forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathamp.html
1542 lines (1314 loc) · 127 KB
/
amp.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="generator" content="Docutils 0.18.1: http://docutils.sourceforge.net/" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Automatic Mixed Precision package - torch.amp — PyTorch 1.13 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/amp.html"/>
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="_static/sphinx-dropdown.css" type="text/css" />
<link rel="stylesheet" href="_static/panels-bootstrap.min.css" type="text/css" />
<link rel="stylesheet" href="_static/css/jit.css" type="text/css" />
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="Automatic differentiation package - torch.autograd" href="autograd.html" />
<link rel="prev" title="Tensor Views" href="tensor_view.html" />
<!-- Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-117752657-2"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-117752657-2');
</script>
<!-- End Google Analytics -->
<script src="_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<a href="https://pytorch.org/mobile">Mobile</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="active docs-active">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-orange-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/audio/stable/index.html">
<span class="dropdown-title">torchaudio</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/text/stable/index.html">
<span class="dropdown-title">torchtext</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/vision/stable/index.html">
<span class="dropdown-title">torchvision</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torcharrow">
<span class="dropdown-title">torcharrow</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/data">
<span class="dropdown-title">TorchData</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torchrec">
<span class="dropdown-title">TorchRec</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/serve/">
<span class="dropdown-title">TorchServe</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torchx/">
<span class="dropdown-title">TorchX</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/xla">
<span class="dropdown-title">PyTorch on XLA Devices</span>
<p></p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-arrow">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class="dropdown-title">About</span>
<p>Learn about PyTorch’s features and capabilities</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn about the PyTorch foundation</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class="dropdown-title">Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered.</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/community-stories">
<span class="dropdown-title">Community Stories</span>
<p>Learn how our community solves real, everyday machine learning problems with PyTorch.</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class="dropdown-title">Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class="dropdown-title">Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/hub">
<span class="dropdown-title">Models (Beta)</span>
<p>Discover, publish, and reuse pre-trained models</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">GitHub</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>1.13 ▼</a>
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption" role="heading"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="community/build_ci_governance.html">PyTorch Governance | Build + CI</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/design.html">PyTorch Design Philosophy</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/governance.html">PyTorch Governance | Mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/persons_of_interest.html">PyTorch Governance | Maintainers</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Developer Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="notes/amp_examples.html">CUDA Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/mps.html">MPS backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/numerical_accuracy.html">Numerical accuracy</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
<li class="toctree-l1"><a class="reference internal" href="deploy.html">torch::deploy</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python API</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">torch.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="library.html">torch.library</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="fsdp.html">torch.distributed.fsdp</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="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="monitor.html">torch.monitor</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"><a class="reference internal" href="quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="masked.html">torch.masked</a></li>
<li class="toctree-l1"><a class="reference internal" href="nested.html">torch.nested</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit_utils.html">torch.utils.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="config_mod.html">torch.__config__</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/data">TorchData</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/torchrec">TorchRec</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="http://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="index.html">
Docs
</a> >
</li>
<li>Automatic Mixed Precision package - torch.amp</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/amp.rst.txt" rel="nofollow"><img src="_static/images/view-page-source-icon.svg"></a>
</li>
</ul>
</div>
</div>
<div class="pytorch-shortcuts-wrapper" id="pytorch-shortcuts-wrapper">
Shortcuts
</div>
</div>
<section data-toggle="wy-nav-shift" id="pytorch-content-wrap" class="pytorch-content-wrap">
<div class="pytorch-content-left">
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<section id="automatic-mixed-precision-package-torch-amp">
<h1>Automatic Mixed Precision package - torch.amp<a class="headerlink" href="#automatic-mixed-precision-package-torch-amp" title="Permalink to this heading">¶</a></h1>
<span class="target" id="module-torch.cpu"></span><span class="target" id="module-torch.cpu.amp"></span><span class="target" id="module-torch.cuda.amp"></span><span class="target" id="module-torch.amp"></span><p><a class="reference internal" href="#module-torch.amp" title="torch.amp"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.amp</span></code></a> provides convenience methods for mixed precision,
where some operations use the <code class="docutils literal notranslate"><span class="pre">torch.float32</span></code> (<code class="docutils literal notranslate"><span class="pre">float</span></code>) datatype and other operations
use lower precision floating point datatype (<code class="docutils literal notranslate"><span class="pre">lower_precision_fp</span></code>): <code class="docutils literal notranslate"><span class="pre">torch.float16</span></code> (<code class="docutils literal notranslate"><span class="pre">half</span></code>) or <code class="docutils literal notranslate"><span class="pre">torch.bfloat16</span></code>. Some ops, like linear layers and convolutions,
are much faster in <code class="docutils literal notranslate"><span class="pre">lower_precision_fp</span></code>. Other ops, like reductions, often require the dynamic
range of <code class="docutils literal notranslate"><span class="pre">float32</span></code>. Mixed precision tries to match each op to its appropriate datatype.</p>
<p>Ordinarily, “automatic mixed precision training” with datatype of <code class="docutils literal notranslate"><span class="pre">torch.float16</span></code> uses <a class="reference internal" href="#torch.autocast" title="torch.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.autocast</span></code></a> and
<a class="reference internal" href="#torch.cuda.amp.GradScaler" title="torch.cuda.amp.GradScaler"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.cuda.amp.GradScaler</span></code></a> together, as shown in the <a class="reference internal" href="notes/amp_examples.html#amp-examples"><span class="std std-ref">CUDA Automatic Mixed Precision examples</span></a>
and <a class="reference external" href="https://pytorch.org/tutorials/recipes/recipes/amp_recipe.html">CUDA Automatic Mixed Precision recipe</a>.
However, <a class="reference internal" href="#torch.autocast" title="torch.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.autocast</span></code></a> and <a class="reference internal" href="#torch.cuda.amp.GradScaler" title="torch.cuda.amp.GradScaler"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.cuda.amp.GradScaler</span></code></a> are modular, and may be used separately if desired.
As shown in the CPU example section of <a class="reference internal" href="#torch.autocast" title="torch.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.autocast</span></code></a>, “automatic mixed precision training/inference” on CPU with
datatype of <code class="docutils literal notranslate"><span class="pre">torch.bfloat16</span></code> only uses <a class="reference internal" href="#torch.autocast" title="torch.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.autocast</span></code></a>.</p>
<p>For CUDA and CPU, APIs are also provided separately:</p>
<ul class="simple">
<li><p><code class="docutils literal notranslate"><span class="pre">torch.autocast("cuda",</span> <span class="pre">args...)</span></code> is equivalent to <code class="docutils literal notranslate"><span class="pre">torch.cuda.amp.autocast(args...)</span></code>.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">torch.autocast("cpu",</span> <span class="pre">args...)</span></code> is equivalent to <code class="docutils literal notranslate"><span class="pre">torch.cpu.amp.autocast(args...)</span></code>. For CPU, only lower precision floating point datatype of <code class="docutils literal notranslate"><span class="pre">torch.bfloat16</span></code> is supported for now.</p></li>
</ul>
<nav class="contents local" id="contents">
<ul class="simple">
<li><p><a class="reference internal" href="#autocasting" id="id4">Autocasting</a></p></li>
<li><p><a class="reference internal" href="#gradient-scaling" id="id5">Gradient Scaling</a></p></li>
<li><p><a class="reference internal" href="#autocast-op-reference" id="id6">Autocast Op Reference</a></p>
<ul>
<li><p><a class="reference internal" href="#op-eligibility" id="id7">Op Eligibility</a></p></li>
<li><p><a class="reference internal" href="#cuda-op-specific-behavior" id="id8">CUDA Op-Specific Behavior</a></p>
<ul>
<li><p><a class="reference internal" href="#cuda-ops-that-can-autocast-to-float16" id="id9">CUDA Ops that can autocast to <code class="docutils literal notranslate"><span class="pre">float16</span></code></a></p></li>
<li><p><a class="reference internal" href="#cuda-ops-that-can-autocast-to-float32" id="id10">CUDA Ops that can autocast to <code class="docutils literal notranslate"><span class="pre">float32</span></code></a></p></li>
<li><p><a class="reference internal" href="#cuda-ops-that-promote-to-the-widest-input-type" id="id11">CUDA Ops that promote to the widest input type</a></p></li>
<li><p><a class="reference internal" href="#prefer-binary-cross-entropy-with-logits-over-binary-cross-entropy" id="id12">Prefer <code class="docutils literal notranslate"><span class="pre">binary_cross_entropy_with_logits</span></code> over <code class="docutils literal notranslate"><span class="pre">binary_cross_entropy</span></code></a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#cpu-op-specific-behavior" id="id13">CPU Op-Specific Behavior</a></p>
<ul>
<li><p><a class="reference internal" href="#cpu-ops-that-can-autocast-to-bfloat16" id="id14">CPU Ops that can autocast to <code class="docutils literal notranslate"><span class="pre">bfloat16</span></code></a></p></li>
<li><p><a class="reference internal" href="#cpu-ops-that-can-autocast-to-float32" id="id15">CPU Ops that can autocast to <code class="docutils literal notranslate"><span class="pre">float32</span></code></a></p></li>
<li><p><a class="reference internal" href="#cpu-ops-that-promote-to-the-widest-input-type" id="id16">CPU Ops that promote to the widest input type</a></p></li>
</ul>
</li>
</ul>
</li>
</ul>
</nav>
<section id="autocasting">
<span id="id1"></span><h2><a class="toc-backref" href="#id4" role="doc-backlink">Autocasting</a><a class="headerlink" href="#autocasting" title="Permalink to this heading">¶</a></h2>
<dl class="py class">
<dt class="sig sig-object py" id="torch.autocast">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.</span></span><span class="sig-name descname"><span class="pre">autocast</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">device_type</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">dtype</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">enabled</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">cache_enabled</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/amp/autocast_mode.html#autocast"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.autocast" title="Permalink to this definition">¶</a></dt>
<dd><p>Instances of <a class="reference internal" href="#torch.autocast" title="torch.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">autocast</span></code></a> serve as context managers or decorators that
allow regions of your script to run in mixed precision.</p>
<p>In these regions, ops run in an op-specific dtype chosen by autocast
to improve performance while maintaining accuracy.
See the <a class="reference internal" href="#autocast-op-reference"><span class="std std-ref">Autocast Op Reference</span></a> for details.</p>
<p>When entering an autocast-enabled region, Tensors may be any type.
You should not call <code class="docutils literal notranslate"><span class="pre">half()</span></code> or <code class="docutils literal notranslate"><span class="pre">bfloat16()</span></code> on your model(s) or inputs when using autocasting.</p>
<p><a class="reference internal" href="#torch.autocast" title="torch.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">autocast</span></code></a> should wrap only the forward pass(es) of your network, including the loss
computation(s). Backward passes under autocast are not recommended.
Backward ops run in the same type that autocast used for corresponding forward ops.</p>
<p>Example for CUDA Devices:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Creates model and optimizer in default precision</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">Net</span><span class="p">()</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
<span class="n">optimizer</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="o">...</span><span class="p">)</span>
<span class="k">for</span> <span class="nb">input</span><span class="p">,</span> <span class="n">target</span> <span class="ow">in</span> <span class="n">data</span><span class="p">:</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="c1"># Enables autocasting for the forward pass (model + loss)</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">model</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="n">loss</span> <span class="o">=</span> <span class="n">loss_fn</span><span class="p">(</span><span class="n">output</span><span class="p">,</span> <span class="n">target</span><span class="p">)</span>
<span class="c1"># Exits the context manager before backward()</span>
<span class="n">loss</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<p>See the <a class="reference internal" href="notes/amp_examples.html#amp-examples"><span class="std std-ref">CUDA Automatic Mixed Precision examples</span></a> for usage (along with gradient scaling)
in more complex scenarios (e.g., gradient penalty, multiple models/losses, custom autograd functions).</p>
<p><a class="reference internal" href="#torch.autocast" title="torch.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">autocast</span></code></a> can also be used as a decorator, e.g., on the <code class="docutils literal notranslate"><span class="pre">forward</span></code> method of your model:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">AutocastModel</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="o">...</span>
<span class="nd">@autocast</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="nb">input</span><span class="p">):</span>
<span class="o">...</span>
</pre></div>
</div>
<p>Floating-point Tensors produced in an autocast-enabled region may be <code class="docutils literal notranslate"><span class="pre">float16</span></code>.
After returning to an autocast-disabled region, using them with floating-point
Tensors of different dtypes may cause type mismatch errors. If so, cast the Tensor(s)
produced in the autocast region back to <code class="docutils literal notranslate"><span class="pre">float32</span></code> (or other dtype if desired).
If a Tensor from the autocast region is already <code class="docutils literal notranslate"><span class="pre">float32</span></code>, the cast is a no-op,
and incurs no additional overhead.
CUDA Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Creates some tensors in default dtype (here assumed to be float32)</span>
<span class="n">a_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">((</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="n">b_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">((</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="n">c_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">((</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="n">d_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">((</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="c1"># torch.mm is on autocast's list of ops that should run in float16.</span>
<span class="c1"># Inputs are float32, but the op runs in float16 and produces float16 output.</span>
<span class="c1"># No manual casts are required.</span>
<span class="n">e_float16</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">mm</span><span class="p">(</span><span class="n">a_float32</span><span class="p">,</span> <span class="n">b_float32</span><span class="p">)</span>
<span class="c1"># Also handles mixed input types</span>
<span class="n">f_float16</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">mm</span><span class="p">(</span><span class="n">d_float32</span><span class="p">,</span> <span class="n">e_float16</span><span class="p">)</span>
<span class="c1"># After exiting autocast, calls f_float16.float() to use with d_float32</span>
<span class="n">g_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">mm</span><span class="p">(</span><span class="n">d_float32</span><span class="p">,</span> <span class="n">f_float16</span><span class="o">.</span><span class="n">float</span><span class="p">())</span>
</pre></div>
</div>
<p>CPU Training Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Creates model and optimizer in default precision</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">Net</span><span class="p">()</span>
<span class="n">optimizer</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="o">...</span><span class="p">)</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="n">epochs</span><span class="p">:</span>
<span class="k">for</span> <span class="nb">input</span><span class="p">,</span> <span class="n">target</span> <span class="ow">in</span> <span class="n">data</span><span class="p">:</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="c1"># Runs the forward pass with autocasting.</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">autocast</span><span class="p">(</span><span class="n">device_type</span><span class="o">=</span><span class="s2">"cpu"</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">bfloat16</span><span class="p">):</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">model</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="n">loss</span> <span class="o">=</span> <span class="n">loss_fn</span><span class="p">(</span><span class="n">output</span><span class="p">,</span> <span class="n">target</span><span class="p">)</span>
<span class="n">loss</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<p>CPU Inference Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Creates model in default precision</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">Net</span><span class="p">()</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">autocast</span><span class="p">(</span><span class="n">device_type</span><span class="o">=</span><span class="s2">"cpu"</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">bfloat16</span><span class="p">):</span>
<span class="k">for</span> <span class="nb">input</span> <span class="ow">in</span> <span class="n">data</span><span class="p">:</span>
<span class="c1"># Runs the forward pass with autocasting.</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">model</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
</pre></div>
</div>
<p>CPU Inference Example with Jit Trace:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">TestModel</span><span class="p">(</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="n">input_size</span><span class="p">,</span> <span class="n">num_classes</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">TestModel</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">fc1</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="n">input_size</span><span class="p">,</span> <span class="n">num_classes</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">fc1</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">input_size</span> <span class="o">=</span> <span class="mi">2</span>
<span class="n">num_classes</span> <span class="o">=</span> <span class="mi">2</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">TestModel</span><span class="p">(</span><span class="n">input_size</span><span class="p">,</span> <span class="n">num_classes</span><span class="p">)</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="c1"># For now, we suggest to disable the Jit Autocast Pass,</span>
<span class="c1"># As the issue: https://github.com/pytorch/pytorch/issues/75956</span>
<span class="n">torch</span><span class="o">.</span><span class="n">_C</span><span class="o">.</span><span class="n">_jit_set_autocast_mode</span><span class="p">(</span><span class="kc">False</span><span class="p">)</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">cpu</span><span class="o">.</span><span class="n">amp</span><span class="o">.</span><span class="n">autocast</span><span class="p">(</span><span class="n">cache_enabled</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">trace</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">input_size</span><span class="p">))</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">freeze</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="c1"># Models Run</span>
<span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">3</span><span class="p">):</span>
<span class="n">model</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">input_size</span><span class="p">))</span>
</pre></div>
</div>
<p>Type mismatch errors <em>in</em> an autocast-enabled region are a bug; if this is what you observe,
please file an issue.</p>
<p><code class="docutils literal notranslate"><span class="pre">autocast(enabled=False)</span></code> subregions can be nested in autocast-enabled regions.
Locally disabling autocast can be useful, for example, if you want to force a subregion
to run in a particular <code class="docutils literal notranslate"><span class="pre">dtype</span></code>. Disabling autocast gives you explicit control over
the execution type. In the subregion, inputs from the surrounding region
should be cast to <code class="docutils literal notranslate"><span class="pre">dtype</span></code> before use:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Creates some tensors in default dtype (here assumed to be float32)</span>
<span class="n">a_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">((</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="n">b_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">((</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="n">c_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">((</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="n">d_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">((</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="n">e_float16</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">mm</span><span class="p">(</span><span class="n">a_float32</span><span class="p">,</span> <span class="n">b_float32</span><span class="p">)</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">(</span><span class="n">enabled</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
<span class="c1"># Calls e_float16.float() to ensure float32 execution</span>
<span class="c1"># (necessary because e_float16 was created in an autocasted region)</span>
<span class="n">f_float32</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">mm</span><span class="p">(</span><span class="n">c_float32</span><span class="p">,</span> <span class="n">e_float16</span><span class="o">.</span><span class="n">float</span><span class="p">())</span>
<span class="c1"># No manual casts are required when re-entering the autocast-enabled region.</span>
<span class="c1"># torch.mm again runs in float16 and produces float16 output, regardless of input types.</span>
<span class="n">g_float16</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">mm</span><span class="p">(</span><span class="n">d_float32</span><span class="p">,</span> <span class="n">f_float32</span><span class="p">)</span>
</pre></div>
</div>
<p>The autocast state is thread-local. If you want it enabled in a new thread, the context manager or decorator
must be invoked in that thread. This affects <a class="reference internal" href="generated/torch.nn.DataParallel.html#torch.nn.DataParallel" title="torch.nn.DataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.DataParallel</span></code></a> and
<a class="reference internal" href="generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel" title="torch.nn.parallel.DistributedDataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.parallel.DistributedDataParallel</span></code></a> when used with more than one GPU per process
(see <a class="reference internal" href="notes/amp_examples.html#amp-multigpu"><span class="std std-ref">Working with Multiple GPUs</span></a>).</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>device_type</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#str" title="(in Python v3.10)"><em>str</em></a><em>, </em><em>required</em>) – Whether to use ‘cuda’ or ‘cpu’ device</p></li>
<li><p><strong>enabled</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.10)"><em>bool</em></a><em>, </em><em>optional</em>) – Whether autocasting should be enabled in the region.
Default: <code class="docutils literal notranslate"><span class="pre">True</span></code></p></li>
<li><p><strong>dtype</strong> (<em>torch_dtype</em><em>, </em><em>optional</em>) – Whether to use torch.float16 or torch.bfloat16.</p></li>
<li><p><strong>cache_enabled</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.10)"><em>bool</em></a><em>, </em><em>optional</em>) – Whether the weight cache inside autocast should be enabled.
Default: <code class="docutils literal notranslate"><span class="pre">True</span></code></p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.cuda.amp.autocast">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.cuda.amp.</span></span><span class="sig-name descname"><span class="pre">autocast</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">enabled</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">dtype</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">torch.float16</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">cache_enabled</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/autocast_mode.html#autocast"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.autocast" title="Permalink to this definition">¶</a></dt>
<dd><p>See <a class="reference internal" href="#torch.autocast" title="torch.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.autocast</span></code></a>.
<code class="docutils literal notranslate"><span class="pre">torch.cuda.amp.autocast(args...)</span></code> is equivalent to <code class="docutils literal notranslate"><span class="pre">torch.autocast("cuda",</span> <span class="pre">args...)</span></code></p>
<dl class="field-list simple">
</dl>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.cuda.amp.custom_fwd">
<span class="sig-prename descclassname"><span class="pre">torch.cuda.amp.</span></span><span class="sig-name descname"><span class="pre">custom_fwd</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">fwd</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">cast_inputs</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/autocast_mode.html#custom_fwd"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.custom_fwd" title="Permalink to this definition">¶</a></dt>
<dd><p>Helper decorator for <code class="docutils literal notranslate"><span class="pre">forward</span></code> methods of custom autograd functions (subclasses of
<a class="reference internal" href="autograd.html#torch.autograd.Function" title="torch.autograd.Function"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.autograd.Function</span></code></a>). See the <a class="reference internal" href="notes/amp_examples.html#amp-custom-examples"><span class="std std-ref">example page</span></a> for more detail.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>cast_inputs</strong> (<a class="reference internal" href="tensor_attributes.html#torch.dtype" title="torch.dtype"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.dtype</span></code></a> or None, optional, default=None) – If not <code class="docutils literal notranslate"><span class="pre">None</span></code>,
when <code class="docutils literal notranslate"><span class="pre">forward</span></code> runs in an autocast-enabled region, casts incoming
floating-point CUDA Tensors to the target dtype (non-floating-point Tensors are not affected),
then executes <code class="docutils literal notranslate"><span class="pre">forward</span></code> with autocast disabled.
If <code class="docutils literal notranslate"><span class="pre">None</span></code>, <code class="docutils literal notranslate"><span class="pre">forward</span></code>’s internal ops execute with the current autocast state.</p>
</dd>
</dl>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>If the decorated <code class="docutils literal notranslate"><span class="pre">forward</span></code> is called outside an autocast-enabled region,
<a class="reference internal" href="#torch.cuda.amp.custom_fwd" title="torch.cuda.amp.custom_fwd"><code class="xref py py-func docutils literal notranslate"><span class="pre">custom_fwd</span></code></a> is a no-op and <code class="docutils literal notranslate"><span class="pre">cast_inputs</span></code> has no effect.</p>
</div>
</dd></dl>
<dl class="py function">
<dt class="sig sig-object py" id="torch.cuda.amp.custom_bwd">
<span class="sig-prename descclassname"><span class="pre">torch.cuda.amp.</span></span><span class="sig-name descname"><span class="pre">custom_bwd</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">bwd</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/autocast_mode.html#custom_bwd"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.custom_bwd" title="Permalink to this definition">¶</a></dt>
<dd><p>Helper decorator for backward methods of custom autograd functions (subclasses of
<a class="reference internal" href="autograd.html#torch.autograd.Function" title="torch.autograd.Function"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.autograd.Function</span></code></a>).
Ensures that <code class="docutils literal notranslate"><span class="pre">backward</span></code> executes with the same autocast state as <code class="docutils literal notranslate"><span class="pre">forward</span></code>.
See the <a class="reference internal" href="notes/amp_examples.html#amp-custom-examples"><span class="std std-ref">example page</span></a> for more detail.</p>
</dd></dl>
<dl class="py class">
<dt class="sig sig-object py" id="torch.cpu.amp.autocast">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.cpu.amp.</span></span><span class="sig-name descname"><span class="pre">autocast</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">enabled</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">dtype</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">torch.bfloat16</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">cache_enabled</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cpu/amp/autocast_mode.html#autocast"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cpu.amp.autocast" title="Permalink to this definition">¶</a></dt>
<dd><p>See <a class="reference internal" href="#torch.autocast" title="torch.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.autocast</span></code></a>.
<code class="docutils literal notranslate"><span class="pre">torch.cpu.amp.autocast(args...)</span></code> is equivalent to <code class="docutils literal notranslate"><span class="pre">torch.autocast("cpu",</span> <span class="pre">args...)</span></code></p>
<dl class="field-list simple">
</dl>
</dd></dl>
</section>
<section id="gradient-scaling">
<span id="id2"></span><h2><a class="toc-backref" href="#id5" role="doc-backlink">Gradient Scaling</a><a class="headerlink" href="#gradient-scaling" title="Permalink to this heading">¶</a></h2>
<p>If the forward pass for a particular op has <code class="docutils literal notranslate"><span class="pre">float16</span></code> inputs, the backward pass for
that op will produce <code class="docutils literal notranslate"><span class="pre">float16</span></code> gradients.
Gradient values with small magnitudes may not be representable in <code class="docutils literal notranslate"><span class="pre">float16</span></code>.
These values will flush to zero (“underflow”), so the update for the corresponding parameters will be lost.</p>
<p>To prevent underflow, “gradient scaling” multiplies the network’s loss(es) by a scale factor and
invokes a backward pass on the scaled loss(es). Gradients flowing backward through the network are
then scaled by the same factor. In other words, gradient values have a larger magnitude,
so they don’t flush to zero.</p>
<p>Each parameter’s gradient (<code class="docutils literal notranslate"><span class="pre">.grad</span></code> attribute) should be unscaled before the optimizer
updates the parameters, so the scale factor does not interfere with the learning rate.</p>
<dl class="py class">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.cuda.amp.</span></span><span class="sig-name descname"><span class="pre">GradScaler</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">init_scale</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">65536.0</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">growth_factor</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">2.0</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">backoff_factor</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0.5</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">growth_interval</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">2000</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">enabled</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">True</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler" title="Permalink to this definition">¶</a></dt>
<dd><dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.get_backoff_factor">
<span class="sig-name descname"><span class="pre">get_backoff_factor</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.get_backoff_factor"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.get_backoff_factor" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns a Python float containing the scale backoff factor.</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.get_growth_factor">
<span class="sig-name descname"><span class="pre">get_growth_factor</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.get_growth_factor"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.get_growth_factor" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns a Python float containing the scale growth factor.</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.get_growth_interval">
<span class="sig-name descname"><span class="pre">get_growth_interval</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.get_growth_interval"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.get_growth_interval" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns a Python int containing the growth interval.</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.get_scale">
<span class="sig-name descname"><span class="pre">get_scale</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.get_scale"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.get_scale" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns a Python float containing the current scale, or 1.0 if scaling is disabled.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p><a class="reference internal" href="#torch.cuda.amp.GradScaler.get_scale" title="torch.cuda.amp.GradScaler.get_scale"><code class="xref py py-meth docutils literal notranslate"><span class="pre">get_scale()</span></code></a> incurs a CPU-GPU sync.</p>
</div>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.is_enabled">
<span class="sig-name descname"><span class="pre">is_enabled</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.is_enabled"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.is_enabled" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns a bool indicating whether this instance is enabled.</p>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.load_state_dict">
<span class="sig-name descname"><span class="pre">load_state_dict</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">state_dict</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.load_state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.load_state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Loads the scaler state. If this instance is disabled, <a class="reference internal" href="#torch.cuda.amp.GradScaler.load_state_dict" title="torch.cuda.amp.GradScaler.load_state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">load_state_dict()</span></code></a> is a no-op.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>state_dict</strong> (<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.10)"><em>dict</em></a>) – scaler state. Should be an object returned from a call to <a class="reference internal" href="#torch.cuda.amp.GradScaler.state_dict" title="torch.cuda.amp.GradScaler.state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">state_dict()</span></code></a>.</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.scale">
<span class="sig-name descname"><span class="pre">scale</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">outputs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.scale"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.scale" title="Permalink to this definition">¶</a></dt>
<dd><p>Multiplies (‘scales’) a tensor or list of tensors by the scale factor.</p>
<p>Returns scaled outputs. If this instance of <a class="reference internal" href="#torch.cuda.amp.GradScaler" title="torch.cuda.amp.GradScaler"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradScaler</span></code></a> is not enabled, outputs are returned
unmodified.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>outputs</strong> (<a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>Tensor</em></a><em> or </em><em>iterable of Tensors</em>) – Outputs to scale.</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.set_backoff_factor">
<span class="sig-name descname"><span class="pre">set_backoff_factor</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">new_factor</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.set_backoff_factor"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.set_backoff_factor" title="Permalink to this definition">¶</a></dt>
<dd><dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>new_scale</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#float" title="(in Python v3.10)"><em>float</em></a>) – Value to use as the new scale backoff factor.</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.set_growth_factor">
<span class="sig-name descname"><span class="pre">set_growth_factor</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">new_factor</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.set_growth_factor"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.set_growth_factor" title="Permalink to this definition">¶</a></dt>
<dd><dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>new_scale</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#float" title="(in Python v3.10)"><em>float</em></a>) – Value to use as the new scale growth factor.</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.set_growth_interval">
<span class="sig-name descname"><span class="pre">set_growth_interval</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">new_interval</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.set_growth_interval"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.set_growth_interval" title="Permalink to this definition">¶</a></dt>
<dd><dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>new_interval</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a>) – Value to use as the new growth interval.</p>
</dd>
</dl>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.state_dict">
<span class="sig-name descname"><span class="pre">state_dict</span></span><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.state_dict"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.state_dict" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns the state of the scaler as a <a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.10)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a>. It contains five entries:</p>
<ul class="simple">
<li><p><code class="docutils literal notranslate"><span class="pre">"scale"</span></code> - a Python float containing the current scale</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">"growth_factor"</span></code> - a Python float containing the current growth factor</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">"backoff_factor"</span></code> - a Python float containing the current backoff factor</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">"growth_interval"</span></code> - a Python int containing the current growth interval</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">"_growth_tracker"</span></code> - a Python int containing the number of recent consecutive unskipped steps.</p></li>
</ul>
<p>If this instance is not enabled, returns an empty dict.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>If you wish to checkpoint the scaler’s state after a particular iteration, <a class="reference internal" href="#torch.cuda.amp.GradScaler.state_dict" title="torch.cuda.amp.GradScaler.state_dict"><code class="xref py py-meth docutils literal notranslate"><span class="pre">state_dict()</span></code></a>
should be called after <a class="reference internal" href="#torch.cuda.amp.GradScaler.update" title="torch.cuda.amp.GradScaler.update"><code class="xref py py-meth docutils literal notranslate"><span class="pre">update()</span></code></a>.</p>
</div>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.step">
<span class="sig-name descname"><span class="pre">step</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">optimizer</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwargs</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.step"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.step" title="Permalink to this definition">¶</a></dt>
<dd><p><a class="reference internal" href="#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code></a> carries out the following two operations:</p>
<ol class="arabic simple">
<li><p>Internally invokes <code class="docutils literal notranslate"><span class="pre">unscale_(optimizer)</span></code> (unless <a class="reference internal" href="#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_()</span></code></a> was explicitly called for <code class="docutils literal notranslate"><span class="pre">optimizer</span></code>
earlier in the iteration). As part of the <a class="reference internal" href="#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_()</span></code></a>, gradients are checked for infs/NaNs.</p></li>
<li><p>If no inf/NaN gradients are found, invokes <code class="docutils literal notranslate"><span class="pre">optimizer.step()</span></code> using the unscaled
gradients. Otherwise, <code class="docutils literal notranslate"><span class="pre">optimizer.step()</span></code> is skipped to avoid corrupting the params.</p></li>
</ol>
<p><code class="docutils literal notranslate"><span class="pre">*args</span></code> and <code class="docutils literal notranslate"><span class="pre">**kwargs</span></code> are forwarded to <code class="docutils literal notranslate"><span class="pre">optimizer.step()</span></code>.</p>
<p>Returns the return value of <code class="docutils literal notranslate"><span class="pre">optimizer.step(*args,</span> <span class="pre">**kwargs)</span></code>.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>optimizer</strong> (<a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.Optimizer"><em>torch.optim.Optimizer</em></a>) – Optimizer that applies the gradients.</p></li>
<li><p><strong>args</strong> – Any arguments.</p></li>
<li><p><strong>kwargs</strong> – Any keyword arguments.</p></li>
</ul>
</dd>
</dl>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Closure use is not currently supported.</p>
</div>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.unscale_">
<span class="sig-name descname"><span class="pre">unscale_</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">optimizer</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.unscale_"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.unscale_" title="Permalink to this definition">¶</a></dt>
<dd><p>Divides (“unscales”) the optimizer’s gradient tensors by the scale factor.</p>
<p><a class="reference internal" href="#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_()</span></code></a> is optional, serving cases where you need to
<a class="reference internal" href="notes/amp_examples.html#working-with-unscaled-gradients"><span class="std std-ref">modify or inspect gradients</span></a>
between the backward pass(es) and <a class="reference internal" href="#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code></a>.
If <a class="reference internal" href="#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_()</span></code></a> is not called explicitly, gradients will be unscaled automatically during <a class="reference internal" href="#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code></a>.</p>
<p>Simple example, using <a class="reference internal" href="#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_()</span></code></a> to enable clipping of unscaled gradients:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="o">...</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">scale</span><span class="p">(</span><span class="n">loss</span><span class="p">)</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">unscale_</span><span class="p">(</span><span class="n">optimizer</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">utils</span><span class="o">.</span><span class="n">clip_grad_norm_</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">max_norm</span><span class="p">)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">optimizer</span><span class="p">)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">update</span><span class="p">()</span>
</pre></div>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>optimizer</strong> (<a class="reference internal" href="optim.html#torch.optim.Optimizer" title="torch.optim.Optimizer"><em>torch.optim.Optimizer</em></a>) – Optimizer that owns the gradients to be unscaled.</p>
</dd>
</dl>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p><a class="reference internal" href="#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_()</span></code></a> does not incur a CPU-GPU sync.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p><a class="reference internal" href="#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_()</span></code></a> should only be called once per optimizer per <a class="reference internal" href="#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code></a> call,
and only after all gradients for that optimizer’s assigned parameters have been accumulated.
Calling <a class="reference internal" href="#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_()</span></code></a> twice for a given optimizer between each <a class="reference internal" href="#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code></a> triggers a RuntimeError.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p><a class="reference internal" href="#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_()</span></code></a> may unscale sparse gradients out of place, replacing the <code class="docutils literal notranslate"><span class="pre">.grad</span></code> attribute.</p>
</div>
</dd></dl>
<dl class="py method">
<dt class="sig sig-object py" id="torch.cuda.amp.GradScaler.update">
<span class="sig-name descname"><span class="pre">update</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">new_scale</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/cuda/amp/grad_scaler.html#GradScaler.update"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.cuda.amp.GradScaler.update" title="Permalink to this definition">¶</a></dt>
<dd><p>Updates the scale factor.</p>
<p>If any optimizer steps were skipped the scale is multiplied by <code class="docutils literal notranslate"><span class="pre">backoff_factor</span></code>
to reduce it. If <code class="docutils literal notranslate"><span class="pre">growth_interval</span></code> unskipped iterations occurred consecutively,
the scale is multiplied by <code class="docutils literal notranslate"><span class="pre">growth_factor</span></code> to increase it.</p>
<p>Passing <code class="docutils literal notranslate"><span class="pre">new_scale</span></code> sets the new scale value manually. (<code class="docutils literal notranslate"><span class="pre">new_scale</span></code> is not
used directly, it’s used to fill GradScaler’s internal scale tensor. So if
<code class="docutils literal notranslate"><span class="pre">new_scale</span></code> was a tensor, later in-place changes to that tensor will not further
affect the scale GradScaler uses internally.)</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><p><strong>new_scale</strong> (float or <code class="xref py py-class docutils literal notranslate"><span class="pre">torch.cuda.FloatTensor</span></code>, optional, default=None) – New scale factor.</p>
</dd>
</dl>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p><a class="reference internal" href="#torch.cuda.amp.GradScaler.update" title="torch.cuda.amp.GradScaler.update"><code class="xref py py-meth docutils literal notranslate"><span class="pre">update()</span></code></a> should only be called at the end of the iteration, after <code class="docutils literal notranslate"><span class="pre">scaler.step(optimizer)</span></code> has
been invoked for all optimizers used this iteration.</p>
</div>
</dd></dl>
</dd></dl>
</section>
<section id="autocast-op-reference">
<span id="id3"></span><h2><a class="toc-backref" href="#id6" role="doc-backlink">Autocast Op Reference</a><a class="headerlink" href="#autocast-op-reference" title="Permalink to this heading">¶</a></h2>
<section id="op-eligibility">
<span id="autocast-eligibility"></span><h3><a class="toc-backref" href="#id7" role="doc-backlink">Op Eligibility</a><a class="headerlink" href="#op-eligibility" title="Permalink to this heading">¶</a></h3>
<p>Ops that run in <code class="docutils literal notranslate"><span class="pre">float64</span></code> or non-floating-point dtypes are not eligible, and will
run in these types whether or not autocast is enabled.</p>
<p>Only out-of-place ops and Tensor methods are eligible.
In-place variants and calls that explicitly supply an <code class="docutils literal notranslate"><span class="pre">out=...</span></code> Tensor
are allowed in autocast-enabled regions, but won’t go through autocasting.
For example, in an autocast-enabled region <code class="docutils literal notranslate"><span class="pre">a.addmm(b,</span> <span class="pre">c)</span></code> can autocast,
but <code class="docutils literal notranslate"><span class="pre">a.addmm_(b,</span> <span class="pre">c)</span></code> and <code class="docutils literal notranslate"><span class="pre">a.addmm(b,</span> <span class="pre">c,</span> <span class="pre">out=d)</span></code> cannot.
For best performance and stability, prefer out-of-place ops in autocast-enabled
regions.</p>
<p>Ops called with an explicit <code class="docutils literal notranslate"><span class="pre">dtype=...</span></code> argument are not eligible,
and will produce output that respects the <code class="docutils literal notranslate"><span class="pre">dtype</span></code> argument.</p>
</section>
<section id="cuda-op-specific-behavior">
<span id="autocast-cuda-op-reference"></span><h3><a class="toc-backref" href="#id8" role="doc-backlink">CUDA Op-Specific Behavior</a><a class="headerlink" href="#cuda-op-specific-behavior" title="Permalink to this heading">¶</a></h3>
<p>The following lists describe the behavior of eligible ops in autocast-enabled regions.
These ops always go through autocasting whether they are invoked as part of a <a class="reference internal" href="generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.Module</span></code></a>,
as a function, or as a <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a> method. If functions are exposed in multiple namespaces,
they go through autocasting regardless of the namespace.</p>
<p>Ops not listed below do not go through autocasting. They run in the type
defined by their inputs. However, autocasting may still change the type
in which unlisted ops run if they’re downstream from autocasted ops.</p>
<p>If an op is unlisted, we assume it’s numerically stable in <code class="docutils literal notranslate"><span class="pre">float16</span></code>.
If you believe an unlisted op is numerically unstable in <code class="docutils literal notranslate"><span class="pre">float16</span></code>,
please file an issue.</p>
<section id="cuda-ops-that-can-autocast-to-float16">
<h4><a class="toc-backref" href="#id9" role="doc-backlink">CUDA Ops that can autocast to <code class="docutils literal notranslate"><span class="pre">float16</span></code></a><a class="headerlink" href="#cuda-ops-that-can-autocast-to-float16" title="Permalink to this heading">¶</a></h4>
<p><code class="docutils literal notranslate"><span class="pre">__matmul__</span></code>,
<code class="docutils literal notranslate"><span class="pre">addbmm</span></code>,
<code class="docutils literal notranslate"><span class="pre">addmm</span></code>,
<code class="docutils literal notranslate"><span class="pre">addmv</span></code>,
<code class="docutils literal notranslate"><span class="pre">addr</span></code>,
<code class="docutils literal notranslate"><span class="pre">baddbmm</span></code>,
<code class="docutils literal notranslate"><span class="pre">bmm</span></code>,
<code class="docutils literal notranslate"><span class="pre">chain_matmul</span></code>,
<code class="docutils literal notranslate"><span class="pre">multi_dot</span></code>,
<code class="docutils literal notranslate"><span class="pre">conv1d</span></code>,
<code class="docutils literal notranslate"><span class="pre">conv2d</span></code>,
<code class="docutils literal notranslate"><span class="pre">conv3d</span></code>,
<code class="docutils literal notranslate"><span class="pre">conv_transpose1d</span></code>,
<code class="docutils literal notranslate"><span class="pre">conv_transpose2d</span></code>,
<code class="docutils literal notranslate"><span class="pre">conv_transpose3d</span></code>,
<code class="docutils literal notranslate"><span class="pre">GRUCell</span></code>,
<code class="docutils literal notranslate"><span class="pre">linear</span></code>,
<code class="docutils literal notranslate"><span class="pre">LSTMCell</span></code>,
<code class="docutils literal notranslate"><span class="pre">matmul</span></code>,
<code class="docutils literal notranslate"><span class="pre">mm</span></code>,
<code class="docutils literal notranslate"><span class="pre">mv</span></code>,
<code class="docutils literal notranslate"><span class="pre">prelu</span></code>,
<code class="docutils literal notranslate"><span class="pre">RNNCell</span></code></p>
</section>
<section id="cuda-ops-that-can-autocast-to-float32">
<h4><a class="toc-backref" href="#id10" role="doc-backlink">CUDA Ops that can autocast to <code class="docutils literal notranslate"><span class="pre">float32</span></code></a><a class="headerlink" href="#cuda-ops-that-can-autocast-to-float32" title="Permalink to this heading">¶</a></h4>
<p><code class="docutils literal notranslate"><span class="pre">__pow__</span></code>,
<code class="docutils literal notranslate"><span class="pre">__rdiv__</span></code>,
<code class="docutils literal notranslate"><span class="pre">__rpow__</span></code>,
<code class="docutils literal notranslate"><span class="pre">__rtruediv__</span></code>,
<code class="docutils literal notranslate"><span class="pre">acos</span></code>,
<code class="docutils literal notranslate"><span class="pre">asin</span></code>,
<code class="docutils literal notranslate"><span class="pre">binary_cross_entropy_with_logits</span></code>,
<code class="docutils literal notranslate"><span class="pre">cosh</span></code>,
<code class="docutils literal notranslate"><span class="pre">cosine_embedding_loss</span></code>,
<code class="docutils literal notranslate"><span class="pre">cdist</span></code>,
<code class="docutils literal notranslate"><span class="pre">cosine_similarity</span></code>,
<code class="docutils literal notranslate"><span class="pre">cross_entropy</span></code>,
<code class="docutils literal notranslate"><span class="pre">cumprod</span></code>,
<code class="docutils literal notranslate"><span class="pre">cumsum</span></code>,
<code class="docutils literal notranslate"><span class="pre">dist</span></code>,
<code class="docutils literal notranslate"><span class="pre">erfinv</span></code>,
<code class="docutils literal notranslate"><span class="pre">exp</span></code>,
<code class="docutils literal notranslate"><span class="pre">expm1</span></code>,
<code class="docutils literal notranslate"><span class="pre">group_norm</span></code>,
<code class="docutils literal notranslate"><span class="pre">hinge_embedding_loss</span></code>,
<code class="docutils literal notranslate"><span class="pre">kl_div</span></code>,
<code class="docutils literal notranslate"><span class="pre">l1_loss</span></code>,
<code class="docutils literal notranslate"><span class="pre">layer_norm</span></code>,
<code class="docutils literal notranslate"><span class="pre">log</span></code>,
<code class="docutils literal notranslate"><span class="pre">log_softmax</span></code>,
<code class="docutils literal notranslate"><span class="pre">log10</span></code>,
<code class="docutils literal notranslate"><span class="pre">log1p</span></code>,
<code class="docutils literal notranslate"><span class="pre">log2</span></code>,
<code class="docutils literal notranslate"><span class="pre">margin_ranking_loss</span></code>,
<code class="docutils literal notranslate"><span class="pre">mse_loss</span></code>,
<code class="docutils literal notranslate"><span class="pre">multilabel_margin_loss</span></code>,
<code class="docutils literal notranslate"><span class="pre">multi_margin_loss</span></code>,
<code class="docutils literal notranslate"><span class="pre">nll_loss</span></code>,
<code class="docutils literal notranslate"><span class="pre">norm</span></code>,
<code class="docutils literal notranslate"><span class="pre">normalize</span></code>,
<code class="docutils literal notranslate"><span class="pre">pdist</span></code>,
<code class="docutils literal notranslate"><span class="pre">poisson_nll_loss</span></code>,
<code class="docutils literal notranslate"><span class="pre">pow</span></code>,
<code class="docutils literal notranslate"><span class="pre">prod</span></code>,
<code class="docutils literal notranslate"><span class="pre">reciprocal</span></code>,