-
Notifications
You must be signed in to change notification settings - Fork 362
/
Copy pathgetting_started_with_fx_path.html
1047 lines (840 loc) · 68.8 KB
/
getting_started_with_fx_path.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 0.17.1: http://docutils.sourceforge.net/" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Torch-TensorRT (FX Frontend) User Guide — Torch-TensorRT v2.2.0.dev0+f617898 documentation</title>
<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/sg_gallery.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-binder.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-dataframe.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-rendered-html.css" type="text/css" />
<link rel="stylesheet" href="../_static/collapsible-lists/css/tree_view.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.10.0-beta/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/custom.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-TensorRT torch.compile Backend" href="torch_compile.html" />
<link rel="prev" title="Creating a TorchScript Module" href="creating_torchscript_module_in_python.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','');</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/katex@0.10.0/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/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>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-arrow">
PyTorch 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>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch">
<span class="dropdown-title">ExecuTorch</span>
</a>
</div>
</div>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li>
<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">
v2.2.0.dev0+f617898
</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">Getting Started</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../getting_started/installation.html">Installation</a></li>
<li class="toctree-l1"><a class="reference internal" href="../getting_started/getting_started_with_python_api.html">Using Torch-TensorRT in Python</a></li>
<li class="toctree-l1"><a class="reference internal" href="../getting_started/getting_started_with_cpp_api.html">Using Torch-TensorRT in C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="../getting_started/getting_started_with_windows.html">Building Torch-TensorRT on Windows</a></li>
<li class="toctree-l1"><a class="reference internal" href="../getting_started/getting_started_with_windows.html#building-with-visual-studio-code">Building With Visual Studio Code</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">User Guide</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="creating_torchscript_module_in_python.html">Creating a TorchScript Module</a></li>
<li class="toctree-l1"><a class="reference internal" href="creating_torchscript_module_in_python.html#working-with-torchscript-in-python">Working with TorchScript in Python</a></li>
<li class="toctree-l1"><a class="reference internal" href="creating_torchscript_module_in_python.html#saving-torchscript-module-to-disk">Saving TorchScript Module to Disk</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Torch-TensorRT (FX Frontend) User Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_compile.html">Torch-TensorRT <cite>torch.compile</cite> Backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamo_export.html">Torch-TensorRT Dynamo Backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="ptq.html">Post Training Quantization (PTQ)</a></li>
<li class="toctree-l1"><a class="reference internal" href="runtime.html">Deploying Torch-TensorRT Programs</a></li>
<li class="toctree-l1"><a class="reference internal" href="saving_models.html">Saving models compiled with Torch-TensorRT</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamic_shapes.html">Dynamic shapes with Torch-TensorRT</a></li>
<li class="toctree-l1"><a class="reference internal" href="use_from_pytorch.html">Using Torch-TensorRT Directly From PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="using_dla.html">DLA</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Tutorials</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../tutorials/serving_torch_tensorrt_with_triton.html">Serving a Torch-TensorRT model with Triton</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tutorials/notebooks.html">Example notebooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tutorials/_rendered_examples/dynamo/torch_compile_resnet_example.html">Compiling ResNet using the Torch-TensorRT <cite>torch.compile</cite> Backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tutorials/_rendered_examples/dynamo/torch_compile_transformers_example.html">Compiling a Transformer using torch.compile and TensorRT</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tutorials/_rendered_examples/dynamo/torch_compile_advanced_usage.html">Torch Compile Advanced Usage</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python API Documenation</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../py_api/torch_tensorrt.html">torch_tensorrt</a></li>
<li class="toctree-l1"><a class="reference internal" href="../py_api/logging.html">torch_tensorrt.logging</a></li>
<li class="toctree-l1"><a class="reference internal" href="../py_api/ptq.html">torch_tensorrt.ptq</a></li>
<li class="toctree-l1"><a class="reference internal" href="../py_api/ts.html">torch_tensorrt.ts</a></li>
<li class="toctree-l1"><a class="reference internal" href="../py_api/fx.html">torch_tensorrt.fx</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">C++ API Documenation</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../_cpp_api/torch_tensort_cpp.html">Torch-TensorRT C++ API</a></li>
<li class="toctree-l1"><a class="reference internal" href="../_cpp_api/namespace_torch_tensorrt.html">Namespace torch_tensorrt</a></li>
<li class="toctree-l1"><a class="reference internal" href="../_cpp_api/namespace_torch_tensorrt__logging.html">Namespace torch_tensorrt::logging</a></li>
<li class="toctree-l1"><a class="reference internal" href="../_cpp_api/namespace_torch_tensorrt__torchscript.html">Namespace torch_tensorrt::torchscript</a></li>
<li class="toctree-l1"><a class="reference internal" href="../_cpp_api/namespace_torch_tensorrt__ptq.html">Namespace torch_tensorrt::ptq</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">CLI Documenation</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../cli/torchtrtc.html">torchtrtc</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Contributor Documentation</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../contributors/system_overview.html">System Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="../contributors/writing_converters.html">Writing Converters</a></li>
<li class="toctree-l1"><a class="reference internal" href="../contributors/writing_dynamo_aten_lowering_passes.html">Writing Dynamo ATen Lowering Passes</a></li>
<li class="toctree-l1"><a class="reference internal" href="../contributors/useful_links.html">Useful Links for Torch-TensorRT Development</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Indices</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../indices/supported_ops.html">Operators Supported</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-TensorRT (FX Frontend) User Guide</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/user_guide/getting_started_with_fx_path.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="
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">
<section id="torch-tensorrt-fx-frontend-user-guide">
<span id="getting-started-with-fx"></span><h1>Torch-TensorRT (FX Frontend) User Guide<a class="headerlink" href="#torch-tensorrt-fx-frontend-user-guide" title="Permalink to this headline">¶</a></h1>
<p>Torch-TensorRT (FX Frontend) is a tool that can convert a PyTorch model through <code class="docutils literal notranslate"><span class="pre">torch.fx</span></code> to an
TensorRT engine optimized targeting running on Nvidia GPUs. TensorRT is the inference engine
developed by NVIDIA which composed of various kinds of optimization including kernel fusion,
graph optimization, low precision, etc.. This tool is developed in Python environment which allows this
workflow to be very accessible to researchers and engineers. There are a few stages that a
user want to use this tool and we will introduce them here.</p>
<p>> Torch-TensorRT (FX Frontend) is in <code class="docutils literal notranslate"><span class="pre">Beta</span></code> and currently it is recommended to work with PyTorch nightly.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span><span class="c1"># Test an example by</span>
$<span class="w"> </span>python<span class="w"> </span>py/torch_tensorrt/fx/example/lower_example.py
</pre></div>
</div>
<section id="converting-a-pytorch-model-to-tensorrt-engine">
<h2>Converting a PyTorch Model to TensorRT Engine<a class="headerlink" href="#converting-a-pytorch-model-to-tensorrt-engine" title="Permalink to this headline">¶</a></h2>
<p>In general, users are welcome to use the <code class="docutils literal notranslate"><span class="pre">compile()</span></code> to finish the conversion from a model to tensorRT engine. It is a
wrapper API that consists of the major steps needed to finish this converison. Please refer to an example usage in <code class="docutils literal notranslate"><span class="pre">lower_example.py</span></code> file under <code class="docutils literal notranslate"><span class="pre">examples/fx</span></code>.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>def<span class="w"> </span>compile<span class="o">(</span>
<span class="w"> </span>module:<span class="w"> </span>nn.Module,
<span class="w"> </span>input,
<span class="w"> </span><span class="nv">max_batch_size</span><span class="o">=</span><span class="m">2048</span>,
<span class="w"> </span><span class="nv">max_workspace_size</span><span class="o">=</span><span class="m">33554432</span>,
<span class="w"> </span><span class="nv">explicit_batch_dimension</span><span class="o">=</span>False,
<span class="w"> </span><span class="nv">lower_precision</span><span class="o">=</span>LowerPrecision.FP16,
<span class="w"> </span><span class="nv">verbose_log</span><span class="o">=</span>False,
<span class="w"> </span><span class="nv">timing_cache_prefix</span><span class="o">=</span><span class="s2">""</span>,
<span class="w"> </span><span class="nv">save_timing_cache</span><span class="o">=</span>False,
<span class="w"> </span><span class="nv">cuda_graph_batch_size</span><span class="o">=</span>-1,
<span class="w"> </span><span class="nv">dynamic_batch</span><span class="o">=</span>True,
<span class="o">)</span><span class="w"> </span>-><span class="w"> </span>nn.Module:
<span class="w"> </span><span class="s2">"""</span>
<span class="s2"> Takes in original module, input and lowering setting, run lowering workflow to turn module</span>
<span class="s2"> into lowered module, or so called TRTModule.</span>
<span class="s2"> Args:</span>
<span class="s2"> module: Original module for lowering.</span>
<span class="s2"> input: Input for module.</span>
<span class="s2"> max_batch_size: Maximum batch size (must be >= 1 to be set, 0 means not set)</span>
<span class="s2"> max_workspace_size: Maximum size of workspace given to TensorRT.</span>
<span class="s2"> explicit_batch_dimension: Use explicit batch dimension in TensorRT if set True, otherwise use implicit batch dimension.</span>
<span class="s2"> lower_precision: lower_precision config given to TRTModule.</span>
<span class="s2"> verbose_log: Enable verbose log for TensorRT if set True.</span>
<span class="s2"> timing_cache_prefix: Timing cache file name for timing cache used by fx2trt.</span>
<span class="s2"> save_timing_cache: Update timing cache with current timing cache data if set to True.</span>
<span class="s2"> cuda_graph_batch_size: Cuda graph batch size, default to be -1.</span>
<span class="s2"> dynamic_batch: batch dimension (dim=0) is dynamic.</span>
<span class="s2"> Returns:</span>
<span class="s2"> A torch.nn.Module lowered by TensorRT.</span>
<span class="s2"> """</span>
</pre></div>
</div>
<p>In this section, we will go through an example to illustrate the major steps that fx path uses.
Users can refer to <code class="docutils literal notranslate"><span class="pre">fx2trt_example.py</span></code> file in <code class="docutils literal notranslate"><span class="pre">examples/fx</span></code>.</p>
<ul class="simple">
<li><p><strong>Step 1: Trace the model with acc_tracer</strong></p></li>
</ul>
<p>Acc_tracer is a tracer inheritated from FX tracer. It comes with args normalizer to convert all args to kwargs and pass to TRT converters.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>import<span class="w"> </span>torch_tensorrt.fx.tracer.acc_tracer.acc_tracer<span class="w"> </span>as<span class="w"> </span>acc_tracer
<span class="c1"># Build the model which needs to be a PyTorch nn.Module.</span>
<span class="nv">my_pytorch_model</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>build_model<span class="o">()</span>
<span class="c1"># Prepare inputs to the model. Inputs have to be a List of Tensors</span>
<span class="nv">inputs</span><span class="w"> </span><span class="o">=</span><span class="w"> </span><span class="o">[</span>Tensor,<span class="w"> </span>Tensor,<span class="w"> </span>...<span class="o">]</span>
<span class="c1"># Trace the model with acc_tracer.</span>
<span class="nv">acc_mod</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>acc_tracer.trace<span class="o">(</span>my_pytorch_model,<span class="w"> </span>inputs<span class="o">)</span>
</pre></div>
</div>
<p><em>Common Errors:</em></p>
<p>symbolically traced variables cannot be used as inputs to control flow
This means the model contains dynamic control flow. Please refer to section “Dynamic Control Flow” in <a class="reference external" href="https://pytorch.org/docs/stable/fx.html#dynamic-control-flow">FX guide</a>.</p>
<ul class="simple">
<li><p><strong>Step 2: Build TensorRT engine</strong></p></li>
</ul>
<p>There are <a class="reference external" href="https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#explicit-implicit-batch">two different modes</a> for how TensorRT handles batch dimension, explicit batch dimension and implicit batch dimension. This mode was used by early versions of TensorRT, and is now deprecated but continues to be supported for backwards compatibility. In explicit batch mode, all dimensions are explicit and can be dynamic, that is their length can change at execution time. Many new features, such as dynamic shapes and loops, are available only in this mode. User can still choose to use implicit batch mode when they set <code class="docutils literal notranslate"><span class="pre">explicit_batch_dimension=False</span></code> in <code class="docutils literal notranslate"><span class="pre">compile()</span></code>. We do not recommend to use it since it will lack of support in future TensorRT versions.</p>
<p>Explicit batch is the default mode and it must be set for dynamic shape. For most of vision task, user can choose to enable <code class="docutils literal notranslate"><span class="pre">dynamic_batch</span></code> in <code class="docutils literal notranslate"><span class="pre">compile()</span></code> if they want to get the similar effects as implicit mode where only batch dimension changes. It has some requirements:
1. Shapes of inputs, outputs and activations are fixed except batch dimension.
2. Inputs, outputs and activations have batch dimension as the major dimension.
3. All the operators in the model do not modify batch dimension (permute, transpose, split, etc.) or compute over batch dimension (sum, softmax, etc.).</p>
<p>For examples of the last path, if we have a 3D tensor t shaped as (batch, sequence, dimension), operations such as torch.transpose(0, 2). If any of these three are not satisfied, we’ll need to specify InputTensorSpec as inputs with dynamic range.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>import<span class="w"> </span>deeplearning.trt.fx2trt.converter.converters
from<span class="w"> </span>torch.fx.experimental.fx2trt.fx2trt<span class="w"> </span>import<span class="w"> </span>InputTensorSpec,<span class="w"> </span>TRTInterpreter
<span class="c1"># InputTensorSpec is a dataclass we use to store input information.</span>
<span class="c1"># There're two ways we can build input_specs.</span>
<span class="c1"># Option 1, build it manually.</span>
<span class="nv">input_specs</span><span class="w"> </span><span class="o">=</span><span class="w"> </span><span class="o">[</span>
<span class="w"> </span>InputTensorSpec<span class="o">(</span><span class="nv">shape</span><span class="o">=(</span><span class="m">1</span>,<span class="w"> </span><span class="m">2</span>,<span class="w"> </span><span class="m">3</span><span class="o">)</span>,<span class="w"> </span><span class="nv">dtype</span><span class="o">=</span>torch.float32<span class="o">)</span>,
<span class="w"> </span>InputTensorSpec<span class="o">(</span><span class="nv">shape</span><span class="o">=(</span><span class="m">1</span>,<span class="w"> </span><span class="m">4</span>,<span class="w"> </span><span class="m">5</span><span class="o">)</span>,<span class="w"> </span><span class="nv">dtype</span><span class="o">=</span>torch.float32<span class="o">)</span>,
<span class="o">]</span>
<span class="c1"># Option 2, build it using sample_inputs where user provide a sample</span>
<span class="nv">inputs</span><span class="w"> </span><span class="o">=</span><span class="w"> </span><span class="o">[</span>
torch.rand<span class="o">((</span><span class="m">1</span>,2,3<span class="o">)</span>,<span class="w"> </span><span class="nv">dtype</span><span class="o">=</span>torch.float32<span class="o">)</span>,
torch.rand<span class="o">((</span><span class="m">1</span>,4,5<span class="o">)</span>,<span class="w"> </span><span class="nv">dtype</span><span class="o">=</span>torch.float32<span class="o">)</span>,
<span class="o">]</span>
<span class="nv">input_specs</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>InputTensorSpec.from_tensors<span class="o">(</span>inputs<span class="o">)</span>
<span class="c1"># IMPORTANT: If dynamic shape is needed, we need to build it slightly differently.</span>
<span class="nv">input_specs</span><span class="w"> </span><span class="o">=</span><span class="w"> </span><span class="o">[</span>
<span class="w"> </span>InputTensorSpec<span class="o">(</span>
<span class="w"> </span><span class="nv">shape</span><span class="o">=(</span>-1,<span class="w"> </span><span class="m">2</span>,<span class="w"> </span><span class="m">3</span><span class="o">)</span>,
<span class="w"> </span><span class="nv">dtype</span><span class="o">=</span>torch.float32,
<span class="w"> </span><span class="c1"># Currently we only support one set of dynamic range. User may set other dimensions but it is not promised to work for any models</span>
<span class="w"> </span><span class="c1"># (min_shape, optimize_target_shape, max_shape)</span>
<span class="w"> </span><span class="c1"># For more information refer to fx/input_tensor_spec.py</span>
<span class="w"> </span><span class="nv">shape_ranges</span><span class="w"> </span><span class="o">=</span><span class="w"> </span><span class="o">[</span>
<span class="w"> </span><span class="o">((</span><span class="m">1</span>,<span class="w"> </span><span class="m">2</span>,<span class="w"> </span><span class="m">3</span><span class="o">)</span>,<span class="w"> </span><span class="o">(</span><span class="m">4</span>,<span class="w"> </span><span class="m">2</span>,<span class="w"> </span><span class="m">3</span><span class="o">)</span>,<span class="w"> </span><span class="o">(</span><span class="m">100</span>,<span class="w"> </span><span class="m">2</span>,<span class="w"> </span><span class="m">3</span><span class="o">))</span>,
<span class="w"> </span><span class="o">]</span>,
<span class="w"> </span><span class="o">)</span>,
<span class="w"> </span>InputTensorSpec<span class="o">(</span><span class="nv">shape</span><span class="o">=(</span><span class="m">1</span>,<span class="w"> </span><span class="m">4</span>,<span class="w"> </span><span class="m">5</span><span class="o">)</span>,<span class="w"> </span><span class="nv">dtype</span><span class="o">=</span>torch.float32<span class="o">)</span>,
<span class="o">]</span>
<span class="c1"># Build a TRT interpreter. Set explicit_batch_dimension accordingly.</span>
<span class="nv">interpreter</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>TRTInterpreter<span class="o">(</span>
<span class="w"> </span>acc_mod,<span class="w"> </span>input_specs,<span class="w"> </span><span class="nv">explicit_batch_dimension</span><span class="o">=</span>True/False
<span class="o">)</span>
<span class="c1"># The output of TRTInterpreter run() is wrapped as TRTInterpreterResult.</span>
<span class="c1"># The TRTInterpreterResult contains required parameter to build TRTModule,</span>
<span class="c1"># and other informational output from TRTInterpreter run.</span>
class<span class="w"> </span>TRTInterpreterResult<span class="o">(</span>NamedTuple<span class="o">)</span>:
<span class="w"> </span>engine:<span class="w"> </span>Any
<span class="w"> </span>input_names:<span class="w"> </span>Sequence<span class="o">[</span>str<span class="o">]</span>
<span class="w"> </span>output_names:<span class="w"> </span>Sequence<span class="o">[</span>str<span class="o">]</span>
<span class="w"> </span>serialized_cache:<span class="w"> </span>bytearray
<span class="c1">#max_batch_size: set accordingly for maximum batch size you will use.</span>
<span class="c1">#max_workspace_size: set to the maximum size we can afford for temporary buffer</span>
<span class="c1">#lower_precision: the precision model layers are running on (TensorRT will choose the best perforamnce precision).</span>
<span class="c1">#sparse_weights: allow the builder to examine weights and use optimized functions when weights have suitable sparsity</span>
<span class="c1">#force_fp32_output: force output to be fp32</span>
<span class="c1">#strict_type_constraints: Usually we should set it to False unless we want to control the precision of certain layer for numeric #reasons.</span>
<span class="c1">#algorithm_selector: set up algorithm selection for certain layer</span>
<span class="c1">#timing_cache: enable timing cache for TensorRT</span>
<span class="c1">#profiling_verbosity: TensorRT logging level</span>
<span class="nv">trt_interpreter_result</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>interpreter.run<span class="o">(</span>
<span class="w"> </span><span class="nv">max_batch_size</span><span class="o">=</span><span class="m">64</span>,
<span class="w"> </span><span class="nv">max_workspace_size</span><span class="o">=</span><span class="m">1</span><span class="w"> </span><span class="s"><< 25,</span>
<span class="s"> sparse_weights=False,</span>
<span class="s"> force_fp32</span>_output<span class="o">=</span>False,
<span class="w"> </span><span class="nv">strict_type_constraints</span><span class="o">=</span>False,
<span class="w"> </span><span class="nv">algorithm_selector</span><span class="o">=</span>None,
<span class="w"> </span><span class="nv">timing_cache</span><span class="o">=</span>None,
<span class="w"> </span><span class="nv">profiling_verbosity</span><span class="o">=</span>None,
<span class="o">)</span>
</pre></div>
</div>
<p><em>Common Errors:</em></p>
<p>RuntimeError: Conversion of function xxx not currently supported!
- This means we don’t have the support for this xxx operator. Please refer to section “How to add a missing op” below for further instructions.</p>
<ul class="simple">
<li><p><strong>Step 3: Run the model</strong></p></li>
</ul>
<p>One way is using TRTModule, which is basically a PyTorch nn.Module.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>from<span class="w"> </span>torch_tensorrt.fx<span class="w"> </span>import<span class="w"> </span>TRTModule
<span class="nv">mod</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>TRTModule<span class="o">(</span>
<span class="w"> </span>trt_interpreter_result.engine,
<span class="w"> </span>trt_interpreter_result.input_names,
<span class="w"> </span>trt_interpreter_result.output_names<span class="o">)</span>
<span class="c1"># Just like all other PyTorch modules</span>
<span class="nv">outputs</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>mod<span class="o">(</span>*inputs<span class="o">)</span>
torch.save<span class="o">(</span>mod,<span class="w"> </span><span class="s2">"trt.pt"</span><span class="o">)</span>
<span class="nv">reload_trt_mod</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>torch.load<span class="o">(</span><span class="s2">"trt.pt"</span><span class="o">)</span>
<span class="nv">reload_model_output</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>reload_trt_mod<span class="o">(</span>*inputs<span class="o">)</span>
</pre></div>
</div>
<p>So far, we give a detailed explanation of major steps in convterting a PyTorch model into TensorRT engine. Users are welcome to refer to the source code for some parameters explanations. In the converting scheme, there are two important actions in it. One is acc tracer which helps us to convert a PyTorch model to acc graph. The other is FX path converter which helps to convert the acc graph’s operation to corresponding TensorRT operation and build up the TensoRT engine for it.</p>
</section>
<section id="acc-tracer">
<h2>Acc Tracer<a class="headerlink" href="#acc-tracer" title="Permalink to this headline">¶</a></h2>
<p>Acc tracer is a custom FX symbolic tracer. It does a couple more things compare to the vanilla FX symbolic tracer. We mainly depend on it to convert PyTorch ops or builtin ops to acc ops. There are two main purposes for fx2trt to use acc ops:</p>
<ol class="arabic simple">
<li><p>there’re many ops that do similar things in PyTorch ops and builtin ops such like torch.add, builtin.add and torch.Tensor.add. Using acc tracer, we normalize these three ops to a single acc_ops.add. This helps reduce the number of converters we need to write.</p></li>
<li><p>acc ops only have kwargs which makes writing converter easier as we don’t need to add additional logic to find arguments in args and kwargs.</p></li>
</ol>
</section>
<section id="fx2trt">
<h2>FX2TRT<a class="headerlink" href="#fx2trt" title="Permalink to this headline">¶</a></h2>
<p>After symbolic tracing, we have the graph representation of a PyTorch model. fx2trt leverages the power of fx.Interpreter. fx.Interpreter goes through the whole graph node by node and calls the function that node represents. fx2trt overrides the original behavior of calling the function with invoking corresponding converts for each node. Each converter function adds corresponding TensorRT layer(s).</p>
<p>Below is an example of a converter function. The decorator is used to register this converter function with the corresponding node. In this example, we register this converter to a fx node whose target is acc_ops.sigmoid.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>@tensorrt_converter<span class="o">(</span>acc_ops.sigmoid<span class="o">)</span>
def<span class="w"> </span>acc_ops_sigmoid<span class="o">(</span>network,<span class="w"> </span>target,<span class="w"> </span>args,<span class="w"> </span>kwargs,<span class="w"> </span>name<span class="o">)</span>:
<span class="w"> </span><span class="s2">"""</span>
<span class="s2"> network: TensorRT network. We'll be adding layers to it.</span>
<span class="s2"> The rest arguments are attributes of fx node.</span>
<span class="s2"> """</span>
<span class="w"> </span><span class="nv">input_val</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>kwargs<span class="o">[</span><span class="s1">'input'</span><span class="o">]</span>
<span class="w"> </span><span class="k">if</span><span class="w"> </span>not<span class="w"> </span>isinstance<span class="o">(</span>input_val,<span class="w"> </span>trt.tensorrt.ITensor<span class="o">)</span>:
<span class="w"> </span>raise<span class="w"> </span>RuntimeError<span class="o">(</span>f<span class="s1">'Sigmoid received input {input_val} that is not part '</span>
<span class="w"> </span><span class="s1">'of the TensorRT region!'</span><span class="o">)</span>
<span class="w"> </span><span class="nv">layer</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>network.add_activation<span class="o">(</span><span class="nv">input</span><span class="o">=</span>input_val,<span class="w"> </span><span class="nv">type</span><span class="o">=</span>trt.ActivationType.SIGMOID<span class="o">)</span>
<span class="w"> </span>layer.name<span class="w"> </span><span class="o">=</span><span class="w"> </span>name
<span class="w"> </span><span class="k">return</span><span class="w"> </span>layer.get_output<span class="o">(</span><span class="m">0</span><span class="o">)</span>
</pre></div>
</div>
<section id="how-to-add-a-missing-op">
<h3>How to Add a Missing Op<a class="headerlink" href="#how-to-add-a-missing-op" title="Permalink to this headline">¶</a></h3>
<p>You can actually add it wherever you want just need to remember import the file so that all acc ops and mapper will be registered before tracing with acc_tracer.</p>
<ul class="simple">
<li><p><strong>Step 1. Add a new acc op</strong></p></li>
</ul>
<p>TODO: Need to explain more on the logistic of acc op like when we want to break down an op and when we want to reuse other ops.</p>
<p>In <a class="reference external" href="https://github.com/pytorch/TensorRT/blob/master/py/torch_tensorrt/fx/tracer/acc_tracer/acc_tracer.py">acc tracer</a>, we convert nodes in the graph to acc ops if there’s a mapping registered for the node to an acc op.</p>
<p>In order to make the conversion to acc ops to happen, there’re two things required. One is that there should be an acc op function defined and the other is there should be a mapping registered.</p>
<p>Defining an acc op is simple, we first just need a function and register the function as an acc op via this decorator <a class="reference external" href="https://github.com/pytorch/TensorRT/blob/master/py/torch_tensorrt/fx/tracer/acc_tracer/acc_normalizer.py">acc_normalizer.py</a>. e.g. the following code adds an acc op named foo() which adds two given inputs.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span><span class="c1"># NOTE: all acc ops should only take kwargs as inputs, therefore we need the "*"</span>
<span class="c1"># at the beginning.</span>
@register_acc_op
def<span class="w"> </span>foo<span class="o">(</span>*,<span class="w"> </span>input,<span class="w"> </span>other,<span class="w"> </span>alpha<span class="o">)</span>:
<span class="w"> </span><span class="k">return</span><span class="w"> </span>input<span class="w"> </span>+<span class="w"> </span>alpha<span class="w"> </span>*<span class="w"> </span>other
</pre></div>
</div>
<p>There’re two ways to register a mapping. One is <a class="reference external" href="https://github.com/pytorch/TensorRT/blob/1a22204fecec690bc3c2a318dab4f57b98c57f05/py/torch_tensorrt/fx/tracer/acc_tracer/acc_normalizer.py#L164">register_acc_op_mapping()</a>. Let’s register a mapping from torch.add to foo() we just created above. We need to add decorator register_acc_op_mapping to it.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span><span class="nv">this_arg_is_optional</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>True
@register_acc_op_mapping<span class="o">(</span>
<span class="w"> </span><span class="nv">op_and_target</span><span class="o">=(</span><span class="s2">"call_function"</span>,<span class="w"> </span>torch.add<span class="o">)</span>,
<span class="w"> </span><span class="nv">arg_replacement_tuples</span><span class="o">=[</span>
<span class="w"> </span><span class="o">(</span><span class="s2">"input"</span>,<span class="w"> </span><span class="s2">"input"</span><span class="o">)</span>,
<span class="w"> </span><span class="o">(</span><span class="s2">"other"</span>,<span class="w"> </span><span class="s2">"other"</span><span class="o">)</span>,
<span class="w"> </span><span class="o">(</span><span class="s2">"alpha"</span>,<span class="w"> </span><span class="s2">"alpha"</span>,<span class="w"> </span>this_arg_is_optional<span class="o">)</span>,
<span class="w"> </span><span class="o">]</span>,
<span class="o">)</span>
@register_acc_op
def<span class="w"> </span>foo<span class="o">(</span>*,<span class="w"> </span>input,<span class="w"> </span>other,<span class="w"> </span><span class="nv">alpha</span><span class="o">=</span><span class="m">1</span>.0<span class="o">)</span>:
<span class="w"> </span><span class="k">return</span><span class="w"> </span>input<span class="w"> </span>+<span class="w"> </span>alpha<span class="w"> </span>*<span class="w"> </span>other
</pre></div>
</div>
<p><code class="docutils literal notranslate"><span class="pre">op_and_target</span></code> determines which node will trigger this mapping. op and target are the attributes of FX node. In acc_normalization when we see a node with the same op and target as set in the <code class="docutils literal notranslate"><span class="pre">op_and_target</span></code>, we will trigger the mapping. Since we want to map from <code class="docutils literal notranslate"><span class="pre">torch.add</span></code>, then op would be call_function and target would be <code class="docutils literal notranslate"><span class="pre">torch.add</span></code>. <code class="docutils literal notranslate"><span class="pre">arg_replacement_tuples</span></code> determines how we construct kwargs for new acc op node using args and kwargs from original node. Each tuple in <code class="docutils literal notranslate"><span class="pre">arg_replacement_tuples</span></code> represents one argument mapping rule. It contains two or three elements. The third element is a boolean variable that determines whether this kwarg is optional in <em>original node</em>. We only need to specify the third element if it’s True. The first element is the argument name in original node which will be used as the acc op node’s argument whose name is the second element in the tuple. The sequence of the tuples does matter because the position of the tuple determines where the argument is in original node’s args. We use this information to map args from original node to kwargs in acc op node.
We don’t have to specify arg_replacement_tuples if none of the followings are true.</p>
<ol class="arabic simple">
<li><p>kwargs of original nodes and acc op nodes have different name.</p></li>
<li><p>there’re optional arguments.</p></li>
</ol>
<p>The other way to register a mapping is through <a class="reference external" href="https://github.com/pytorch/TensorRT/blob/1a22204fecec690bc3c2a318dab4f57b98c57f05/py/torch_tensorrt/fx/tracer/acc_tracer/acc_normalizer.py#L206">register_custom_acc_mapper_fn()</a>. This one is designed to reduce the redundant op registration as it allows you to use a function to map to one or more existing acc ops throught some combinations. In the function, you can do basically whatever you want. Let’s use an example to explain how it works.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>@register_acc_op
def<span class="w"> </span>foo<span class="o">(</span>*,<span class="w"> </span>input,<span class="w"> </span>other,<span class="w"> </span><span class="nv">alpha</span><span class="o">=</span><span class="m">1</span>.0<span class="o">)</span>:
<span class="w"> </span><span class="k">return</span><span class="w"> </span>input<span class="w"> </span>+<span class="w"> </span>alpha<span class="w"> </span>*<span class="w"> </span>other
@register_custom_acc_mapper_fn<span class="o">(</span>
<span class="w"> </span><span class="nv">op_and_target</span><span class="o">=(</span><span class="s2">"call_function"</span>,<span class="w"> </span>torch.add<span class="o">)</span>,
<span class="w"> </span><span class="nv">arg_replacement_tuples</span><span class="o">=[</span>
<span class="w"> </span><span class="o">(</span><span class="s2">"input"</span>,<span class="w"> </span><span class="s2">"input"</span><span class="o">)</span>,
<span class="w"> </span><span class="o">(</span><span class="s2">"other"</span>,<span class="w"> </span><span class="s2">"other"</span><span class="o">)</span>,
<span class="w"> </span><span class="o">(</span><span class="s2">"alpha"</span>,<span class="w"> </span><span class="s2">"alpha"</span>,<span class="w"> </span>this_arg_is_optional<span class="o">)</span>,
<span class="w"> </span><span class="o">]</span>,
<span class="o">)</span>
def<span class="w"> </span>custom_mapper<span class="o">(</span>node:<span class="w"> </span>torch.fx.Node,<span class="w"> </span>_:<span class="w"> </span>nn.Module<span class="o">)</span><span class="w"> </span>-><span class="w"> </span>torch.fx.Node:
<span class="w"> </span><span class="s2">"""</span>
<span class="s2"> `node` is original node, which is a call_function node with target</span>
<span class="s2"> being torch.add.</span>
<span class="s2"> """</span>
<span class="w"> </span><span class="nv">alpha</span><span class="w"> </span><span class="o">=</span><span class="w"> </span><span class="m">1</span>
<span class="w"> </span><span class="k">if</span><span class="w"> </span><span class="s2">"alpha"</span><span class="w"> </span><span class="k">in</span><span class="w"> </span>node.kwargs:
<span class="w"> </span><span class="nv">alpha</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>node.kwargs<span class="o">[</span><span class="s2">"alpha"</span><span class="o">]</span>
<span class="w"> </span><span class="nv">foo_kwargs</span><span class="w"> </span><span class="o">=</span><span class="w"> </span><span class="o">{</span><span class="s2">"input"</span>:<span class="w"> </span>node<span class="o">[</span><span class="s2">"input"</span><span class="o">]</span>,<span class="w"> </span><span class="s2">"other"</span>:<span class="w"> </span>node<span class="o">[</span><span class="s2">"other"</span><span class="o">]</span>,<span class="w"> </span><span class="s2">"alpha"</span>:<span class="w"> </span>alpha<span class="o">}</span>
<span class="w"> </span>with<span class="w"> </span>node.graph.inserting_before<span class="o">(</span>node<span class="o">)</span>:
<span class="w"> </span><span class="nv">foo_node</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>node.graph.call_function<span class="o">(</span>foo,<span class="w"> </span><span class="nv">kwargs</span><span class="o">=</span>foo_kwargs<span class="o">)</span>
<span class="w"> </span>foo_node.meta<span class="w"> </span><span class="o">=</span><span class="w"> </span>node.meta.copy<span class="o">()</span>
<span class="w"> </span><span class="k">return</span><span class="w"> </span>foo_node
</pre></div>
</div>
<p>In the custom mapper function, we construct an acc op node and return it. The node we returns here would take over all the children nodes of original nodes <a class="reference external" href="https://github.com/pytorch/TensorRT/blob/1a22204fecec690bc3c2a318dab4f57b98c57f05/py/torch_tensorrt/fx/tracer/acc_tracer/acc_normalizer.py#L379">acc_normalizer.py</a>.</p>
<p>The last step would be <em>adding unit test</em> for the new acc op or mapper function we added. The place to add the unit test is here <a class="reference external" href="https://github.com/pytorch/TensorRT/blob/master/py/torch_tensorrt/fx/test/tracer/test_acc_tracer.py">test_acc_tracer.py</a>.</p>
<ul class="simple">
<li><p><strong>Step 2. Add a new converter</strong></p></li>
</ul>
<p>All the developed converters for acc ops are all in <a class="reference external" href="https://github.com/pytorch/TensorRT/blob/master/py/torch_tensorrt/fx/converters/acc_ops_converters.py">acc_op_converter.py</a>. It could give you a good example of how the converter is added.</p>
<p>Essentially, the converter is the mapping mechanism that maps the acc ops to a TensorRT layer. If we are able to find all the TensorRT layers we need we can get start to add a converter for the node using <a class="reference external" href="https://docs.nvidia.com/deeplearning/tensorrt/api/python_api/infer/Graph/Network.html">TensorRT APIs</a>.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>@tensorrt_converter<span class="o">(</span>acc_ops.sigmoid<span class="o">)</span>
def<span class="w"> </span>acc_ops_sigmoid<span class="o">(</span>network,<span class="w"> </span>target,<span class="w"> </span>args,<span class="w"> </span>kwargs,<span class="w"> </span>name<span class="o">)</span>:
<span class="w"> </span><span class="s2">"""</span>
<span class="s2"> network: TensorRT network. We'll be adding layers to it.</span>
<span class="s2"> The rest arguments are attributes of fx node.</span>
<span class="s2"> """</span>
<span class="w"> </span><span class="nv">input_val</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>kwargs<span class="o">[</span><span class="s1">'input'</span><span class="o">]</span>
<span class="w"> </span><span class="k">if</span><span class="w"> </span>not<span class="w"> </span>isinstance<span class="o">(</span>input_val,<span class="w"> </span>trt.tensorrt.ITensor<span class="o">)</span>:
<span class="w"> </span>raise<span class="w"> </span>RuntimeError<span class="o">(</span>f<span class="s1">'Sigmoid received input {input_val} that is not part '</span>
<span class="w"> </span><span class="s1">'of the TensorRT region!'</span><span class="o">)</span>
<span class="w"> </span><span class="nv">layer</span><span class="w"> </span><span class="o">=</span><span class="w"> </span>network.add_activation<span class="o">(</span><span class="nv">input</span><span class="o">=</span>input_val,<span class="w"> </span><span class="nv">type</span><span class="o">=</span>trt.ActivationType.SIGMOID<span class="o">)</span>
<span class="w"> </span>layer.name<span class="w"> </span><span class="o">=</span><span class="w"> </span>name
<span class="w"> </span><span class="k">return</span><span class="w"> </span>layer.get_output<span class="o">(</span><span class="m">0</span><span class="o">)</span>
</pre></div>
</div>
<p>We need to use <code class="docutils literal notranslate"><span class="pre">tensorrt_converter</span></code> decorator to register the converter. The argument for the decorator is the target of the fx node that we need to convert. In the converter, we can find the inputs to the fx node in kwargs. As in the example, the original node is <cite>acc_ops.sigmoid</cite> which only has one argument “input” in acc_ops.py. We get the input and check if it’s a TensorRT tensor. After that, we add a sigmoid layer to TensorRT network and return the output of the layer. The output we returned will be passed to the children nodes of acc_ops.sigmoid by fx.Interpreter.</p>
<p><strong>What if we can not find corresponding layers in TensorRT that do the same thing as the node.</strong></p>
<p>In this case, we would need to do a bit more work. TensorRT provides plugins which serves as custom layers. <em>We have not implement this feature yet. We will update once it is enabled</em>.</p>
<p>Last step would be adding the unit test for the new converter we added. User could add corresponding unit test in this <a class="reference external" href="https://github.com/pytorch/TensorRT/tree/master/py/torch_tensorrt/fx/test/converters/acc_op">folder</a>.</p>
</section>
</section>
</section>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="torch_compile.html" class="btn btn-neutral float-right" title="Torch-TensorRT torch.compile Backend" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="creating_torchscript_module_in_python.html" class="btn btn-neutral" title="Creating a TorchScript Module" accesskey="p" rel="prev"><img src="../_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright 2022, NVIDIA Corporation.
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
</div>
<div class="pytorch-content-right" id="pytorch-content-right">
<div class="pytorch-right-menu" id="pytorch-right-menu">
<div class="pytorch-side-scroll" id="pytorch-side-scroll-right">
<ul>
<li><a class="reference internal" href="#">Torch-TensorRT (FX Frontend) User Guide</a><ul>
<li><a class="reference internal" href="#converting-a-pytorch-model-to-tensorrt-engine">Converting a PyTorch Model to TensorRT Engine</a></li>
<li><a class="reference internal" href="#acc-tracer">Acc Tracer</a></li>
<li><a class="reference internal" href="#fx2trt">FX2TRT</a><ul>
<li><a class="reference internal" href="#how-to-add-a-missing-op">How to Add a Missing Op</a></li>
</ul>
</li>
</ul>
</li>
</ul>
</div>
</div>
</div>
</section>
</div>
<script type="text/javascript" id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script data-url_root="../" id="documentation_options" src="../_static/documentation_options.js"></script>
<script src="../_static/jquery.js"></script>
<script src="../_static/underscore.js"></script>
<script src="../_static/doctools.js"></script>
<script src="../_static/collapsible-lists/js/CollapsibleLists.compressed.js"></script>
<script src="../_static/collapsible-lists/js/apply-collapsible-lists.js"></script>
<script crossorigin="anonymous" integrity="sha256-Ae2Vz/4ePdIu6ZyI/5ZGsYnb+m0JlOmKPjt6XZ9JJkA=" src="https://cdnjs.cloudflare.com/ajax/libs/require.js/2.3.4/require.min.js"></script>
<script type="text/javascript" src="../_static/js/vendor/popper.min.js"></script>
<script type="text/javascript" src="../_static/js/vendor/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/list.js/1.5.0/list.min.js"></script>
<script type="text/javascript" src="../_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<!-- Begin Footer -->
<div class="container-fluid docs-tutorials-resources" id="docs-tutorials-resources">
<div class="container">
<div class="row">
<div class="col-md-4 text-center">
<h2>Docs</h2>
<p>Access comprehensive developer documentation for PyTorch</p>
<a class="with-right-arrow" href="https://pytorch.org/docs/stable/index.html">View Docs</a>
</div>
<div class="col-md-4 text-center">
<h2>Tutorials</h2>
<p>Get in-depth tutorials for beginners and advanced developers</p>
<a class="with-right-arrow" href="https://pytorch.org/tutorials">View Tutorials</a>
</div>
<div class="col-md-4 text-center">
<h2>Resources</h2>
<p>Find development resources and get your questions answered</p>
<a class="with-right-arrow" href="https://pytorch.org/resources">View Resources</a>
</div>
</div>
</div>
</div>
<footer class="site-footer">
<div class="container footer-container">
<div class="footer-logo-wrapper">
<a href="https://pytorch.org/" class="footer-logo"></a>
</div>
<div class="footer-links-wrapper">
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/">PyTorch</a></li>
<li><a href="https://pytorch.org/get-started">Get Started</a></li>
<li><a href="https://pytorch.org/features">Features</a></li>
<li><a href="https://pytorch.org/ecosystem">Ecosystem</a></li>
<li><a href="https://pytorch.org/blog/">Blog</a></li>
<li><a href="https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md">Contributing</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/resources">Resources</a></li>
<li><a href="https://pytorch.org/tutorials">Tutorials</a></li>
<li><a href="https://pytorch.org/docs/stable/index.html">Docs</a></li>
<li><a href="https://discuss.pytorch.org" target="_blank">Discuss</a></li>
<li><a href="https://github.com/pytorch/pytorch/issues" target="_blank">Github Issues</a></li>
<li><a href="https://pytorch.org/assets/brand-guidelines/PyTorch-Brand-Guidelines.pdf" target="_blank">Brand Guidelines</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title">Stay up to date</li>
<li><a href="https://www.facebook.com/pytorch" target="_blank">Facebook</a></li>
<li><a href="https://twitter.com/pytorch" target="_blank">Twitter</a></li>
<li><a href="https://www.youtube.com/pytorch" target="_blank">YouTube</a></li>
<li><a href="https://www.linkedin.com/company/pytorch" target="_blank">LinkedIn</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title">PyTorch Podcasts</li>
<li><a href="https://open.spotify.com/show/6UzHKeiy368jKfQMKKvJY5" target="_blank">Spotify</a></li>
<li><a href="https://podcasts.apple.com/us/podcast/pytorch-developer-podcast/id1566080008" target="_blank">Apple</a></li>
<li><a href="https://www.google.com/podcasts?feed=aHR0cHM6Ly9mZWVkcy5zaW1wbGVjYXN0LmNvbS9PQjVGa0lsOA%3D%3D" target="_blank">Google</a></li>
<li><a href="https://music.amazon.com/podcasts/7a4e6f0e-26c2-49e9-a478-41bd244197d0/PyTorch-Developer-Podcast?" target="_blank">Amazon</a></li>
</ul>
</div>
</div>
<div class="privacy-policy">
<ul>
<li class="privacy-policy-links"><a href="https://www.linuxfoundation.org/terms/" target="_blank">Terms</a></li>
<li class="privacy-policy-links">|</li>
<li class="privacy-policy-links"><a href="https://www.linuxfoundation.org/privacy-policy/" target="_blank">Privacy</a></li>
</ul>
</div>
<div class="copyright">
<p>© Copyright The Linux Foundation. The PyTorch Foundation is a project of The Linux Foundation.
For web site terms of use, trademark policy and other policies applicable to The PyTorch Foundation please see
<a href="https://www.linuxfoundation.org/policies/">www.linuxfoundation.org/policies/</a>. The PyTorch Foundation supports the PyTorch open source
project, which has been established as PyTorch Project a Series of LF Projects, LLC. For policies applicable to the PyTorch Project a Series of LF Projects, LLC,
please see <a href="https://www.lfprojects.org/policies/">www.lfprojects.org/policies/</a>.</p>
</div>
</div>
</footer>
<div class="cookie-banner-wrapper">
<div class="container">
<p class="gdpr-notice">To analyze traffic and optimize your experience, we serve cookies on this site. By clicking or navigating, you agree to allow our usage of cookies. As the current maintainers of this site, Facebook’s Cookies Policy applies. Learn more, including about available controls: <a href="https://www.facebook.com/policies/cookies/">Cookies Policy</a>.</p>
<img class="close-button" src="../_static/images/pytorch-x.svg">
</div>
</div>
<!-- End Footer -->
<!-- Begin Mobile Menu -->
<div class="mobile-main-menu">
<div class="container-fluid">
<div class="container">
<div class="mobile-main-menu-header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<a class="main-menu-close-button" href="#" data-behavior="close-mobile-menu"></a>
</div>
</div>
</div>
<div class="mobile-main-menu-links-container">
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<a href="">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="resources-mobile-menu-title">
Docs
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/docs/stable/index.html">PyTorch</a>
</li>
<li>
<a href="https://pytorch.org/audio/stable/index.html">torchaudio</a>
</li>
<li>
<a href="https://pytorch.org/text/stable/index.html">torchtext</a>
</li>
<li>
<a href="https://pytorch.org/vision/stable/index.html">torchvision</a>
</li>
<li>
<a href="https://pytorch.org/torcharrow">torcharrow</a>
</li>
<li>
<a href="https://pytorch.org/data">TorchData</a>
</li>
<li>
<a href="https://pytorch.org/torchrec">TorchRec</a>
</li>
<li>
<a href="https://pytorch.org/serve/">TorchServe</a>
</li>
<li>
<a href="https://pytorch.org/torchx/">TorchX</a>
</li>
<li>
<a href="https://pytorch.org/xla">PyTorch on XLA Devices</a>
</li>
</ul>
<li class="resources-mobile-menu-title">
Resources
</li>
<ul class="resources-mobile-menu-items">
<li>
<a href="https://pytorch.org/features">About</a>
</li>
<li>
<a href="https://pytorch.org/foundation">PyTorch Foundation</a>
</li>
<li>
<a href="https://pytorch.org/#community-module">Community</a>
</li>
<li>
<a href="https://pytorch.org/community-stories">Community Stories</a>
</li>
<li>