read overhead. Whereas in torch compile mode, operators are first synthesized into a graph, wherein one operator is merged with another to reduce and localize memory reads and total kernel launch overhead.</p>
<p>The goal for the AWS Graviton team was to optimize torch.compile backend for Graviton3 processors. PyTorch eager mode was already optimized for Graviton3 processors with <a href="https://github.com/ARM-software/ComputeLibrary">Arm Compute Library (ACL)</a> kernels using oneDNN (also known as MKLDNN). So, the question was, how to reuse those kernels in torch.compile mode to get the best of graph compilation and the optimized kernel performance together?</p>
<h2 id="results">Results</h2>
<p>The AWS Graviton team extended the torch inductor and oneDNN primitives that reused the ACL kernels and optimized compile mode performance on Graviton3 processors. Starting with PyTorch 2.3.1, the optimizations are available in the torch Python wheels and AWS Graviton DLC. Please see the <strong>Running an inference</strong> section that follows for the instructions on installation, runtime configuration, and how to run the tests.</p>
<p>To demonstrate the performance improvements, we used NLP, CV, and recommendation models from <a href="https://github.com/pytorch/benchmark">TorchBench</a> and the most downloaded NLP models from <a href="https://huggingface.co/models">Hugging Face</a> across Question Answering, Text Classification, Token Classification, Translation, Zero-Shot Classification, Translation, Summarization, Feature Extraction, Text Generation, Text2Text Generation, Fill-Mask, and Sentence Similarity tasks to cover a wide variety of customer use cases.</p>
<p>We started with measuring TorchBench model inference latency, in milliseconds (msec), for the eager mode, which is marked 1.0 with a red dotted line in the following graph. Then we compared the improvements from torch.compile for the same model inference, the normalized results are plotted in the graph. You can see that for the 45 models we benchmarked, there is a 1.35x latency improvement (geomean for the 45 models).</p>
<p><img src="/assets/images/accelerated-pytorch-inference/fg1.png" alt="PyTorch model inference performance improvement with torch.compile on AWS Graviton3-based c7g instance using TorchBench framework" style="width:100%" /></p>
<p><em><strong>Image 1</strong>: PyTorch model inference performance improvement with torch.compile on AWS Graviton3-based c7g instance using TorchBench framework. The reference eager mode performance is marked as 1.0. (higher is better)</em></p>
<p>Similar to the preceding TorchBench inference performance graph, we started with measuring the Hugging Face NLP model inference latency, in msec, for the eager mode, which is marked 1.0 with a red dotted line in the following graph. Then we compared the improvements from torch.compile for the same model inference, the normalized results are plotted in the graph. You can see that for the 33 models we benchmarked, there is around 2x performance improvement (geomean for the 33 models).</p>
<p><img src="/assets/images/accelerated-pytorch-inference/fg2.png" alt="Hugging Face NLP model inference performance improvement with torch.compile on AWS Graviton3-based c7g instance using Hugging Face example scripts" style="width:100%" /></p>
<p><em><strong>Image 2</strong>: Hugging Face NLP model inference performance improvement with torch.compile on AWS Graviton3-based c7g instance using Hugging Face example scripts. The reference eager mode performance is marked as 1.0. (higher is better)</em></p>
<h2 id="running-an-inference">Running an inference</h2>
<p>Starting with PyTorch 2.3.1, the optimizations are available in the torch Python wheel and in AWS Graviton PyTorch DLC. This section shows how to run inference in eager and torch.compile modes using torch Python wheels and benchmarking scripts from Hugging Face and TorchBench repos.</p>
<p>To successfully run the scripts and reproduce the speedup numbers mentioned in this post, you need an instance from the Graviton3 family (<code class="language-plaintext highlighter-rouge">c7g/r7g/m7g/hpc7g</code>) of hardware. For this post, we used the <a href="https://aws.amazon.com/ec2/instance-types/c7g/">c7g.4xl (16 vcpu) instance</a>. The instance, the AMI details, and the required torch library versions are mentioned in the following snippet.</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>Instance: c7g.4xl instance
Region: us-west-2
AMI: ami-05cc25bfa725a144a (Ubuntu 22.04/Jammy with 6.5.0-1017-aws kernel)
# Install Python
sudo apt-get update
sudo apt-get install -y python3 python3-pip
# Upgrade pip3 to the latest version
python3 -m pip install --upgrade pip
# Install PyTorch and extensions
python3 -m pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1
</code></pre></div></div>
<p>The generic runtime tunings implemented for eager mode inference are equally applicable for the torch.compile mode, so, we set the following environment variables to further improve the torch.compile performance on AWS Graviton3 processors.</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code># Enable the fast math GEMM kernels, to accelerate fp32 inference with bfloat16 gemm
export DNNL_DEFAULT_FPMATH_MODE=BF16
# Enable Linux Transparent Huge Page (THP) allocations,
# to reduce the tensor memory allocation latency
export THP_MEM_ALLOC_ENABLE=1
# Set LRU Cache capacity to cache the primitives and avoid redundant
# memory allocations
export LRU_CACHE_CAPACITY=1024
</code></pre></div></div>
<h4 id="torchbench-benchmarking-scripts">TORCHBENCH BENCHMARKING SCRIPTS</h4>
<p>TorchBench is a collection of open source benchmarks used to evaluate PyTorch performance. We benchmarked 45 models using the scripts from the TorchBench repo. Following code shows how to run the scripts for the eager mode and the compile mode with inductor backend.</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code># Set OMP_NUM_THREADS to number of vcpus, 16 for c7g.4xl instance
export OMP_NUM_THREADS=16
# Install the dependencies
sudo apt-get install -y libgl1-mesa-glx
sudo apt-get install -y libpangocairo-1.0-0
python3 -m pip install psutil numpy transformers pynvml numba onnx onnxruntime scikit-learn timm effdet gym doctr opencv-python h5py==3.10.0 python-doctr
# Clone pytorch benchmark repo
git clone https://github.com/pytorch/benchmark.git
cd benchmark
# PyTorch benchmark repo doesn't have any release tags. So,
# listing the commit we used for collecting the performance numbers
git checkout 9a5e4137299741e1b6fb7aa7f5a6a853e5dd2295
# Setup the models
python3 install.py
# Colect eager mode performance using the following command. The results will be
# stored at .userbenchmark/cpu/metric-<timestamp>.json.
python3 run_benchmark.py cpu --model BERT_pytorch,hf_Bert,hf_Bert_large,hf_GPT2,hf_Albert,hf_Bart,hf_BigBird,hf_DistilBert,hf_GPT2_large,dlrm,hf_T5,mnasnet1_0,mobilenet_v2,mobilenet_v3_large,squeezenet1_1,timm_efficientnet,shufflenet_v2_x1_0,timm_regnet,resnet50,soft_actor_critic,phlippe_densenet,resnet152,resnet18,resnext50_32x4d,densenet121,phlippe_resnet,doctr_det_predictor,timm_vovnet,alexnet,doctr_reco_predictor,vgg16,dcgan,yolov3,pytorch_stargan,hf_Longformer,timm_nfnet,timm_vision_transformer,timm_vision_transformer_large,nvidia_deeprecommender,demucs,tts_angular,hf_Reformer,pytorch_CycleGAN_and_pix2pix,functorch_dp_cifar10,pytorch_unet --test eval --metrics="latencies,cpu_peak_mem"
# Collect torch.compile mode performance with inductor backend
# and weights pre-packing enabled. The results will be stored at
# .userbenchmark/cpu/metric-<timestamp>.json
python3 run_benchmark.py cpu --model BERT_pytorch,hf_Bert,hf_Bert_large,hf_GPT2,hf_Albert,hf_Bart,hf_BigBird,hf_DistilBert,hf_GPT2_large,dlrm,hf_T5,mnasnet1_0,mobilenet_v2,mobilenet_v3_large,squeezenet1_1,timm_efficientnet,shufflenet_v2_x1_0,timm_regnet,resnet50,soft_actor_critic,phlippe_densenet,resnet152,resnet18,resnext50_32x4d,densenet121,phlippe_resnet,doctr_det_predictor,timm_vovnet,alexnet,doctr_reco_predictor,vgg16,dcgan,yolov3,pytorch_stargan,hf_Longformer,timm_nfnet,timm_vision_transformer,timm_vision_transformer_large,nvidia_deeprecommender,demucs,tts_angular,hf_Reformer,pytorch_CycleGAN_and_pix2pix,functorch_dp_cifar10,pytorch_unet --test eval --torchdynamo inductor --freeze_prepack_weights --metrics="latencies,cpu_peak_mem"
</code></pre></div></div>
<p>On successful completion of the inference runs, the script stores the results in JSON format. The following is the sample output:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>{
"name": "cpu"
"environ": {
"pytorch_git_version": "d44533f9d073df13895333e70b66f81c513c1889"
},
"metrics": {
"BERT_pytorch-eval_latency": 56.3769865,
"BERT_pytorch-eval_cmem": 0.4169921875
}
}
</code></pre></div></div>
<h4 id="hugging-face-benchmarking-scripts">HUGGING FACE BENCHMARKING SCRIPTS</h4>
<p>Google T5 Small Text Translation model is one of the around 30 Hugging Face models we benchmarked. We’re using it as a sample model to demonstrate how to run inference in eager and compile modes. The additional configurations and APIs required to run it in compile mode are highlighted in <strong>BOLD</strong>. Save the following script as <code class="language-plaintext highlighter-rouge">google_t5_small_text_translation.py</code>.</p>
<pre><code>import argparse
from transformers import T5Tokenizer, T5Model
import torch
from torch.profiler import profile, record_function, ProfilerActivity
<b>import torch._inductor.config as config
config.cpp.weight_prepack=True
config.freezing=True</b>
def test_inference(mode, num_iter):
tokenizer = T5Tokenizer.from_pretrained("t5-small")
model = T5Model.from_pretrained("t5-small")
input_ids = tokenizer(
"Studies have been shown that owning a dog is good for you", return_tensors="pt"
).input_ids # Batch size 1
decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids # Batch size 1
<b>if (mode == 'compile'):
model = torch.compile(model)</b>
with torch.no_grad():
for _ in range(50):
outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
with profile(activities=[ProfilerActivity.CPU]) as prof:
with record_function("model_inference"):
for _ in range(num_iter):
outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
print(prof.key_averages().table(sort_by="self_cpu_time_total"))
def main() -> None:
global m, args
parser = argparse.ArgumentParser(__doc__)
parser.add_argument(
"-m",
"--mode",
choices=["eager", "compile"],
default="eager",
help="Which test to run.",
)
parser.add_argument(
"-n",
"--number",
type=int,
default=100,
help="how many iterations to run.",
)
args = parser.parse_args()
test_inference(args.mode, args.number)
if __name__ == "__main__":
main()
</code></pre>
<p>Run the script with the following steps:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code># Set OMP_NUM_THREADS to number of vcpus to 4 because
# the scripts are running inference in sequence, and
# they don't need large number of vcpus
export OMP_NUM_THREADS=4
# Install the dependencies
python3 -m pip install transformers
# Run the inference script in Eager mode
# using number of iterations as 1 just to show the torch profiler output
# but for the benchmarking, we used 1000 iterations.
python3 google_t5_small_text_translation.py -n 1 -m eager
# Run the inference script in torch compile mode
python3 google_t5_small_text_translation.py -n 1 -m compile
</code></pre></div></div>
<p>On successful completion of the inference runs, the script prints the torch profiler output with the latency breakdown for the torch operators. The following is the sample output from torch profiler:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code># Torch profiler output for the eager mode run on c7g.xl (4vcpu)
------------------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
------------------------ ------------ ------------ ------------ ------------ ------------ ------------
aten::mm 40.71% 12.502ms 40.71% 12.502ms 130.229us 96
model_inference 26.44% 8.118ms 100.00% 30.708ms 30.708ms 1
aten::bmm 6.85% 2.102ms 9.47% 2.908ms 80.778us 36
aten::matmul 3.73% 1.146ms 57.26% 17.583ms 133.205us 132
aten::select 1.88% 576.000us 1.90% 583.000us 0.998us 584
aten::transpose 1.51% 464.000us 1.83% 563.000us 3.027us 186
------------------------ ------------ ------------ ------------ ------------ ------------ -------------------
Self CPU time total: 30.708ms
# Torch profiler output for the compile mode run for the same model on the same instance
--------------------------------- ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
--------------------------------- ------------ ------------ ------------ ------------ ------------ ------------
mkldnn::_linear_pointwise 37.98% 5.461ms 45.91% 6.602ms 68.771us 96
Torch-Compiled Region 29.56% 4.251ms 98.53% 14.168ms 14.168ms 1
aten::bmm 14.90% 2.143ms 21.73% 3.124ms 86.778us 36
aten::select 4.51% 648.000us 4.62% 665.000us 1.155us 576
aten::view 3.29% 473.000us 3.29% 473.000us 1.642us 288
aten::empty 2.53% 364.000us 2.53% 364.000us 3.165us 115
--------------------------------- ------------ ------------ ------------ ------------ ------------ --------------------
Self CPU time total: 14.379ms
</code></pre></div></div>
<h2 id="technical-deep-dive-what-are-the-challenges-and-optimization-details">Technical deep dive: What are the challenges and optimization details</h2>
<p>Underpinning torch.compile are new technologies – TorchDynamo, AOTDispatcher, and TorchInductor.</p>
<p><strong>TorchDynamo</strong> captures PyTorch programs safely using Python Frame Evaluation Hooks<br />
<strong>AOTDispatcher</strong> overloads PyTorch’s autograd engine as a tracing autodiff for generating ahead-of-time backward traces.<br />
<strong>TorchInductor</strong> is a deep learning compiler that generates fast code for multiple accelerators and backends.</p>
<p><img src="/assets/images/accelerated-pytorch-inference/fg3.png" alt="The PyTorch compilation process source" style="width:100%" /></p>
<p><em><strong>Image 3</strong>: The PyTorch compilation process</em></p>
<p>When torch.compile is invoked, torch dynamo rewrites Python bytecode to extract sequences of PyTorch operations into an <a href="https://pytorch.org/docs/stable/fx.html">FX</a> <a href="https://pytorch.org/docs/stable/fx.html">Graph</a>, which is then compiled with inductor backend. For a typical inference scenario where the graph is frozen and gradient calculations are disabled, the inductor invokes platform specific optimizations like graph rewrite into more performant operators, operator fusion, and weights pre-packing.</p>
<p>However, on Graviton3, the inductor wasn’t able to perform any of those optimizations because there was no aarch64 backend defined. To fix this, we extended the inductor’s FX passes to pick oneDNN operators for linear layer compilation on Graviton3 processors with ACL backend. The code snippet for this follows:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>packed_weight_op = (
mkldnn._reorder_linear_weight
if (is_bf16_weight or mkldnn._is_mkldnn_acl_supported())
packed_linear_inputs: Tuple[Any, ...] = (input, packed_weight_node)
if is_bf16_weight or mkldnn._is_mkldnn_acl_supported():
packed_linear_inputs += (bias, "none", [], "")
packed_linear_op = mkldnn._linear_pointwise.default
</code></pre></div></div>
<p>After this was done, the FX pass was successful in compiling the <code class="language-plaintext highlighter-rouge">matmul </code>operators to <code class="language-plaintext highlighter-rouge">linear_pointwise </code>. The following snippet highlights the matmul operator in the original model:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code> %attention_scores : [num_users=1] = call_function[target=torch.matmul](args = (%query_layer, %transpose), kwargs = {})
%attention_scores_1 : [num_users=1] = call_function[target=operator.truediv](args = (%attention_scores, 8.0), kwargs = {})
%attention_scores_2 : [num_users=1] = call_function[target=operator.add](args = (%attention_scores_1, %extended_attention_mask_3), kwargs = {})
</code></pre></div></div>
<p>The following snippet highlights the linear_pointwise operator in the compiled graph:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>%_linear_pointwise_default_140 : [num_users=2] = call_function[target=torch.ops.mkldnn._linear_pointwise.default](args = (%add_7, %_frozen_param278, %_frozen_param16, none, [], ), kwargs = {})
%mul_5 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%_linear_pointwise_default_140, 0.5), kwargs = {})
%mul_6 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%_linear_pointwise_default_140, 0.7071067811865476), kwargs = {})
%erf : [num_users=1] = call_function[target=torch.ops.aten.erf.default](args = (%mul_6,), kwargs = {})
%add_8 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%erf, 1), kwargs = {})
</code></pre></div></div>
<p>This completes the torch inductor changes required to compile the graph into optimized operators on AWS Graviton3 processors. Next comes the actual inference where the compiled graph is dispatched to be run. OneDNN with ACL was the backend we chose during the inductor compilation, so, the new operators were dispatched to oneDNN as expected, for example, <code class="language-plaintext highlighter-rouge">mkldnn._linear_pointwise</code>. However, due to gaps in oneDNN ACL primitives, the operators were run with C++ reference kernels instead of the optimized ACL kernels. Hence, the compile performance was still significantly behind the eager mode performance.</p>
<p>There were mainly three areas where oneDNN ACL primitives lack support for torch.compile mode. The following section talks about them in detail.</p>
<p><strong>1. ACL primitives didn’t have support for weights in blocked layout</strong></p>
<p>ACL primitives originally designed for eager mode supported weights only in the standard channels last (<a href="https://oneapi-src.github.io/oneDNN/dev_guide_understanding_memory_formats.html#nhwc">NHWC</a>) format, without any pre-packing. Whereas weights pre-packing into blocked layout is one of the main optimizations in the inductor compilation passes where the weights are reordered into blocks specific to the runtime platform. This avoids the redundant and on-the-fly reorders when running the General Matrix Multiplication (GEMM), which otherwise would be the bottleneck for inference performance. But the ACL primitives didn’t have support for blocked layout and hence the operators were run with oneDNN C++ reference kernels instead.</p>
<p><strong>2. Mixed precision primitives weren’t supported in oneDNN</strong></p>
<p>AWS Graviton3 processors support <a href="https://developer.arm.com/documentation/ddi0596/2020-12/SVE-Instructions/BFMMLA--BFloat16-floating-point-matrix-multiply-accumulate-">bfloat16 MMLA instructions</a> which can be used to accelerate fp32 inference with bfloat16 GEMM as a mixed precision compute. ACL supports bfloat16 mixed precision GEMM kernels, and are integrated into oneDNN as a fast math compute option for the existing fp32 operators. However, the fast math approach didn’t work for compile mode because of weights pre-packing optimization. The compile mode requires explicit mixed precision primitive implementation in oneDNN in order to use bfloat16 acceleration.</p>
<p><strong>3. ACL primitives didn’t support fused kernels for some of the activation functions</strong></p>
<p>In eager mode, operators are dispatched individually because the model is run independently as soon as it’s reached. Whereas in compile mode, operator fusion is another important optimization where the operators are fused for runtime efficiency. For example, Gaussian Error Linear Unit (<a href="https://arxiv.org/pdf/1606.08415.pdf#%3A~%3Atext%3DWe%20propose%20the%20Gaussian%20Error%2Cstandard%20Gaussian%20cumulative%20distribution%20function">GELU</a>) is one of the most widely used activation functions in transformers-based neural network architectures. So, it’s typical to have a linear layer (with matrix multiplications) followed by GELU activation. As part of compiling the model into efficient operators, the torch inductor fuses matmul and GELU into a single linearpointwise+gelu operator. However, oneDNN ACL primitives didn’t have the support for fused kernels with GELU.</p>
<p>We addressed these gaps by extending oneDNN primitives to handle the additional layouts and new primitive definitions. The following sections talk about the optimizations in detail.</p>
<h3 id="optimization-1-extended-acl-primitives-to-accept-weight-tensors-in-blocked-layout">Optimization 1: Extended ACL primitives to accept weight tensors in blocked layout</h3>
<p>We extended the ACL primitives to accept blocked layout in addition to the the standard NHWC format. The code snippet for this is as follows:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>const bool is_weights_md_format_ok
= utils::one_of(weights_format_kind_received,
format_kind::any, format_kind::blocked);
const memory_desc_t weights_md_received = weights_md_;
acl_utils::reorder_to_weight_format(aip.wei_tensor_info,
weights_md_, expected_weight_format, inner_dim, o_dim,
remaining_dims, {});
ACL_CHECK_SUPPORT(
(weights_format_kind_received == format_kind::blocked)
&& !(dnnl_memory_desc_equal(
&weights_md_received, &weights_md_)),
"specified blocked format not supported by ACL, use "
"format_kind_t::any to find a supported blocked format for "
"your platform");
</code></pre></div></div>
<h3 id="optimization-2-defined-new-acl-primitives-to-handle-mixed-precision-operators-weights-in-bfloat16-and-activations-in-fp32">Optimization 2: Defined new ACL primitives to handle mixed precision operators (weights in bfloat16 and activations in fp32)</h3>
<p>We defined mixed precision primitive definitions and updated the existing oneDNN ACL fp32 primitives to handle bfloat16 tensors.</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code> /* With graph compilation, we are able to reorder and pre-pack the weights during the model load
* and compilation phase itself so that redundant and on-the-fly reorders can be avoided.
* This primitive definition is to support gemm fastmath mode for the compile scenario where src is
* in fp32 and weights are in bf16
*/
{{forward, f32, bf16, f32}, {
CPU_INSTANCE_AARCH64_ACL(acl_inner_product_fwd_t)
nullptr,
}},
</code></pre></div></div>
<h3 id="optimization-3-disabled-operator-fusion-pass-in-torch-inductor">Optimization 3: Disabled operator fusion pass in torch inductor</h3>
<p>We bypassed the operator fusion pass in torch inductor so that the compiled graph doesn’t contain GELU fused operators. This is a temporary solution to enable ACL kernels in torch.compile. There is a work in progress to enable operator fusion pass for the future PyTorch releases. With this workaround, we were able to successfully dispatch the linear layer to ACL. As shown in the following torch.profiler output, the <code class="language-plaintext highlighter-rouge">aten::addmm </code>(one of the variants of the matmul operator) and <code class="language-plaintext highlighter-rouge">aten::gelu </code>in the original model (as highlighted in <em>Image 4</em>) was compiled to <code class="language-plaintext highlighter-rouge">mkldnn::_linear_pointwise </code>without <code class="language-plaintext highlighter-rouge">gelu </code>operator fusion (as highlighted in <em>Image 5</em>).</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>--------------------------- ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
--------------------------- ------------ ------------ ------------ ------------ ------------ ------------
aten::addmm 73.32% 46.543ms 74.49% 47.287ms 647.767us 73
model_inference 9.92% 6.296ms 100.00% 63.479ms 63.479ms 1
aten::bmm 4.37% 2.776ms 5.46% 3.467ms 144.458us 24
aten::copy_ 1.74% 1.102ms 1.74% 1.102ms 8.103us 136
aten::gelu 1.50% 950.000us 1.50% 950.000us 79.167us 12
</code></pre></div></div>
<p><em><strong>Image 4</strong>: torch.profiler output for Hugging Face bert base model inference in Eager mode, showing addmm and gelu operators</em></p>
<div class="mt-3 mb-3"> </div>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>----------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
----------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------
mkldnn::_linear_pointwise 53.61% 15.529ms 57.53% 16.665ms 228.288us 73
Torch-Compiled Region 36.95% 10.705ms 99.31% 28.769ms 28.769ms 1
aten::_scaled_dot_product_flash_attention_for_cpu 3.67% 1.064ms 4.43% 1.284ms 107.000us 12
aten::view 1.97% 572.000us 1.97% 572.000us 2.509us 228
aten::empty 1.38% 399.000us 1.38% 399.000us 3.270us 122
</code></pre></div></div>
<p><em><strong>Image 5</strong>: torch.profiler output for Hugging Face Bert base model inference in torch.compile mode, showing linear_pointwise operator without gelu fusion</em></p>
<p>Lastly, the <code class="language-plaintext highlighter-rouge">gelu </code>operator was compiled into <code class="language-plaintext highlighter-rouge">erf </code>(error function) and was dispatched to an inductor auto vectorization backend. The following snippets show the <code class="language-plaintext highlighter-rouge">erf </code>operator in the compiled graph and running it using <code class="language-plaintext highlighter-rouge">libm.so</code>.</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>%_linear_pointwise_default_140 : [num_users=2] = call_function[target=torch.ops.mkldnn._linear_pointwise.default](args = (%add_7, %_frozen_param278, %_frozen_param16, none, [], ), kwargs = {})
%mul_5 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%_linear_pointwise_default_140, 0.5), kwargs = {})
%mul_6 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%_linear_pointwise_default_140, 0.7071067811865476), kwargs = {})
%erf : [num_users=1] = call_function[target=torch.ops.aten.erf.default](args = (%mul_6,), kwargs = {})
%add_8 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%erf, 1), kwargs = {})
</code></pre></div></div>
<p><em><strong>Image 6</strong>: snippet after post grad pass showing erf function in the compiled graph</em></p>
<div class="mt-3 mb-3"> </div>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code> 0.82% 0.40% python3 libm.so.6 [.] erff32
0.05% 0.00% python3 libtorch_python.so [.] torch::autograd::THPVariable_erf
0.05% 0.00% python3 libtorch_cpu.so [.] at::_ops::erf::call
</code></pre></div></div>
<p><em><strong>Image 7</strong>: Linux perf report showing erf dispatch to libm.so</em></p>
<p>With this work, we were able to optimize <code class="language-plaintext highlighter-rouge">torch.compile </code>performance on Graviton3 processors by using inductor graph compilation along with the oneDNN+ACL backend.</p>
<h3 id="torchbench-enhancements">TorchBench enhancements</h3>
<p>To demonstrate the torch.compile performance improvements on AWS Graviton3 processors, we extended TorchBench framework to add a new argument to enable graph freeze and weights pre-packing and disable torch auto grad for eval test mode. The code snippet for this is as follows:</p>
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>parser.add_argument(
"—freeze_prepack_weights",
action='store_true',
help="set to freeze the graph and prepack weights",
)
if args.freeze_prepack_weights:
torch._inductor.config.freezing=True
torch._inductor.config.cpp.weight_prepack=True
</code></pre></div></div>
<p><em><strong>Image 8</strong>: Added freeze_prepack_weights option for torchdynamo backend in TorchBench to demonstrate torch.compile performance improvements on AWS Graviton3 processors</em></p>
<p>We have upstreamed all the optimizations, and starting with PyTorch 2.3.1, these are supported in torch Python wheels and AWS Graviton PyTorch DLC.</p>
<h2 id="whats-next">What’s next</h2>
<p>Next, we’re extending the torch inductor CPU backend support to compile Llama model, and adding support for fused GEMM kernels to enable torch inductor operator fusion optimization on AWS Graviton3 processors.</p>
<h2 id="conclusion">Conclusion</h2>
<p>In this tutorial, we covered how we optimized torch.compile performance on AWS Graviton3-based EC2 instances, how to use the optimizations to improve PyTorch model inference performance, and demonstrated the resulting speedups. We hope that you will give it a try! If you need any support with ML software on Graviton, please open an issue on the AWS Graviton Technical Guide <a href="https://github.com/aws/aws-graviton-getting-started">GitHub</a>.</p>
<h2 id="acknowledgements">Acknowledgements</h2>
<p>We would like to thank the PyTorch community for the baseline torch.compile framework and their continued efforts to optimize it further.</p>
<p>References: <a href="https://pytorch.org/assets/pytorch2-2.pdf">https://pytorch.org/assets/pytorch2-2.pdf</a></p>
<h2 id="author">Author</h2>
<p>Sunita Nadampalli is a Software Development Manager and AI/ML expert at AWS. She leads AWS Graviton software performance optimizations for AI/ML and HPC workloads. She is passionate about open source software development and delivering high-performance and sustainable software solutions for SoCs based on the Arm ISA.</p>
</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>
<a href="/events">Events</a>
</li>
<li>
<a href="/newsletter">Newsletter</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a>About</a>
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="/foundation">PyTorch Foundation</a>
</li>
<li>
<a href="/governing-board">Governing Board</a>
</li>
<li>
<a href="/credits">Cloud Credit Program</a>
</li>
<li>
<a href="/tac">Technical Advisory Council</a>
</li>
<li>
<a href="/staff">Staff</a>
</li>
<li>
<a href="/contact-us">Contact Us</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
<a href="/join">Become a Member</a>
</li>
<li class="resources-mobile-menu-title">
<a href="https://github.com/pytorch/pytorch" title="Go to PyTorch GitHub"><div id="topnav-gh-icon"></div></a>
</li>
</ul>
</div>
</div>
</div>
<script src="/assets/mobile-menu.js"></script>
<script src="/assets/scroll-to-anchor.js"></script>
<script src="/assets/external-links-new-tab.js"></script>
<script src="/assets/search-bar.js"></script>
<script src="/assets/cookie-banner.js"></script>