-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathfx.html
4461 lines (4077 loc) · 422 KB
/
fx.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta name="robots" content="noindex">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>torch.fx — PyTorch 2.5 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/fx.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="torch.fx.experimental" href="fx.experimental.html" />
<link rel="prev" title="torch.futures" href="futures.html" />
<!-- Google Tag Manager -->
<script>(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':
new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],
j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src=
'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);
})(window,document,'script','dataLayer','GTM-T8XT4PS');</script>
<!-- End Google Tag Manager -->
<script src="_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Learn
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/get-started">
<span class=dropdown-title>Get Started</span>
<p>Run PyTorch locally or get started quickly with one of the supported cloud platforms</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials">
<span class="dropdown-title">Tutorials</span>
<p>Whats new in PyTorch tutorials</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/basics/intro.html">
<span class="dropdown-title">Learn the Basics</span>
<p>Familiarize yourself with PyTorch concepts and modules</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/recipes/recipes_index.html">
<span class="dropdown-title">PyTorch Recipes</span>
<p>Bite-size, ready-to-deploy PyTorch code examples</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tutorials/beginner/introyt.html">
<span class="dropdown-title">Intro to PyTorch - YouTube Series</span>
<p>Master PyTorch basics with our engaging YouTube tutorial series</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Ecosystem
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem">
<span class="dropdown-title">Tools</span>
<p>Learn about the tools and frameworks in the PyTorch Ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class=dropdown-title>Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class=dropdown-title>Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class=dropdown-title>Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem/contributor-awards-2023">
<span class="dropdown-title">Contributor Awards - 2023</span>
<p>Award winners announced at this year's PyTorch Conference</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Edge
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/edge">
<span class="dropdown-title">About PyTorch Edge</span>
<p>Build innovative and privacy-aware AI experiences for edge devices</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch-overview">
<span class="dropdown-title">ExecuTorch</span>
<p>End-to-end solution for enabling on-device inference capabilities across mobile and edge devices</p>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p>Explore the documentation for comprehensive guidance on how to use PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/pytorch-domains">
<span class="dropdown-title">PyTorch Domains</span>
<p>Read the PyTorch Domains documentation to learn more about domain-specific libraries</p>
</a>
</div>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
Blogs & News
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/blog/">
<span class="dropdown-title">PyTorch Blog</span>
<p>Catch up on the latest technical news and happenings</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/community-blog">
<span class="dropdown-title">Community Blog</span>
<p>Stories from the PyTorch ecosystem</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/videos">
<span class="dropdown-title">Videos</span>
<p>Learn about the latest PyTorch tutorials, new, and more </p>
<a class="nav-dropdown-item" href="https://pytorch.org/community-stories">
<span class="dropdown-title">Community Stories</span>
<p>Learn how our community solves real, everyday machine learning problems with PyTorch</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="with-down-arrow">
About
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn more about the PyTorch Foundation</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/governing-board">
<span class="dropdown-title">Governing Board</span>
<p></p>
</a>
</div>
</div>
</li>
<li class="main-menu-item">
<div class="no-dropdown">
<a href="https://pytorch.org/join" data-cta="join">
Become a Member
</a>
</div>
</li>
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="github-icon">
</a>
</div>
</li>
<!--- TODO: This block adds the search icon to the nav bar. We will enable it later.
<li>
<div class="main-menu-item">
<a href="https://github.com/pytorch/pytorch" class="search-icon">
</a>
</div>
</li>
--->
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>2.5 ▼</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">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/custom_operators.html">PyTorch Custom Operators Landing Page</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.func.html">Extending torch.func with autograd.Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/fsdp.html">FSDP Notes</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/get_start_xpu.html">Getting Started on Intel GPU</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/mps.html">MPS backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/numerical_accuracy.html">Numerical accuracy</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
<li class="toctree-l1"><a class="reference internal" href="deploy.html">torch::deploy</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python API</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="amp.html">torch.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="library.html">torch.library</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpu.html">torch.cpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html">Understanding CUDA Memory Usage</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#generating-a-snapshot">Generating a Snapshot</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#using-the-visualizer">Using the visualizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_cuda_memory.html#snapshot-api-reference">Snapshot API Reference</a></li>
<li class="toctree-l1"><a class="reference internal" href="mps.html">torch.mps</a></li>
<li class="toctree-l1"><a class="reference internal" href="xpu.html">torch.xpu</a></li>
<li class="toctree-l1"><a class="reference internal" href="mtia.html">torch.mtia</a></li>
<li class="toctree-l1"><a class="reference internal" href="meta.html">Meta device</a></li>
<li class="toctree-l1"><a class="reference internal" href="backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="export.html">torch.export</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.tensor.html">torch.distributed.tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="fsdp.html">torch.distributed.fsdp</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.tensor.parallel.html">torch.distributed.tensor.parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.pipelining.html">torch.distributed.pipelining</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.checkpoint.html">torch.distributed.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.compiler.html">torch.compiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="func.html">torch.func</a></li>
<li class="toctree-l1"><a class="reference internal" href="futures.html">torch.futures</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.experimental.html">torch.fx.experimental</a></li>
<li class="toctree-l1"><a class="reference internal" href="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="monitor.html">torch.monitor</a></li>
<li class="toctree-l1"><a class="reference internal" href="signal.html">torch.signal</a></li>
<li class="toctree-l1"><a class="reference internal" href="special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.attention.html">torch.nn.attention</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="masked.html">torch.masked</a></li>
<li class="toctree-l1"><a class="reference internal" href="nested.html">torch.nested</a></li>
<li class="toctree-l1"><a class="reference internal" href="size.html">torch.Size</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="utils.html">torch.utils</a></li>
<li class="toctree-l1"><a class="reference internal" href="benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><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="deterministic.html">torch.utils.deterministic</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit_utils.html">torch.utils.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="module_tracker.html">torch.utils.module_tracker</a></li>
<li class="toctree-l1"><a class="reference internal" href="type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="config_mod.html">torch.__config__</a></li>
<li class="toctree-l1"><a class="reference internal" href="future_mod.html">torch.__future__</a></li>
<li class="toctree-l1"><a class="reference internal" href="logging.html">torch._logging</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_environment_variables.html">Torch Environment Variables</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/data">TorchData</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/torchrec">TorchRec</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="index.html">
Docs
</a> >
</li>
<li>torch.fx</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/fx.rst.txt" rel="nofollow"><img src="_static/images/view-page-source-icon.svg"></a>
</li>
</ul>
</div>
</div>
<div class="pytorch-shortcuts-wrapper" id="pytorch-shortcuts-wrapper">
Shortcuts
</div>
</div>
<section data-toggle="wy-nav-shift" id="pytorch-content-wrap" class="pytorch-content-wrap">
<div class="pytorch-content-left">
<!-- Google Tag Manager (noscript) -->
<noscript><iframe src="https://www.googletagmanager.com/ns.html?id=GTM-T8XT4PS"
height="0" width="0" style="display:none;visibility:hidden"></iframe></noscript>
<!-- End Google Tag Manager (noscript) -->
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<div class="section" id="torch-fx">
<h1>torch.fx<a class="headerlink" href="#torch-fx" title="Permalink to this heading">¶</a></h1>
<div class="section" id="module-torch.fx">
<span id="overview"></span><h2>Overview<a class="headerlink" href="#module-torch.fx" title="Permalink to this heading">¶</a></h2>
<p>FX is a toolkit for developers to use to transform <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code>
instances. FX consists of three main components: a <strong>symbolic tracer,</strong>
an <strong>intermediate representation</strong>, and <strong>Python code generation</strong>. A
demonstration of these components in action:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="c1"># Simple module for demonstration</span>
<span class="k">class</span> <span class="nc">MyModule</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="kc">None</span><span class="p">:</span>
<span class="nb">super</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">param</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Parameter</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">4</span><span class="p">))</span>
<span class="bp">self</span><span class="o">.</span><span class="n">linear</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">5</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">linear</span><span class="p">(</span><span class="n">x</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">param</span><span class="p">)</span><span class="o">.</span><span class="n">clamp</span><span class="p">(</span><span class="nb">min</span><span class="o">=</span><span class="mf">0.0</span><span class="p">,</span> <span class="nb">max</span><span class="o">=</span><span class="mf">1.0</span><span class="p">)</span>
<span class="n">module</span> <span class="o">=</span> <span class="n">MyModule</span><span class="p">()</span>
<span class="kn">from</span> <span class="nn">torch.fx</span> <span class="kn">import</span> <span class="n">symbolic_trace</span>
<span class="c1"># Symbolic tracing frontend - captures the semantics of the module</span>
<span class="n">symbolic_traced</span> <span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span> <span class="o">=</span> <span class="n">symbolic_trace</span><span class="p">(</span><span class="n">module</span><span class="p">)</span>
<span class="c1"># High-level intermediate representation (IR) - Graph representation</span>
<span class="nb">print</span><span class="p">(</span><span class="n">symbolic_traced</span><span class="o">.</span><span class="n">graph</span><span class="p">)</span>
<span class="sd">"""</span>
<span class="sd">graph():</span>
<span class="sd"> %x : [num_users=1] = placeholder[target=x]</span>
<span class="sd"> %param : [num_users=1] = get_attr[target=param]</span>
<span class="sd"> %add : [num_users=1] = call_function[target=operator.add](args = (%x, %param), kwargs = {})</span>
<span class="sd"> %linear : [num_users=1] = call_module[target=linear](args = (%add,), kwargs = {})</span>
<span class="sd"> %clamp : [num_users=1] = call_method[target=clamp](args = (%linear,), kwargs = {min: 0.0, max: 1.0})</span>
<span class="sd"> return clamp</span>
<span class="sd">"""</span>
<span class="c1"># Code generation - valid Python code</span>
<span class="nb">print</span><span class="p">(</span><span class="n">symbolic_traced</span><span class="o">.</span><span class="n">code</span><span class="p">)</span>
<span class="sd">"""</span>
<span class="sd">def forward(self, x):</span>
<span class="sd"> param = self.param</span>
<span class="sd"> add = x + param; x = param = None</span>
<span class="sd"> linear = self.linear(add); add = None</span>
<span class="sd"> clamp = linear.clamp(min = 0.0, max = 1.0); linear = None</span>
<span class="sd"> return clamp</span>
<span class="sd">"""</span>
</pre></div>
</div>
<p>The <strong>symbolic tracer</strong> performs “symbolic execution” of the Python
code. It feeds fake values, called Proxies, through the code. Operations
on theses Proxies are recorded. More information about symbolic tracing
can be found in the <a class="reference internal" href="#torch.fx.symbolic_trace" title="torch.fx.symbolic_trace"><code class="xref py py-func docutils literal notranslate"><span class="pre">symbolic_trace()</span></code></a> and <a class="reference internal" href="#torch.fx.Tracer" title="torch.fx.Tracer"><code class="xref py py-class docutils literal notranslate"><span class="pre">Tracer</span></code></a>
documentation.</p>
<p>The <strong>intermediate representation</strong> is the container for the operations
that were recorded during symbolic tracing. It consists of a list of
Nodes that represent function inputs, callsites (to functions, methods,
or <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> instances), and return values. More information
about the IR can be found in the documentation for <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>. The
IR is the format on which transformations are applied.</p>
<p><strong>Python code generation</strong> is what makes FX a Python-to-Python (or
Module-to-Module) transformation toolkit. For each Graph IR, we can
create valid Python code matching the Graph’s semantics. This
functionality is wrapped up in <a class="reference internal" href="#torch.fx.GraphModule" title="torch.fx.GraphModule"><code class="xref py py-class docutils literal notranslate"><span class="pre">GraphModule</span></code></a>, which is 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> instance that holds a <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a> as well as a
<code class="docutils literal notranslate"><span class="pre">forward</span></code> method generated from the Graph.</p>
<p>Taken together, this pipeline of components (symbolic tracing ->
intermediate representation -> transforms -> Python code generation)
constitutes the Python-to-Python transformation pipeline of FX. In
addition, these components can be used separately. For example,
symbolic tracing can be used in isolation to capture a form of
the code for analysis (and not transformation) purposes. Code
generation can be used for programmatically generating models, for
example from a config file. There are many uses for FX!</p>
<p>Several example transformations can be found at the
<a class="reference external" href="https://github.com/pytorch/examples/tree/master/fx">examples</a>
repository.</p>
</div>
<div class="section" id="writing-transformations">
<span id="id1"></span><h2>Writing Transformations<a class="headerlink" href="#writing-transformations" title="Permalink to this heading">¶</a></h2>
<p>What is an FX transform? Essentially, it’s a function that looks like this.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.fx</span>
<span class="k">def</span> <span class="nf">transform</span><span class="p">(</span><span class="n">m</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="n">tracer_class</span> <span class="p">:</span> <span class="nb">type</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">Tracer</span><span class="p">)</span> <span class="o">-></span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">:</span>
<span class="c1"># Step 1: Acquire a Graph representing the code in `m`</span>
<span class="c1"># NOTE: torch.fx.symbolic_trace is a wrapper around a call to</span>
<span class="c1"># fx.Tracer.trace and constructing a GraphModule. We'll</span>
<span class="c1"># split that out in our transform to allow the caller to</span>
<span class="c1"># customize tracing behavior.</span>
<span class="n">graph</span> <span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">Graph</span> <span class="o">=</span> <span class="n">tracer_class</span><span class="p">()</span><span class="o">.</span><span class="n">trace</span><span class="p">(</span><span class="n">m</span><span class="p">)</span>
<span class="c1"># Step 2: Modify this Graph or create a new one</span>
<span class="n">graph</span> <span class="o">=</span> <span class="o">...</span>
<span class="c1"># Step 3: Construct a Module to return</span>
<span class="k">return</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">graph</span><span class="p">)</span>
</pre></div>
</div>
<p>Your transform will take in 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>, acquire a <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>
from it, do some modifications, and return a new
<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>. You should think of the <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> that your FX
transform returns as identical to a regular <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> – you can pass it to another
FX transform, you can pass it to TorchScript, or you can
run it. Ensuring that the inputs and outputs of your FX transform are 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> will allow for composability.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>It is also possible to modify an existing <a class="reference internal" href="#torch.fx.GraphModule" title="torch.fx.GraphModule"><code class="xref py py-class docutils literal notranslate"><span class="pre">GraphModule</span></code></a> instead of
creating a new one, like so:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.fx</span>
<span class="k">def</span> <span class="nf">transform</span><span class="p">(</span><span class="n">m</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="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">:</span>
<span class="n">gm</span> <span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">symbolic_trace</span><span class="p">(</span><span class="n">m</span><span class="p">)</span>
<span class="c1"># Modify gm.graph</span>
<span class="c1"># <...></span>
<span class="c1"># Recompile the forward() method of `gm` from its Graph</span>
<span class="n">gm</span><span class="o">.</span><span class="n">recompile</span><span class="p">()</span>
<span class="k">return</span> <span class="n">gm</span>
</pre></div>
</div>
<p>Note that you MUST call <a class="reference internal" href="#torch.fx.GraphModule.recompile" title="torch.fx.GraphModule.recompile"><code class="xref py py-meth docutils literal notranslate"><span class="pre">GraphModule.recompile()</span></code></a> to bring the generated
<code class="docutils literal notranslate"><span class="pre">forward()</span></code> method on the <code class="docutils literal notranslate"><span class="pre">GraphModule</span></code> in sync with the modified <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>.</p>
</div>
<p>Given that you’ve passed in 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> that has been traced into a
<a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>, there are now two primary approaches you can take to building a new
<a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>.</p>
<div class="section" id="a-quick-primer-on-graphs">
<h3>A Quick Primer on Graphs<a class="headerlink" href="#a-quick-primer-on-graphs" title="Permalink to this heading">¶</a></h3>
<p>Full treatment of the semantics of graphs can be found in the <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>
documentation, but we are going to cover the basics here. A <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a> is
a data structure that represents a method on a <a class="reference internal" href="#torch.fx.GraphModule" title="torch.fx.GraphModule"><code class="xref py py-class docutils literal notranslate"><span class="pre">GraphModule</span></code></a>. The
information that this requires is:</p>
<ul class="simple">
<li><p>What are the inputs to the method?</p></li>
<li><p>What are the operations that run inside the method?</p></li>
<li><p>What is the output (i.e. return) value from the method?</p></li>
</ul>
<p>All three of these concepts are represented with <a class="reference internal" href="#torch.fx.Node" title="torch.fx.Node"><code class="xref py py-class docutils literal notranslate"><span class="pre">Node</span></code></a> instances.
Let’s see what we mean by that with a short example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.fx</span>
<span class="k">class</span> <span class="nc">MyModule</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">param</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Parameter</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">4</span><span class="p">))</span>
<span class="bp">self</span><span class="o">.</span><span class="n">linear</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">5</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="n">torch</span><span class="o">.</span><span class="n">topk</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span>
<span class="bp">self</span><span class="o">.</span><span class="n">linear</span><span class="p">(</span><span class="n">x</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">linear</span><span class="o">.</span><span class="n">weight</span><span class="p">)</span><span class="o">.</span><span class="n">relu</span><span class="p">(),</span> <span class="n">dim</span><span class="o">=-</span><span class="mi">1</span><span class="p">),</span> <span class="mi">3</span><span class="p">)</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">MyModule</span><span class="p">()</span>
<span class="n">gm</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">symbolic_trace</span><span class="p">(</span><span class="n">m</span><span class="p">)</span>
<span class="n">gm</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">print_tabular</span><span class="p">()</span>
</pre></div>
</div>
<p>Here we define a module <code class="docutils literal notranslate"><span class="pre">MyModule</span></code> for demonstration purposes, instantiate it,
symbolically trace it, then call the <a class="reference internal" href="#torch.fx.Graph.print_tabular" title="torch.fx.Graph.print_tabular"><code class="xref py py-meth docutils literal notranslate"><span class="pre">Graph.print_tabular()</span></code></a> method to print
out a table showing the nodes of this <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>:</p>
<blockquote>
<div><table class="docutils colwidths-auto align-default">
<thead>
<tr class="row-odd"><th class="head"><p>opcode</p></th>
<th class="head"><p>name</p></th>
<th class="head"><p>target</p></th>
<th class="head"><p>args</p></th>
<th class="head"><p>kwargs</p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><p>placeholder</p></td>
<td><p>x</p></td>
<td><p>x</p></td>
<td><p>()</p></td>
<td><p>{}</p></td>
</tr>
<tr class="row-odd"><td><p>get_attr</p></td>
<td><p>linear_weight</p></td>
<td><p>linear.weight</p></td>
<td><p>()</p></td>
<td><p>{}</p></td>
</tr>
<tr class="row-even"><td><p>call_function</p></td>
<td><p>add_1</p></td>
<td><p><built-in function add></p></td>
<td><p>(x, linear_weight)</p></td>
<td><p>{}</p></td>
</tr>
<tr class="row-odd"><td><p>call_module</p></td>
<td><p>linear_1</p></td>
<td><p>linear</p></td>
<td><p>(add_1,)</p></td>
<td><p>{}</p></td>
</tr>
<tr class="row-even"><td><p>call_method</p></td>
<td><p>relu_1</p></td>
<td><p>relu</p></td>
<td><p>(linear_1,)</p></td>
<td><p>{}</p></td>
</tr>
<tr class="row-odd"><td><p>call_function</p></td>
<td><p>sum_1</p></td>
<td><p><built-in method sum …></p></td>
<td><p>(relu_1,)</p></td>
<td><p>{‘dim’: -1}</p></td>
</tr>
<tr class="row-even"><td><p>call_function</p></td>
<td><p>topk_1</p></td>
<td><p><built-in method topk …></p></td>
<td><p>(sum_1, 3)</p></td>
<td><p>{}</p></td>
</tr>
<tr class="row-odd"><td><p>output</p></td>
<td><p>output</p></td>
<td><p>output</p></td>
<td><p>(topk_1,)</p></td>
<td><p>{}</p></td>
</tr>
</tbody>
</table>
</div></blockquote>
<p>We can use this information to answer the questions we posed above.</p>
<ul class="simple">
<li><p>What are the inputs to the method? In FX, method inputs are specified
via special <code class="docutils literal notranslate"><span class="pre">placeholder</span></code> nodes. In this case, we have a single
<code class="docutils literal notranslate"><span class="pre">placeholder</span></code> node with a <code class="docutils literal notranslate"><span class="pre">target</span></code> of <code class="docutils literal notranslate"><span class="pre">x</span></code>, meaning we have
a single (non-self) argument named x.</p></li>
<li><p>What are the operations within the method? The <code class="docutils literal notranslate"><span class="pre">get_attr</span></code>,
<code class="docutils literal notranslate"><span class="pre">call_function</span></code>, <code class="docutils literal notranslate"><span class="pre">call_module</span></code>, and <code class="docutils literal notranslate"><span class="pre">call_method</span></code> nodes
represent the operations in the method. A full treatment of
the semantics of all of these can be found in the <a class="reference internal" href="#torch.fx.Node" title="torch.fx.Node"><code class="xref py py-class docutils literal notranslate"><span class="pre">Node</span></code></a>
documentation.</p></li>
<li><p>What is the return value of the method? The return value in a
<a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a> is specified by a special <code class="docutils literal notranslate"><span class="pre">output</span></code> node.</p></li>
</ul>
<p>Given that we now know the basics of how code is represented in
FX, we can now explore how we would edit a <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>.</p>
</div>
<div class="section" id="graph-manipulation">
<h3>Graph Manipulation<a class="headerlink" href="#graph-manipulation" title="Permalink to this heading">¶</a></h3>
<div class="section" id="direct-graph-manipulation">
<h4>Direct Graph Manipulation<a class="headerlink" href="#direct-graph-manipulation" title="Permalink to this heading">¶</a></h4>
<p>One approach to building this new <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a> is to directly manipulate your old
one. To aid in this, we can simply take the <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a> we obtain from symbolic
tracing and modify it. For example, let’s say we desire to replace
<a class="reference internal" href="generated/torch.add.html#torch.add" title="torch.add"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.add()</span></code></a> calls with <a class="reference internal" href="generated/torch.mul.html#torch.mul" title="torch.mul"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.mul()</span></code></a> calls.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.fx</span>
<span class="c1"># Sample module</span>
<span class="k">class</span> <span class="nc">M</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">):</span>
<span class="k">return</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">transform</span><span class="p">(</span><span class="n">m</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">,</span>
<span class="n">tracer_class</span> <span class="p">:</span> <span class="nb">type</span> <span class="o">=</span> <span class="n">fx</span><span class="o">.</span><span class="n">Tracer</span><span class="p">)</span> <span class="o">-></span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">:</span>
<span class="n">graph</span> <span class="p">:</span> <span class="n">fx</span><span class="o">.</span><span class="n">Graph</span> <span class="o">=</span> <span class="n">tracer_class</span><span class="p">()</span><span class="o">.</span><span class="n">trace</span><span class="p">(</span><span class="n">m</span><span class="p">)</span>
<span class="c1"># FX represents its Graph as an ordered list of</span>
<span class="c1"># nodes, so we can iterate through them.</span>
<span class="k">for</span> <span class="n">node</span> <span class="ow">in</span> <span class="n">graph</span><span class="o">.</span><span class="n">nodes</span><span class="p">:</span>
<span class="c1"># Checks if we're calling a function (i.e:</span>
<span class="c1"># torch.add)</span>
<span class="k">if</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s1">'call_function'</span><span class="p">:</span>
<span class="c1"># The target attribute is the function</span>
<span class="c1"># that call_function calls.</span>
<span class="k">if</span> <span class="n">node</span><span class="o">.</span><span class="n">target</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">add</span><span class="p">:</span>
<span class="n">node</span><span class="o">.</span><span class="n">target</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">mul</span>
<span class="n">graph</span><span class="o">.</span><span class="n">lint</span><span class="p">()</span> <span class="c1"># Does some checks to make sure the</span>
<span class="c1"># Graph is well-formed.</span>
<span class="k">return</span> <span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">graph</span><span class="p">)</span>
</pre></div>
</div>
<p>We can also do more involved <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a> rewrites, such as
deleting or appending nodes. To aid in these transformations,
FX has utility functions for transforming the graph that can
be found in the <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a> documentation. An
example of using these APIs to append a <code class="xref py py-func docutils literal notranslate"><span class="pre">torch.relu()</span></code> call
can be found below.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Specifies the insertion point. Any nodes added to the</span>
<span class="c1"># Graph within this scope will be inserted after `node`</span>
<span class="k">with</span> <span class="n">traced</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">inserting_after</span><span class="p">(</span><span class="n">node</span><span class="p">):</span>
<span class="c1"># Insert a new `call_function` node calling `torch.relu`</span>
<span class="n">new_node</span> <span class="o">=</span> <span class="n">traced</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">call_function</span><span class="p">(</span>
<span class="n">torch</span><span class="o">.</span><span class="n">relu</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">node</span><span class="p">,))</span>
<span class="c1"># We want all places that used the value of `node` to</span>
<span class="c1"># now use that value after the `relu` call we've added.</span>
<span class="c1"># We use the `replace_all_uses_with` API to do this.</span>
<span class="n">node</span><span class="o">.</span><span class="n">replace_all_uses_with</span><span class="p">(</span><span class="n">new_node</span><span class="p">)</span>
</pre></div>
</div>
<p>For simple transformations that only consist of substitutions, you can also
make use of the <a class="reference external" href="https://github.com/pytorch/pytorch/blob/main/torch/fx/subgraph_rewriter.py">subgraph rewriter.</a></p>
</div>
<div class="section" id="subgraph-rewriting-with-replace-pattern">
<h4>Subgraph Rewriting With replace_pattern()<a class="headerlink" href="#subgraph-rewriting-with-replace-pattern" title="Permalink to this heading">¶</a></h4>
<p>FX also provides another level of automation on top of direct graph manipulation.
The <a class="reference internal" href="#torch.fx.replace_pattern" title="torch.fx.replace_pattern"><code class="xref py py-func docutils literal notranslate"><span class="pre">replace_pattern()</span></code></a> API is essentially a “find/replace” tool for editing
<a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>s. It allows you to specify a <code class="docutils literal notranslate"><span class="pre">pattern</span></code> and <code class="docutils literal notranslate"><span class="pre">replacement</span></code> function
and it will trace through those functions, find instances of the group of operations
in the <code class="docutils literal notranslate"><span class="pre">pattern</span></code> graph, and replace those instances with copies of the <code class="docutils literal notranslate"><span class="pre">replacement</span></code>
graph. This can help to greatly automate tedious graph manipulation code, which can
get unwieldy as the transformations get more complex.</p>
</div>
<div class="section" id="graph-manipulation-examples">
<h4>Graph Manipulation Examples<a class="headerlink" href="#graph-manipulation-examples" title="Permalink to this heading">¶</a></h4>
<ul class="simple">
<li><p><a class="reference external" href="https://github.com/pytorch/examples/blob/master/fx/replace_op.py">Replace one
op</a></p></li>
<li><p><a class="reference external" href="https://github.com/pytorch/pytorch/blob/40cbf342d3c000712da92cfafeaca651b3e0bd3e/torch/fx/experimental/optimization.py#L50">Conv/Batch Norm
fusion</a></p></li>
<li><p><a class="reference external" href="https://github.com/pytorch/examples/blob/master/fx/subgraph_rewriter_basic_use.py">replace_pattern: Basic usage</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/docs/main/quantization.html#prototype-fx-graph-mode-quantization">Quantization</a></p></li>
<li><p><a class="reference external" href="https://github.com/pytorch/examples/blob/master/fx/invert.py">Invert Transformation</a></p></li>
</ul>
</div>
</div>
<div class="section" id="proxy-retracing">
<h3>Proxy/Retracing<a class="headerlink" href="#proxy-retracing" title="Permalink to this heading">¶</a></h3>
<p>Another way of manipulating <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>s is by reusing the <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a>
machinery used in symbolic tracing. For example, let’s
imagine that we wanted to write a transformation that decomposed
PyTorch functions into smaller operations. It would transform every
<code class="docutils literal notranslate"><span class="pre">F.relu(x)</span></code> call into <code class="docutils literal notranslate"><span class="pre">(x</span> <span class="pre">></span> <span class="pre">0)</span> <span class="pre">*</span> <span class="pre">x</span></code>. One possibility would be to
perform the requisite graph rewriting to insert the comparison and
multiplication after the <code class="docutils literal notranslate"><span class="pre">F.relu</span></code>, and then clean up the original
<code class="docutils literal notranslate"><span class="pre">F.relu</span></code>. However, we can automate this process by using <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a>
objects to automatically record operations into the <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>.</p>
<p>To use this method, we write the operations that we want inserted as regular
PyTorch code and invoke that code with <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a> objects as arguments.
These <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a> objects will capture the operations that are performed
on them and append them to the <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a>.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Note that this decomposition rule can be read as regular Python</span>
<span class="k">def</span> <span class="nf">relu_decomposition</span><span class="p">(</span><span class="n">x</span><span class="p">):</span>
<span class="k">return</span> <span class="p">(</span><span class="n">x</span> <span class="o">></span> <span class="mi">0</span><span class="p">)</span> <span class="o">*</span> <span class="n">x</span>
<span class="n">decomposition_rules</span> <span class="o">=</span> <span class="p">{}</span>
<span class="n">decomposition_rules</span><span class="p">[</span><span class="n">F</span><span class="o">.</span><span class="n">relu</span><span class="p">]</span> <span class="o">=</span> <span class="n">relu_decomposition</span>
<span class="k">def</span> <span class="nf">decompose</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">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">,</span>
<span class="n">tracer_class</span> <span class="p">:</span> <span class="nb">type</span> <span class="o">=</span> <span class="n">fx</span><span class="o">.</span><span class="n">Tracer</span><span class="p">)</span> <span class="o">-></span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">:</span>
<span class="w"> </span><span class="sd">"""</span>
<span class="sd"> Decompose `model` into smaller constituent operations.</span>
<span class="sd"> Currently,this only supports decomposing ReLU into its</span>
<span class="sd"> mathematical definition: (x > 0) * x</span>
<span class="sd"> """</span>
<span class="n">graph</span> <span class="p">:</span> <span class="n">fx</span><span class="o">.</span><span class="n">Graph</span> <span class="o">=</span> <span class="n">tracer_class</span><span class="p">()</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">new_graph</span> <span class="o">=</span> <span class="n">fx</span><span class="o">.</span><span class="n">Graph</span><span class="p">()</span>
<span class="n">env</span> <span class="o">=</span> <span class="p">{}</span>
<span class="n">tracer</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">proxy</span><span class="o">.</span><span class="n">GraphAppendingTracer</span><span class="p">(</span><span class="n">new_graph</span><span class="p">)</span>
<span class="k">for</span> <span class="n">node</span> <span class="ow">in</span> <span class="n">graph</span><span class="o">.</span><span class="n">nodes</span><span class="p">:</span>
<span class="k">if</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s1">'call_function'</span> <span class="ow">and</span> <span class="n">node</span><span class="o">.</span><span class="n">target</span> <span class="ow">in</span> <span class="n">decomposition_rules</span><span class="p">:</span>
<span class="c1"># By wrapping the arguments with proxies,</span>
<span class="c1"># we can dispatch to the appropriate</span>
<span class="c1"># decomposition rule and implicitly add it</span>
<span class="c1"># to the Graph by symbolically tracing it.</span>
<span class="n">proxy_args</span> <span class="o">=</span> <span class="p">[</span>
<span class="n">fx</span><span class="o">.</span><span class="n">Proxy</span><span class="p">(</span><span class="n">env</span><span class="p">[</span><span class="n">x</span><span class="o">.</span><span class="n">name</span><span class="p">],</span> <span class="n">tracer</span><span class="p">)</span> <span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">fx</span><span class="o">.</span><span class="n">Node</span><span class="p">)</span> <span class="k">else</span> <span class="n">x</span> <span class="k">for</span> <span class="n">x</span> <span class="ow">in</span> <span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">]</span>
<span class="n">output_proxy</span> <span class="o">=</span> <span class="n">decomposition_rules</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">target</span><span class="p">](</span><span class="o">*</span><span class="n">proxy_args</span><span class="p">)</span>
<span class="c1"># Operations on `Proxy` always yield new `Proxy`s, and the</span>
<span class="c1"># return value of our decomposition rule is no exception.</span>
<span class="c1"># We need to extract the underlying `Node` from the `Proxy`</span>
<span class="c1"># to use it in subsequent iterations of this transform.</span>
<span class="n">new_node</span> <span class="o">=</span> <span class="n">output_proxy</span><span class="o">.</span><span class="n">node</span>
<span class="n">env</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">new_node</span>
<span class="k">else</span><span class="p">:</span>
<span class="c1"># Default case: we don't have a decomposition rule for this</span>
<span class="c1"># node, so just copy the node over into the new graph.</span>
<span class="n">new_node</span> <span class="o">=</span> <span class="n">new_graph</span><span class="o">.</span><span class="n">node_copy</span><span class="p">(</span><span class="n">node</span><span class="p">,</span> <span class="k">lambda</span> <span class="n">x</span><span class="p">:</span> <span class="n">env</span><span class="p">[</span><span class="n">x</span><span class="o">.</span><span class="n">name</span><span class="p">])</span>
<span class="n">env</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">new_node</span>
<span class="k">return</span> <span class="n">fx</span><span class="o">.</span><span class="n">GraphModule</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">new_graph</span><span class="p">)</span>
</pre></div>
</div>
<p>In addition to avoiding explicit graph manipulation, using <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a>s
also allows you to specify your rewrite rules as native Python code.
For transformations that require a large amount of rewrite rules
(such as vmap or grad), this can often improve readability and
maintainability of the rules. Note that while calling <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a> we also
passed a tracer pointing to the underlying variable <cite>graph</cite>. This is done so
if in case the operations in graph are n-ary (e.g. add is a binary operator)
the call to <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a> does not create multiple instances of a graph
tracer which can lead to unexpected runtime errors. We recommend this method
of using <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a> especially when the underlying operators can not be
safely assumed to be unary.</p>
<p>A worked example of using <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a>s for <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a> manipulation
can be found
<a class="reference external" href="https://github.com/pytorch/examples/blob/master/fx/proxy_based_graph_creation.py">here</a>.</p>
</div>
<div class="section" id="the-interpreter-pattern">
<h3>The Interpreter Pattern<a class="headerlink" href="#the-interpreter-pattern" title="Permalink to this heading">¶</a></h3>
<p>A useful code organizational pattern in FX is to loop over all the <a class="reference internal" href="#torch.fx.Node" title="torch.fx.Node"><code class="xref py py-class docutils literal notranslate"><span class="pre">Node</span></code></a>s
in a <a class="reference internal" href="#torch.fx.Graph" title="torch.fx.Graph"><code class="xref py py-class docutils literal notranslate"><span class="pre">Graph</span></code></a> and execute them. This can be used for several things including
runtime analysis of values flowing through the graph or transformation of the code
via retracing with <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a>s. For example, suppose we want to run a
<a class="reference internal" href="#torch.fx.GraphModule" title="torch.fx.GraphModule"><code class="xref py py-class docutils literal notranslate"><span class="pre">GraphModule</span></code></a> and record the <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> shape and dtype
properties on the nodes as we see them at runtime. That might look like:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.fx</span>
<span class="kn">from</span> <span class="nn">torch.fx.node</span> <span class="kn">import</span> <span class="n">Node</span>
<span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">Dict</span>
<span class="k">class</span> <span class="nc">ShapeProp</span><span class="p">:</span>
<span class="w"> </span><span class="sd">"""</span>
<span class="sd"> Shape propagation. This class takes a `GraphModule`.</span>
<span class="sd"> Then, its `propagate` method executes the `GraphModule`</span>
<span class="sd"> node-by-node with the given arguments. As each operation</span>
<span class="sd"> executes, the ShapeProp class stores away the shape and</span>
<span class="sd"> element type for the output values of each operation on</span>
<span class="sd"> the `shape` and `dtype` attributes of the operation's</span>
<span class="sd"> `Node`.</span>
<span class="sd"> """</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">mod</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">mod</span> <span class="o">=</span> <span class="n">mod</span>
<span class="bp">self</span><span class="o">.</span><span class="n">graph</span> <span class="o">=</span> <span class="n">mod</span><span class="o">.</span><span class="n">graph</span>
<span class="bp">self</span><span class="o">.</span><span class="n">modules</span> <span class="o">=</span> <span class="nb">dict</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">mod</span><span class="o">.</span><span class="n">named_modules</span><span class="p">())</span>
<span class="k">def</span> <span class="nf">propagate</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="o">*</span><span class="n">args</span><span class="p">):</span>
<span class="n">args_iter</span> <span class="o">=</span> <span class="nb">iter</span><span class="p">(</span><span class="n">args</span><span class="p">)</span>
<span class="n">env</span> <span class="p">:</span> <span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">Node</span><span class="p">]</span> <span class="o">=</span> <span class="p">{}</span>
<span class="k">def</span> <span class="nf">load_arg</span><span class="p">(</span><span class="n">a</span><span class="p">):</span>
<span class="k">return</span> <span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">map_arg</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="k">lambda</span> <span class="n">n</span><span class="p">:</span> <span class="n">env</span><span class="p">[</span><span class="n">n</span><span class="o">.</span><span class="n">name</span><span class="p">])</span>
<span class="k">def</span> <span class="nf">fetch_attr</span><span class="p">(</span><span class="n">target</span> <span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
<span class="n">target_atoms</span> <span class="o">=</span> <span class="n">target</span><span class="o">.</span><span class="n">split</span><span class="p">(</span><span class="s1">'.'</span><span class="p">)</span>
<span class="n">attr_itr</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">mod</span>
<span class="k">for</span> <span class="n">i</span><span class="p">,</span> <span class="n">atom</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">target_atoms</span><span class="p">):</span>
<span class="k">if</span> <span class="ow">not</span> <span class="nb">hasattr</span><span class="p">(</span><span class="n">attr_itr</span><span class="p">,</span> <span class="n">atom</span><span class="p">):</span>
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Node referenced nonexistent target </span><span class="si">{</span><span class="s1">'.'</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">target_atoms</span><span class="p">[:</span><span class="n">i</span><span class="p">])</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="n">attr_itr</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">attr_itr</span><span class="p">,</span> <span class="n">atom</span><span class="p">)</span>
<span class="k">return</span> <span class="n">attr_itr</span>
<span class="k">for</span> <span class="n">node</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">nodes</span><span class="p">:</span>
<span class="k">if</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s1">'placeholder'</span><span class="p">:</span>
<span class="n">result</span> <span class="o">=</span> <span class="nb">next</span><span class="p">(</span><span class="n">args_iter</span><span class="p">)</span>
<span class="k">elif</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s1">'get_attr'</span><span class="p">:</span>
<span class="n">result</span> <span class="o">=</span> <span class="n">fetch_attr</span><span class="p">(</span><span class="n">node</span><span class="o">.</span><span class="n">target</span><span class="p">)</span>
<span class="k">elif</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s1">'call_function'</span><span class="p">:</span>
<span class="n">result</span> <span class="o">=</span> <span class="n">node</span><span class="o">.</span><span class="n">target</span><span class="p">(</span><span class="o">*</span><span class="n">load_arg</span><span class="p">(</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">),</span> <span class="o">**</span><span class="n">load_arg</span><span class="p">(</span><span class="n">node</span><span class="o">.</span><span class="n">kwargs</span><span class="p">))</span>
<span class="k">elif</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s1">'call_method'</span><span class="p">:</span>
<span class="n">self_obj</span><span class="p">,</span> <span class="o">*</span><span class="n">args</span> <span class="o">=</span> <span class="n">load_arg</span><span class="p">(</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">)</span>
<span class="n">kwargs</span> <span class="o">=</span> <span class="n">load_arg</span><span class="p">(</span><span class="n">node</span><span class="o">.</span><span class="n">kwargs</span><span class="p">)</span>
<span class="n">result</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">self_obj</span><span class="p">,</span> <span class="n">node</span><span class="o">.</span><span class="n">target</span><span class="p">)(</span><span class="o">*</span><span class="n">args</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">)</span>
<span class="k">elif</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s1">'call_module'</span><span class="p">:</span>
<span class="n">result</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">modules</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">target</span><span class="p">](</span><span class="o">*</span><span class="n">load_arg</span><span class="p">(</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">),</span> <span class="o">**</span><span class="n">load_arg</span><span class="p">(</span><span class="n">node</span><span class="o">.</span><span class="n">kwargs</span><span class="p">))</span>
<span class="c1"># This is the only code specific to shape propagation.</span>
<span class="c1"># you can delete this `if` branch and this becomes</span>
<span class="c1"># a generic GraphModule interpreter.</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">result</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">):</span>
<span class="n">node</span><span class="o">.</span><span class="n">shape</span> <span class="o">=</span> <span class="n">result</span><span class="o">.</span><span class="n">shape</span>
<span class="n">node</span><span class="o">.</span><span class="n">dtype</span> <span class="o">=</span> <span class="n">result</span><span class="o">.</span><span class="n">dtype</span>
<span class="n">env</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">result</span>
<span class="k">return</span> <span class="n">load_arg</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">result</span><span class="p">)</span>
</pre></div>
</div>
<p>As you can see, a full interpreter for FX is not that complicated
but it can be very useful. To ease using this pattern, we provide
the <a class="reference internal" href="#torch.fx.Interpreter" title="torch.fx.Interpreter"><code class="xref py py-class docutils literal notranslate"><span class="pre">Interpreter</span></code></a> class, which encompasses the above logic
in a way that certain aspects of the interpreter’s execution can
be overridden via method overrides.</p>
<p>In addition to executing operations, we can also generate a new
<cite>Graph</cite> by feeding <a class="reference internal" href="#torch.fx.Proxy" title="torch.fx.Proxy"><code class="xref py py-class docutils literal notranslate"><span class="pre">Proxy</span></code></a> values through an interpreter.
Similarly, we provide the <a class="reference internal" href="#torch.fx.Transformer" title="torch.fx.Transformer"><code class="xref py py-class docutils literal notranslate"><span class="pre">Transformer</span></code></a> class to encompass
this pattern. <a class="reference internal" href="#torch.fx.Transformer" title="torch.fx.Transformer"><code class="xref py py-class docutils literal notranslate"><span class="pre">Transformer</span></code></a> behaves similarly to
<a class="reference internal" href="#torch.fx.Interpreter" title="torch.fx.Interpreter"><code class="xref py py-class docutils literal notranslate"><span class="pre">Interpreter</span></code></a>, but instead of calling the <code class="docutils literal notranslate"><span class="pre">run</span></code> method to
get a concrete output value from the Module, you would call the
<a class="reference internal" href="#torch.fx.Transformer.transform" title="torch.fx.Transformer.transform"><code class="xref py py-meth docutils literal notranslate"><span class="pre">Transformer.transform()</span></code></a> method to return a new
<a class="reference internal" href="#torch.fx.GraphModule" title="torch.fx.GraphModule"><code class="xref py py-class docutils literal notranslate"><span class="pre">GraphModule</span></code></a> which was subject to any transformation rules
you installed as overridden methods.</p>
<div class="section" id="examples-of-the-interpreter-pattern">