forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathquantization.html
1543 lines (1314 loc) · 99.5 KB
/
quantization.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta name="robots" content="noindex">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Quantization — PyTorch 1.10.1 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/quantization.html"/>
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="_static/css/jit.css" type="text/css" />
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="torch.nn.intrinsic" href="torch.nn.intrinsic.html" />
<link rel="prev" title="Pipeline Parallelism" href="pipeline.html" />
<!-- Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-117752657-2"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-117752657-2');
</script>
<!-- End Google Analytics -->
<script src="_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<a href="https://pytorch.org/mobile">Mobile</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="active docs-active">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-orange-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/audio/stable/index.html">
<span class="dropdown-title">torchaudio</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/text/stable/index.html">
<span class="dropdown-title">torchtext</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/vision/stable/index.html">
<span class="dropdown-title">torchvision</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/elastic/">
<span class="dropdown-title">TorchElastic</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/serve/">
<span class="dropdown-title">TorchServe</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/xla">
<span class="dropdown-title">PyTorch on XLA Devices</span>
<p></p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-arrow">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class="dropdown-title">About</span>
<p>Learn about PyTorch’s features and capabilities</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class="dropdown-title">Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered.</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class="dropdown-title">Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class="dropdown-title">Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/hub">
<span class="dropdown-title">Models (Beta)</span>
<p>Discover, publish, and reuse pre-trained models</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">GitHub</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>1.10.1 ▼</a>
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption"><span class="caption-text">Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="notes/amp_examples.html">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
</ul>
<p class="caption"><span class="caption-text">Python API</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="amp.html">torch.cuda.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="pipeline.html">Pipeline Parallelism</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="__config__.html">torch.__config__</a></li>
</ul>
<p class="caption"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="http://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</ul>
<p class="caption"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/governance.html">PyTorch Governance</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/persons_of_interest.html">PyTorch Governance | Persons of Interest</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="index.html">
Docs
</a> >
</li>
<li>Quantization</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/quantization.rst.txt" rel="nofollow"><img src="_static/images/view-page-source-icon.svg"></a>
</li>
</ul>
</div>
</div>
<div class="pytorch-shortcuts-wrapper" id="pytorch-shortcuts-wrapper">
Shortcuts
</div>
</div>
<section data-toggle="wy-nav-shift" id="pytorch-content-wrap" class="pytorch-content-wrap">
<div class="pytorch-content-left">
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<div class="section" id="quantization">
<span id="quantization-doc"></span><h1>Quantization<a class="headerlink" href="#quantization" title="Permalink to this headline">¶</a></h1>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Quantization is in beta and subject to change.</p>
</div>
<div class="section" id="introduction-to-quantization">
<h2>Introduction to Quantization<a class="headerlink" href="#introduction-to-quantization" title="Permalink to this headline">¶</a></h2>
<p>Quantization refers to techniques for performing computations and storing
tensors at lower bitwidths than floating point precision. A quantized model
executes some or all of the operations on tensors with integers rather than
floating point values. This allows for a more compact model representation and
the use of high performance vectorized operations on many hardware platforms.
PyTorch supports INT8 quantization compared to typical FP32 models allowing for
a 4x reduction in the model size and a 4x reduction in memory bandwidth
requirements. Hardware support for INT8 computations is typically 2 to 4
times faster compared to FP32 compute. Quantization is primarily a technique to
speed up inference and only the forward pass is supported for quantized
operators.</p>
<p>PyTorch supports multiple approaches to quantizing a deep learning model. In
most cases the model is trained in FP32 and then the model is converted to
INT8. In addition, PyTorch also supports quantization aware training, which
models quantization errors in both the forward and backward passes using
fake-quantization modules. Note that the entire computation is carried out in
floating point. At the end of quantization aware training, PyTorch provides
conversion functions to convert the trained model into lower precision.</p>
<p>At lower level, PyTorch provides a way to represent quantized tensors and
perform operations with them. They can be used to directly construct models
that perform all or part of the computation in lower precision. Higher-level
APIs are provided that incorporate typical workflows of converting FP32 model
to lower precision with minimal accuracy loss.</p>
<p>Quantization requires users to be aware of three concepts:</p>
<ol class="arabic simple">
<li><p>Quantization Config (Qconfig): Specifies how weights and activations are to be quantized. Qconfig is needed to create a quantized model.</p></li>
<li><p>Backend: Refers to kernels that support quantization, usually with different numerics.</p></li>
<li><p>Quantization engine (torch.backends.quantization.engine): When a quantized model is executed, the qengine specifies which backend is to be used for execution. It is important to ensure that the qengine is consistent with the Qconfig.</p></li>
</ol>
</div>
<div class="section" id="natively-supported-backends">
<h2>Natively supported backends<a class="headerlink" href="#natively-supported-backends" title="Permalink to this headline">¶</a></h2>
<p>Today, PyTorch supports the following backends for running quantized operators efficiently:</p>
<ul class="simple">
<li><p>x86 CPUs with AVX2 support or higher (without AVX2 some operations have
inefficient implementations), via <cite>fbgemm</cite> (<a class="reference external" href="https://github.com/pytorch/FBGEMM">https://github.com/pytorch/FBGEMM</a>).</p></li>
<li><p>ARM CPUs (typically found in mobile/embedded devices), via
<cite>qnnpack</cite> (<a class="reference external" href="https://github.com/pytorch/QNNPACK">https://github.com/pytorch/QNNPACK</a>).</p></li>
</ul>
<p>The corresponding implementation is chosen automatically based on the PyTorch build mode, though users
have the option to override this by setting <cite>torch.backends.quantization.engine</cite> to <cite>fbgemm</cite> or <cite>qnnpack</cite>.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>At the moment PyTorch doesn’t provide quantized operator implementations on CUDA -
this is the direction for future work. Move the model to CPU in order to test the
quantized functionality.</p>
<p>Quantization-aware training (through <a class="reference internal" href="torch.quantization.html#torch.quantization.FakeQuantize" title="torch.quantization.FakeQuantize"><code class="xref py py-class docutils literal notranslate"><span class="pre">FakeQuantize</span></code></a>,
which emulates quantized numerics in fp32) supports both CPU and CUDA.</p>
</div>
<p>When preparing a quantized model, it is necessary to ensure that qconfig
and the engine used for quantized computations match the backend on which
the model will be executed. The qconfig controls the type of observers used
during the quantization passes. The qengine controls whether <cite>fbgemm</cite> or
<cite>qnnpack</cite> specific packing function is used when packing weights for linear
and convolution functions and modules. For example:</p>
<p>Default settings for fbgemm:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># set the qconfig for PTQ</span>
<span class="n">qconfig</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_default_qconfig</span><span class="p">(</span><span class="s1">'fbgemm'</span><span class="p">)</span>
<span class="c1"># or, set the qconfig for QAT</span>
<span class="n">qconfig</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_default_qat_qconfig</span><span class="p">(</span><span class="s1">'fbgemm'</span><span class="p">)</span>
<span class="c1"># set the qengine to control weight packing</span>
<span class="n">torch</span><span class="o">.</span><span class="n">backends</span><span class="o">.</span><span class="n">quantized</span><span class="o">.</span><span class="n">engine</span> <span class="o">=</span> <span class="s1">'fbgemm'</span>
</pre></div>
</div>
<p>Default settings for qnnpack:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># set the qconfig for PTQ</span>
<span class="n">qconfig</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_default_qconfig</span><span class="p">(</span><span class="s1">'qnnpack'</span><span class="p">)</span>
<span class="c1"># or, set the qconfig for QAT</span>
<span class="n">qconfig</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_default_qat_qconfig</span><span class="p">(</span><span class="s1">'qnnpack'</span><span class="p">)</span>
<span class="c1"># set the qengine to control weight packing</span>
<span class="n">torch</span><span class="o">.</span><span class="n">backends</span><span class="o">.</span><span class="n">quantized</span><span class="o">.</span><span class="n">engine</span> <span class="o">=</span> <span class="s1">'qnnpack'</span>
</pre></div>
</div>
</div>
<div class="section" id="quantization-api-summary">
<h2>Quantization API Summary<a class="headerlink" href="#quantization-api-summary" title="Permalink to this headline">¶</a></h2>
<p>PyTorch provides two different modes of quantization: Eager Mode Quantization and FX Graph Mode Quantization.</p>
<p>Eager Mode Quantization is a beta feature. User needs to do fusion and specify where quantization and dequantization happens manually, also it only supports modules and not functionals.</p>
<p>FX Graph Mode Quantization is a new automated quantization framework in PyTorch, and currently it’s a prototype feature. It improves upon Eager Mode Quantization by adding support for functionals and automating the quantization process, although people might need to refactor the model to make the model compatible with FX Graph Mode Quantization (symbolically traceable with <code class="docutils literal notranslate"><span class="pre">torch.fx</span></code>). Note that FX Graph Mode Quantization is not expected to work on arbitrary models since the model might not be symbolically traceable, we will integrate it into domain libraries like torchvision and users will be able to quantize models similar to the ones in supported domain libraries with FX Graph Mode Quantization. For arbitrary models we’ll provide general guidelines, but to actually make it work, users might need to be familiar with <code class="docutils literal notranslate"><span class="pre">torch.fx</span></code>, especially on how to make a model symbolically traceable.</p>
<p>New users of quantization are encouraged to try out FX Graph Mode Quantization first, if it does not work, user may try to follow the guideline of <a class="reference external" href="https://pytorch.org/tutorials/prototype/fx_graph_mode_quant_guide.html">using FX Graph Mode Quantization</a> or fall back to eager mode quantization.</p>
<p>The following table compares the differences between Eager Mode Quantization and FX Graph Mode Quantization:</p>
<table class="docutils colwidths-auto align-default">
<tbody>
<tr class="row-odd"><td></td>
<td><p>Eager Mode
Quantization</p></td>
<td><p>FX Graph
Mode
Quantization</p></td>
</tr>
<tr class="row-even"><td><p>Release
Status</p></td>
<td><p>beta</p></td>
<td><p>prototype</p></td>
</tr>
<tr class="row-odd"><td><p>Operator
Fusion</p></td>
<td><p>Manual</p></td>
<td><p>Automatic</p></td>
</tr>
<tr class="row-even"><td><p>Quant/DeQuant
Placement</p></td>
<td><p>Manual</p></td>
<td><p>Automatic</p></td>
</tr>
<tr class="row-odd"><td><p>Quantizing
Modules</p></td>
<td><p>Supported</p></td>
<td><p>Supported</p></td>
</tr>
<tr class="row-even"><td><p>Quantizing
Functionals/Torch
Ops</p></td>
<td><p>Manual</p></td>
<td><p>Automatic</p></td>
</tr>
<tr class="row-odd"><td><p>Support for
Customization</p></td>
<td><p>Limited Support</p></td>
<td><p>Fully
Supported</p></td>
</tr>
<tr class="row-even"><td><p>Quantization Mode
Support</p></td>
<td><p>Post Training
Quantization:
Static, Dynamic,
Weight Only</p>
<p>Quantiztion Aware
Training:
Static</p>
</td>
<td><p>Post Training
Quantization:
Static, Dynamic,
Weight Only</p>
<p>Quantiztion Aware
Training:
Static</p>
</td>
</tr>
<tr class="row-odd"><td><p>Input/Output
Model Type</p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">torch.nn.Module</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">torch.nn.Module</span></code>
(May need some
refactors to make
the model
compatible with FX
Graph Mode
Quantization)</p></td>
</tr>
</tbody>
</table>
<p>There are three types of quantization supported:</p>
<ol class="arabic simple">
<li><p>dynamic quantization (weights quantized with activations read/stored in
floating point and quantized for compute.)</p></li>
<li><p>static quantization (weights quantized, activations quantized, calibration
required post training)</p></li>
<li><p>static quantization aware training (weights quantized, activations quantized,
quantization numerics modeled during training)</p></li>
</ol>
<p>Please see our <a class="reference external" href="https://pytorch.org/blog/introduction-to-quantization-on-pytorch/">Introduction to Quantization on Pytorch</a> blog post
for a more comprehensive overview of the tradeoffs between these quantization
types.</p>
<p>Operator coverage varies between dynamic and static quantization and is captured in the table below.
Note that for FX quantization, the corresponding functionals are also supported.</p>
<table class="docutils colwidths-auto align-default">
<tbody>
<tr class="row-odd"><td></td>
<td><p>Static
Quantization</p></td>
<td><p>Dynamic
Quantization</p></td>
</tr>
<tr class="row-even"><td><div class="line-block">
<div class="line">nn.Linear</div>
<div class="line">nn.Conv1d/2d/3d</div>
</div>
</td>
<td><div class="line-block">
<div class="line">Y</div>
<div class="line">Y</div>
</div>
</td>
<td><div class="line-block">
<div class="line">Y</div>
<div class="line">N</div>
</div>
</td>
</tr>
<tr class="row-odd"><td><div class="line-block">
<div class="line">nn.LSTM</div>
<div class="line">nn.GRU</div>
</div>
</td>
<td><div class="line-block">
<div class="line">N</div>
<div class="line">N</div>
</div>
</td>
<td><div class="line-block">
<div class="line">Y</div>
<div class="line">Y</div>
</div>
</td>
</tr>
<tr class="row-even"><td><div class="line-block">
<div class="line">nn.RNNCell</div>
<div class="line">nn.GRUCell</div>
<div class="line">nn.LSTMCell</div>
</div>
</td>
<td><div class="line-block">
<div class="line">N</div>
<div class="line">N</div>
<div class="line">N</div>
</div>
</td>
<td><div class="line-block">
<div class="line">Y</div>
<div class="line">Y</div>
<div class="line">Y</div>
</div>
</td>
</tr>
<tr class="row-odd"><td><p>nn.EmbeddingBag</p></td>
<td><p>Y (activations
are in fp32)</p></td>
<td><p>Y</p></td>
</tr>
<tr class="row-even"><td><p>nn.Embedding</p></td>
<td><p>Y</p></td>
<td><p>N</p></td>
</tr>
<tr class="row-odd"><td><p>nn.MultiheadAttention</p></td>
<td><p>Not Supported</p></td>
<td><p>Not supported</p></td>
</tr>
<tr class="row-even"><td><p>Activations</p></td>
<td><p>Broadly supported</p></td>
<td><p>Un-changed,
computations
stay in fp32</p></td>
</tr>
</tbody>
</table>
<div class="section" id="eager-mode-quantization">
<h3>Eager Mode Quantization<a class="headerlink" href="#eager-mode-quantization" title="Permalink to this headline">¶</a></h3>
<div class="section" id="dynamic-quantization">
<h4>Dynamic Quantization<a class="headerlink" href="#dynamic-quantization" title="Permalink to this headline">¶</a></h4>
<p>This is the simplest to apply form of quantization where the weights are
quantized ahead of time but the activations are dynamically quantized
during inference. This is used for situations where the model execution time
is dominated by loading weights from memory rather than computing the matrix
multiplications. This is true for for LSTM and Transformer type models with
small batch size.</p>
<p>Diagram:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># original model</span>
<span class="c1"># all tensors and computations are in floating point</span>
<span class="n">previous_layer_fp32</span> <span class="o">--</span> <span class="n">linear_fp32</span> <span class="o">--</span> <span class="n">activation_fp32</span> <span class="o">--</span> <span class="n">next_layer_fp32</span>
<span class="o">/</span>
<span class="n">linear_weight_fp32</span>
<span class="c1"># dynamically quantized model</span>
<span class="c1"># linear and LSTM weights are in int8</span>
<span class="n">previous_layer_fp32</span> <span class="o">--</span> <span class="n">linear_int8_w_fp32_inp</span> <span class="o">--</span> <span class="n">activation_fp32</span> <span class="o">--</span> <span class="n">next_layer_fp32</span>
<span class="o">/</span>
<span class="n">linear_weight_int8</span>
</pre></div>
</div>
<p>API example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="c1"># define a floating point model</span>
<span class="k">class</span> <span class="nc">M</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">M</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">fc</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">fc</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="n">x</span>
<span class="c1"># create a model instance</span>
<span class="n">model_fp32</span> <span class="o">=</span> <span class="n">M</span><span class="p">()</span>
<span class="c1"># create a quantized model instance</span>
<span class="n">model_int8</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">quantize_dynamic</span><span class="p">(</span>
<span class="n">model_fp32</span><span class="p">,</span> <span class="c1"># the original model</span>
<span class="p">{</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">},</span> <span class="c1"># a set of layers to dynamically quantize</span>
<span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">qint8</span><span class="p">)</span> <span class="c1"># the target dtype for quantized weights</span>
<span class="c1"># run the model</span>
<span class="n">input_fp32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">4</span><span class="p">,</span> <span class="mi">4</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
<span class="n">res</span> <span class="o">=</span> <span class="n">model_int8</span><span class="p">(</span><span class="n">input_fp32</span><span class="p">)</span>
</pre></div>
</div>
<p>To learn more about dynamic quantization please see our <a class="reference external" href="https://pytorch.org/tutorials/recipes/recipes/dynamic_quantization.html">dynamic quantization tutorial</a>.</p>
</div>
<div class="section" id="static-quantization">
<h4>Static Quantization<a class="headerlink" href="#static-quantization" title="Permalink to this headline">¶</a></h4>
<p>Static quantization quantizes the weights and activations of the model. It
fuses activations into preceding layers where possible. It requires
calibration with a representative dataset to determine optimal quantization
parameters for activations. Post Training Quantization is typically used when
both memory bandwidth and compute savings are important with CNNs being a
typical use case. Static quantization is also known as Post Training
Quantization or PTQ.</p>
<p>Diagram:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># original model</span>
<span class="c1"># all tensors and computations are in floating point</span>
<span class="n">previous_layer_fp32</span> <span class="o">--</span> <span class="n">linear_fp32</span> <span class="o">--</span> <span class="n">activation_fp32</span> <span class="o">--</span> <span class="n">next_layer_fp32</span>
<span class="o">/</span>
<span class="n">linear_weight_fp32</span>
<span class="c1"># statically quantized model</span>
<span class="c1"># weights and activations are in int8</span>
<span class="n">previous_layer_int8</span> <span class="o">--</span> <span class="n">linear_with_activation_int8</span> <span class="o">--</span> <span class="n">next_layer_int8</span>
<span class="o">/</span>
<span class="n">linear_weight_int8</span>
</pre></div>
</div>
<p>API Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="c1"># define a floating point model where some layers could be statically quantized</span>
<span class="k">class</span> <span class="nc">M</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">M</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="c1"># QuantStub converts tensors from floating point to quantized</span>
<span class="bp">self</span><span class="o">.</span><span class="n">quant</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">QuantStub</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">conv</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">relu</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">()</span>
<span class="c1"># DeQuantStub converts tensors from quantized to floating point</span>
<span class="bp">self</span><span class="o">.</span><span class="n">dequant</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">DeQuantStub</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="c1"># manually specify where tensors will be converted from floating</span>
<span class="c1"># point to quantized in the quantized model</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="c1"># manually specify where tensors will be converted from quantized</span>
<span class="c1"># to floating point in the quantized model</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">dequant</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="n">x</span>
<span class="c1"># create a model instance</span>
<span class="n">model_fp32</span> <span class="o">=</span> <span class="n">M</span><span class="p">()</span>
<span class="c1"># model must be set to eval mode for static quantization logic to work</span>
<span class="n">model_fp32</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="c1"># attach a global qconfig, which contains information about what kind</span>
<span class="c1"># of observers to attach. Use 'fbgemm' for server inference and</span>
<span class="c1"># 'qnnpack' for mobile inference. Other quantization configurations such</span>
<span class="c1"># as selecting symmetric or assymetric quantization and MinMax or L2Norm</span>
<span class="c1"># calibration techniques can be specified here.</span>
<span class="n">model_fp32</span><span class="o">.</span><span class="n">qconfig</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_default_qconfig</span><span class="p">(</span><span class="s1">'fbgemm'</span><span class="p">)</span>
<span class="c1"># Fuse the activations to preceding layers, where applicable.</span>
<span class="c1"># This needs to be done manually depending on the model architecture.</span>
<span class="c1"># Common fusions include `conv + relu` and `conv + batchnorm + relu`</span>
<span class="n">model_fp32_fused</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">fuse_modules</span><span class="p">(</span><span class="n">model_fp32</span><span class="p">,</span> <span class="p">[[</span><span class="s1">'conv'</span><span class="p">,</span> <span class="s1">'relu'</span><span class="p">]])</span>
<span class="c1"># Prepare the model for static quantization. This inserts observers in</span>
<span class="c1"># the model that will observe activation tensors during calibration.</span>
<span class="n">model_fp32_prepared</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">prepare</span><span class="p">(</span><span class="n">model_fp32_fused</span><span class="p">)</span>
<span class="c1"># calibrate the prepared model to determine quantization parameters for activations</span>
<span class="c1"># in a real world setting, the calibration would be done with a representative dataset</span>
<span class="n">input_fp32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">4</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
<span class="n">model_fp32_prepared</span><span class="p">(</span><span class="n">input_fp32</span><span class="p">)</span>
<span class="c1"># Convert the observed model to a quantized model. This does several things:</span>
<span class="c1"># quantizes the weights, computes and stores the scale and bias value to be</span>
<span class="c1"># used with each activation tensor, and replaces key operators with quantized</span>
<span class="c1"># implementations.</span>
<span class="n">model_int8</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">convert</span><span class="p">(</span><span class="n">model_fp32_prepared</span><span class="p">)</span>
<span class="c1"># run the model, relevant calculations will happen in int8</span>
<span class="n">res</span> <span class="o">=</span> <span class="n">model_int8</span><span class="p">(</span><span class="n">input_fp32</span><span class="p">)</span>
</pre></div>
</div>
<p>To learn more about static quantization, please see the <a class="reference external" href="https://pytorch.org/tutorials/advanced/static_quantization_tutorial.html">static quantization tutorial</a>.</p>
</div>
<div class="section" id="quantization-aware-training">
<h4>Quantization Aware Training<a class="headerlink" href="#quantization-aware-training" title="Permalink to this headline">¶</a></h4>
<p>Quantization Aware Training models the effects of quantization during training
allowing for higher accuracy compared to other quantization methods. During
training, all calculations are done in floating point, with fake_quant modules
modeling the effects of quantization by clamping and rounding to simulate the
effects of INT8. After model conversion, weights and
activations are quantized, and activations are fused into the preceding layer
where possible. It is commonly used with CNNs and yields a higher accuracy
compared to static quantization. Quantization Aware Training is also known as
QAT.</p>
<p>Diagram:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># original model</span>
<span class="c1"># all tensors and computations are in floating point</span>
<span class="n">previous_layer_fp32</span> <span class="o">--</span> <span class="n">linear_fp32</span> <span class="o">--</span> <span class="n">activation_fp32</span> <span class="o">--</span> <span class="n">next_layer_fp32</span>
<span class="o">/</span>
<span class="n">linear_weight_fp32</span>
<span class="c1"># model with fake_quants for modeling quantization numerics during training</span>
<span class="n">previous_layer_fp32</span> <span class="o">--</span> <span class="n">fq</span> <span class="o">--</span> <span class="n">linear_fp32</span> <span class="o">--</span> <span class="n">activation_fp32</span> <span class="o">--</span> <span class="n">fq</span> <span class="o">--</span> <span class="n">next_layer_fp32</span>
<span class="o">/</span>
<span class="n">linear_weight_fp32</span> <span class="o">--</span> <span class="n">fq</span>
<span class="c1"># quantized model</span>
<span class="c1"># weights and activations are in int8</span>
<span class="n">previous_layer_int8</span> <span class="o">--</span> <span class="n">linear_with_activation_int8</span> <span class="o">--</span> <span class="n">next_layer_int8</span>
<span class="o">/</span>
<span class="n">linear_weight_int8</span>
</pre></div>
</div>
<p>API Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="c1"># define a floating point model where some layers could benefit from QAT</span>
<span class="k">class</span> <span class="nc">M</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">M</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="c1"># QuantStub converts tensors from floating point to quantized</span>
<span class="bp">self</span><span class="o">.</span><span class="n">quant</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">QuantStub</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">conv</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">bn</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm2d</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">relu</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">()</span>
<span class="c1"># DeQuantStub converts tensors from quantized to floating point</span>
<span class="bp">self</span><span class="o">.</span><span class="n">dequant</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">DeQuantStub</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">conv</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">bn</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">dequant</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="n">x</span>
<span class="c1"># create a model instance</span>
<span class="n">model_fp32</span> <span class="o">=</span> <span class="n">M</span><span class="p">()</span>
<span class="c1"># model must be set to train mode for QAT logic to work</span>
<span class="n">model_fp32</span><span class="o">.</span><span class="n">train</span><span class="p">()</span>
<span class="c1"># attach a global qconfig, which contains information about what kind</span>
<span class="c1"># of observers to attach. Use 'fbgemm' for server inference and</span>
<span class="c1"># 'qnnpack' for mobile inference. Other quantization configurations such</span>
<span class="c1"># as selecting symmetric or assymetric quantization and MinMax or L2Norm</span>
<span class="c1"># calibration techniques can be specified here.</span>
<span class="n">model_fp32</span><span class="o">.</span><span class="n">qconfig</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_default_qat_qconfig</span><span class="p">(</span><span class="s1">'fbgemm'</span><span class="p">)</span>
<span class="c1"># fuse the activations to preceding layers, where applicable</span>
<span class="c1"># this needs to be done manually depending on the model architecture</span>
<span class="n">model_fp32_fused</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">fuse_modules</span><span class="p">(</span><span class="n">model_fp32</span><span class="p">,</span>
<span class="p">[[</span><span class="s1">'conv'</span><span class="p">,</span> <span class="s1">'bn'</span><span class="p">,</span> <span class="s1">'relu'</span><span class="p">]])</span>
<span class="c1"># Prepare the model for QAT. This inserts observers and fake_quants in</span>
<span class="c1"># the model that will observe weight and activation tensors during calibration.</span>
<span class="n">model_fp32_prepared</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">prepare_qat</span><span class="p">(</span><span class="n">model_fp32_fused</span><span class="p">)</span>
<span class="c1"># run the training loop (not shown)</span>
<span class="n">training_loop</span><span class="p">(</span><span class="n">model_fp32_prepared</span><span class="p">)</span>
<span class="c1"># Convert the observed model to a quantized model. This does several things:</span>
<span class="c1"># quantizes the weights, computes and stores the scale and bias value to be</span>
<span class="c1"># used with each activation tensor, fuses modules where appropriate,</span>
<span class="c1"># and replaces key operators with quantized implementations.</span>
<span class="n">model_fp32_prepared</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="n">model_int8</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">convert</span><span class="p">(</span><span class="n">model_fp32_prepared</span><span class="p">)</span>
<span class="c1"># run the model, relevant calculations will happen in int8</span>
<span class="n">res</span> <span class="o">=</span> <span class="n">model_int8</span><span class="p">(</span><span class="n">input_fp32</span><span class="p">)</span>
</pre></div>
</div>
<p>To learn more about quantization aware training, please see the <a class="reference external" href="https://pytorch.org/tutorials/advanced/static_quantization_tutorial.html">QAT
tutorial</a>.</p>
</div>
</div>
<div class="section" id="prototype-fx-graph-mode-quantization">
<h3>(Prototype) FX Graph Mode Quantization<a class="headerlink" href="#prototype-fx-graph-mode-quantization" title="Permalink to this headline">¶</a></h3>
<p>Quantization types supported by FX Graph Mode can be classified in two ways:</p>
<ol class="arabic simple">
<li><p>Post Training Quantization (apply quantization after training, quantization parameters are calculated based on sample calibration data)</p></li>
<li><p>Quantization Aware Training (simulate quantization during training so that the quantization parameters can be learned together with the model using training data)</p></li>
</ol>
<p>And then each of these two may include any or all of the following types:</p>
<ul class="simple">
<li><p>Weight Only Quantization (only weight is statically quantized)</p></li>
<li><p>Dynamic Quantization (weight is statically quantized, activation is dynamically quantized)</p></li>
<li><p>Static Quantization (both weight and activations are statically quantized)</p></li>
</ul>
<p>These two ways of classification are independent, so theoretically we can have 6 different types of quantization.</p>
<p>The supported quantization types in FX Graph Mode Quantization are:</p>
<ul class="simple">
<li><p>Post Training Quantization</p>
<ul>
<li><p>Weight Only Quantization</p></li>
<li><p>Dynamic Quantization</p></li>
<li><p>Static Quantization</p></li>
</ul>
</li>
<li><p>Quantization Aware Training</p>
<ul>
<li><p>Static Quantization</p></li>
</ul>
</li>
</ul>
<p>There are multiple quantization types in post training quantization (weight only, dynamic and static) and the configuration is done through <cite>qconfig_dict</cite> (an argument of the <cite>prepare_fx</cite> function).</p>
<p>API Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch.quantization.quantize_fx</span> <span class="k">as</span> <span class="nn">quantize_fx</span>
<span class="kn">import</span> <span class="nn">copy</span>
<span class="n">model_fp</span> <span class="o">=</span> <span class="n">UserModel</span><span class="p">(</span><span class="o">...</span><span class="p">)</span>
<span class="c1">#</span>
<span class="c1"># post training dynamic/weight_only quantization</span>
<span class="c1">#</span>
<span class="c1"># we need to deepcopy if we still want to keep model_fp unchanged after quantization since quantization apis change the input model</span>
<span class="n">model_to_quantize</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="n">model_fp</span><span class="p">)</span>
<span class="n">model_to_quantize</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="n">qconfig_dict</span> <span class="o">=</span> <span class="p">{</span><span class="s2">""</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">default_dynamic_qconfig</span><span class="p">}</span>
<span class="c1"># prepare</span>
<span class="n">model_prepared</span> <span class="o">=</span> <span class="n">quantize_fx</span><span class="o">.</span><span class="n">prepare_fx</span><span class="p">(</span><span class="n">model_to_quantize</span><span class="p">,</span> <span class="n">qconfig_dict</span><span class="p">)</span>
<span class="c1"># no calibration needed when we only have dynamici/weight_only quantization</span>
<span class="c1"># quantize</span>
<span class="n">model_quantized</span> <span class="o">=</span> <span class="n">quantize_fx</span><span class="o">.</span><span class="n">convert_fx</span><span class="p">(</span><span class="n">model_prepared</span><span class="p">)</span>
<span class="c1">#</span>
<span class="c1"># post training static quantization</span>
<span class="c1">#</span>
<span class="n">model_to_quantize</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="n">model_fp</span><span class="p">)</span>
<span class="n">qconfig_dict</span> <span class="o">=</span> <span class="p">{</span><span class="s2">""</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_default_qconfig</span><span class="p">(</span><span class="s1">'qnnpack'</span><span class="p">)}</span>
<span class="n">model_to_quantize</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="c1"># prepare</span>
<span class="n">model_prepared</span> <span class="o">=</span> <span class="n">quantize_fx</span><span class="o">.</span><span class="n">prepare_fx</span><span class="p">(</span><span class="n">model_to_quantize</span><span class="p">,</span> <span class="n">qconfig_dict</span><span class="p">)</span>
<span class="c1"># calibrate (not shown)</span>
<span class="c1"># quantize</span>
<span class="n">model_quantized</span> <span class="o">=</span> <span class="n">quantize_fx</span><span class="o">.</span><span class="n">convert_fx</span><span class="p">(</span><span class="n">model_prepared</span><span class="p">)</span>
<span class="c1">#</span>
<span class="c1"># quantization aware training for static quantization</span>
<span class="c1">#</span>
<span class="n">model_to_quantize</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="n">model_fp</span><span class="p">)</span>
<span class="n">qconfig_dict</span> <span class="o">=</span> <span class="p">{</span><span class="s2">""</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_default_qat_qconfig</span><span class="p">(</span><span class="s1">'qnnpack'</span><span class="p">)}</span>
<span class="n">model_to_quantize</span><span class="o">.</span><span class="n">train</span><span class="p">()</span>
<span class="c1"># prepare</span>
<span class="n">model_prepared</span> <span class="o">=</span> <span class="n">quantize_fx</span><span class="o">.</span><span class="n">prepare_qat_fx</span><span class="p">(</span><span class="n">model_to_quantize</span><span class="p">,</span> <span class="n">qconfig_dict</span><span class="p">)</span>
<span class="c1"># training loop (not shown)</span>
<span class="c1"># quantize</span>
<span class="n">model_quantized</span> <span class="o">=</span> <span class="n">quantize_fx</span><span class="o">.</span><span class="n">convert_fx</span><span class="p">(</span><span class="n">model_prepared</span><span class="p">)</span>
<span class="c1">#</span>
<span class="c1"># fusion</span>
<span class="c1">#</span>
<span class="n">model_to_quantize</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="n">model_fp</span><span class="p">)</span>
<span class="n">model_fused</span> <span class="o">=</span> <span class="n">quantize_fx</span><span class="o">.</span><span class="n">fuse_fx</span><span class="p">(</span><span class="n">model_to_quantize</span><span class="p">)</span>
</pre></div>
</div>
<p>Please see the following tutorials for more information about FX Graph Mode Quantization:</p>
<ul class="simple">
<li><p><a class="reference external" href="https://pytorch.org/tutorials/prototype/fx_graph_mode_quant_guide.html">User Guide on Using FX Graph Mode Quantization</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/tutorials/prototype/fx_graph_mode_ptq_static.html">FX Graph Mode Post Training Static Quantization</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/tutorials/prototype/fx_graph_mode_ptq_dynamic.html">FX Graph Mode Post Training Dynamic Quantization</a></p></li>
</ul>
</div>
</div>
<div class="section" id="quantized-tensors">
<h2>Quantized Tensors<a class="headerlink" href="#quantized-tensors" title="Permalink to this headline">¶</a></h2>
<p>PyTorch supports both per tensor and per channel asymmetric linear
quantization. Per tensor means that all the values within the tensor are
scaled the same way. Per channel means that for each dimension, typically
the channel dimension of a tensor, the values
in the tensor are scaled and offset by a different value (effectively
the scale and offset become vectors). This allows for lesser error in converting tensors
to quantized values.</p>
<p>The mapping is performed by converting the floating point tensors using</p>
<a class="reference internal image-reference" href="_images/math-quantizer-equation.png"><img alt="_images/math-quantizer-equation.png" src="_images/math-quantizer-equation.png" style="width: 40%;" /></a>
<p>Note that, we ensure that zero in floating point is represented with no error
after quantization, thereby ensuring that operations like padding do not cause
additional quantization error.</p>
<p>In order to do quantization in PyTorch, we need to be able to represent
quantized data in Tensors. A Quantized Tensor allows for storing
quantized data (represented as int8/uint8/int32) along with quantization
parameters like scale and zero_point. Quantized Tensors allow for many
useful operations making quantized arithmetic easy, in addition to
allowing for serialization of data in a quantized format.</p>
</div>
<div class="section" id="quantization-operation-coverage">
<h2>Quantization Operation coverage<a class="headerlink" href="#quantization-operation-coverage" title="Permalink to this headline">¶</a></h2>
<p>Quantized Tensors support a limited subset of data manipulation methods of the
regular full-precision tensor. For NN operators included in PyTorch, we
restrict support to:</p>
<blockquote>