-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathnested.html
1730 lines (1547 loc) · 156 KB
/
nested.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="viewport" content="width=device-width, initial-scale=1.0">
<title>torch.nested — PyTorch 2.7 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/nested.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="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.Size" href="size.html" />
<link rel="prev" title="torch.masked" href="masked.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-2024">
<span class="dropdown-title">Contributor Awards - 2024</span>
<p>Award winners announced at this year's PyTorch Conference</p>
</a>
</div>
</div>
</li>
<li>
<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>
<a class="nav-dropdown-item" href="https://pytorch.org/executorch/stable/index.html">
<span class="dropdown-title">ExecuTorch Docs</span>
</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>
<a class="nav-dropdown-item" href="https://pytorch.org/newsletter">
<span class="dropdown-title">Newsletter</span>
<p>Stay up-to-date with the latest updates</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>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/credits">
<span class="dropdown-title">Cloud Credit Program</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/tac">
<span class="dropdown-title">Technical Advisory Council</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/staff">
<span class="dropdown-title">Staff</span>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/contact-us">
<span class="dropdown-title">Contact Us</span>
</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'>2.7 ▼</a>
</div>
<div id="searchBox">
<div class="searchbox" id="googleSearchBox">
<script async src="https://cse.google.com/cse.js?cx=e65585f8c3ea1440e"></script>
<div class="gcse-search"></div>
</div>
<div id="sphinxSearchBox" style="display: none;">
<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>
</div>
<form id="searchForm">
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="google" checked>
Google Search
</label>
<label style="margin-bottom: 1rem">
<input type="radio" name="searchType" value="sphinx">
Classic Search
</label>
</form>
<script>
document.addEventListener('DOMContentLoaded', function() {
const searchForm = document.getElementById('searchForm');
const googleSearchBox = document.getElementById('googleSearchBox');
const sphinxSearchBox = document.getElementById('sphinxSearchBox');
// Function to toggle search box visibility
function toggleSearchBox(searchType) {
googleSearchBox.style.display = searchType === 'google' ? 'block' : 'none';
sphinxSearchBox.style.display = searchType === 'sphinx' ? 'block' : 'none';
}
// Determine the default search type
let defaultSearchType;
const currentUrl = window.location.href;
if (currentUrl.startsWith('https://pytorch.org/docs/stable')) {
// For the stable documentation, default to Google
defaultSearchType = localStorage.getItem('searchType') || 'google';
} else {
// For any other version, including docs-preview, default to Sphinx
defaultSearchType = 'sphinx';
}
// Set the default search type
document.querySelector(`input[name="searchType"][value="${defaultSearchType}"]`).checked = true;
toggleSearchBox(defaultSearchType);
// Event listener for changes in search type
searchForm.addEventListener('change', function(event) {
const selectedSearchType = event.target.value;
localStorage.setItem('searchType', selectedSearchType);
toggleSearchBox(selectedSearchType);
});
// Set placeholder text for Google search box
window.onload = function() {
var placeholderText = "Search Docs";
var googleSearchboxText = document.querySelector("#gsc-i-id1");
if (googleSearchboxText) {
googleSearchboxText.placeholder = placeholderText;
googleSearchboxText.style.fontFamily = 'FreightSans';
googleSearchboxText.style.fontSize = "1.2rem";
googleSearchboxText.style.color = '#262626';
}
};
});
</script>
</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/libtorch_stable_abi.html">LibTorch Stable ABI</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="accelerator.html">torch.accelerator</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="mtia.memory.html">torch.mtia.memory</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.fsdp.fully_shard.html">torch.distributed.fsdp.fully_shard</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"><a class="reference internal" href="fx.html">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 current"><a class="current reference internal" href="#">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>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/ao">torchao</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.nested</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/nested.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="module-torch.nested">
<span id="torch-nested"></span><h1>torch.nested<a class="headerlink" href="#module-torch.nested" title="Permalink to this heading">¶</a></h1>
<div class="section" id="introduction">
<h2>Introduction<a class="headerlink" href="#introduction" title="Permalink to this heading">¶</a></h2>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>The PyTorch API of nested tensors is in prototype stage and will change in the near future.</p>
</div>
<p>Nested tensors allow for ragged-shaped data to be contained within and operated upon as a
single tensor. Such data is stored underneath in an efficient packed representation, while exposing
a standard PyTorch tensor interface for applying operations.</p>
<p>A common application of nested tensors is for expressing batches of variable-length sequential data
present in various domains, such as varying sentence lengths, image sizes, and audio / video clip
lengths. Traditionally, such data has been handled by padding sequences to that of the max length
within a batch, performing computation on the padded form, and subsequently masking to remove
padding. This is inefficient and error-prone, and nested tensors exist to address these problems.</p>
<p>The API for calling operations on a nested tensor is no different from that of a regular
<code class="docutils literal notranslate"><span class="pre">torch.Tensor</span></code>, allowing for seamless integration with existing models, with the main
difference being <a class="reference internal" href="#construction"><span class="std std-ref">construction of the inputs</span></a>.</p>
<p>As this is a prototype feature, the set of <a class="reference internal" href="#supported-operations"><span class="std std-ref">operations supported</span></a> is
limited, but growing. We welcome issues, feature requests, and contributions.
More information on contributing can be found
<a class="reference external" href="https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/nested/README.md">in this Readme</a>.</p>
</div>
<div class="section" id="construction">
<span id="id1"></span><h2>Construction<a class="headerlink" href="#construction" title="Permalink to this heading">¶</a></h2>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>There are two forms of nested tensors present within PyTorch, distinguished by layout as
specified during construction. Layout can be one of <code class="docutils literal notranslate"><span class="pre">torch.strided</span></code> or <code class="docutils literal notranslate"><span class="pre">torch.jagged</span></code>.
We recommend utilizing the <code class="docutils literal notranslate"><span class="pre">torch.jagged</span></code> layout whenever possible. While it currently only
supports a single ragged dimension, it has better op coverage, receives active development, and
integrates well with <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code>. These docs adhere to this recommendation and refer to
nested tensors with the <code class="docutils literal notranslate"><span class="pre">torch.jagged</span></code> layout as “NJTs” for brevity throughout.</p>
</div>
<p>Construction is straightforward and involves passing a list of tensors to the
<code class="docutils literal notranslate"><span class="pre">torch.nested.nested_tensor</span></code> constructor. A nested tensor with the <code class="docutils literal notranslate"><span class="pre">torch.jagged</span></code> layout
(AKA an “NJT”) supports a single ragged dimension. This constructor will copy the input tensors
into a packed, contiguous block of memory according to the layout described in the <a class="reference internal" href="#data-layout">data_layout</a>
section below.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span><span class="p">,</span> <span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">3</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">5</span><span class="p">)</span> <span class="o">+</span> <span class="mi">3</span>
<span class="gp">>>> </span><span class="n">a</span>
<span class="go">tensor([0, 1, 2])</span>
<span class="gp">>>> </span><span class="n">b</span>
<span class="go">tensor([3, 4, 5, 6, 7])</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">([</span><span class="n">component</span> <span class="k">for</span> <span class="n">component</span> <span class="ow">in</span> <span class="n">nt</span><span class="p">])</span>
<span class="go">[tensor([0, 1, 2]), tensor([3, 4, 5, 6, 7])]</span>
</pre></div>
</div>
<p>Each tensor in the list must have the same number of dimensions, but the shapes can otherwise vary
along a single dimension. If the dimensionalities of the input components don’t match, the
constructor throws an error.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">50</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span> <span class="c1"># 2D tensor</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">50</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span> <span class="c1"># 3D tensor</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">...</span>
<span class="go">RuntimeError: When constructing a nested tensor, all tensors in list must have the same dim</span>
</pre></div>
</div>
<p>During construction, dtype, device, and whether gradients are required can be chosen via the
usual keyword arguments.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">,</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">([</span><span class="n">component</span> <span class="k">for</span> <span class="n">component</span> <span class="ow">in</span> <span class="n">nt</span><span class="p">])</span>
<span class="go">[tensor([0., 1., 2.], device='cuda:0',</span>
<span class="go"> grad_fn=<UnbindBackwardAutogradNestedTensor0>), tensor([3., 4., 5., 6., 7.], device='cuda:0',</span>
<span class="go"> grad_fn=<UnbindBackwardAutogradNestedTensor0>)]</span>
</pre></div>
</div>
<p><code class="docutils literal notranslate"><span class="pre">torch.nested.as_nested_tensor</span></code> can be used to preserve autograd history from the tensors passed
to the constructor. When this constructor is utilized, gradients will flow through the nested tensor
back into the original components. Note that this constructor still copies the input components into
a packed, contiguous block of memory.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">12</span><span class="p">,</span> <span class="mi">512</span><span class="p">,</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">23</span><span class="p">,</span> <span class="mi">512</span><span class="p">,</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">as_nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">a</span><span class="o">.</span><span class="n">grad</span>
<span class="go">tensor([[1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> ...,</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.]])</span>
<span class="gp">>>> </span><span class="n">b</span><span class="o">.</span><span class="n">grad</span>
<span class="go">tensor([[1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> ...,</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.],</span>
<span class="go"> [1., 1., 1., ..., 1., 1., 1.]])</span>
</pre></div>
</div>
<p>The above functions all create contiguous NJTs, where a chunk of memory is allocated to store
a packed form of the underlying components (see the <a class="reference internal" href="#data-layout">data_layout</a> section below for more
details).</p>
<p>It is also possible to create a non-contiguous NJT view over a pre-existing dense tensor
with padding, avoiding the memory allocation and copying. <code class="docutils literal notranslate"><span class="pre">torch.nested.narrow()</span></code> is the tool
for accomplishing this.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">padded</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">5</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">seq_lens</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mi">3</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="mi">5</span><span class="p">],</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">int64</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">narrow</span><span class="p">(</span><span class="n">padded</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="n">start</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span> <span class="n">length</span><span class="o">=</span><span class="n">seq_lens</span><span class="p">,</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([3, j1, 4])</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">is_contiguous</span><span class="p">()</span>
<span class="go">False</span>
</pre></div>
</div>
<p>Note that the nested tensor acts as a view over the original padded dense tensor, referencing the
same memory without copying / allocation. Operation support for non-contiguous NJTs is somewhat more
limited, so if you run into support gaps, it’s always possible to convert to a contiguous NJT
using <code class="docutils literal notranslate"><span class="pre">contiguous()</span></code>.</p>
</div>
<div class="section" id="data-layout-and-shape">
<span id="data-layout"></span><h2>Data Layout and Shape<a class="headerlink" href="#data-layout-and-shape" title="Permalink to this heading">¶</a></h2>
<p>For efficiency, nested tensors generally pack their tensor components into a contiguous chunk of
memory and maintain additional metadata to specify batch item boundaries. For the <code class="docutils literal notranslate"><span class="pre">torch.jagged</span></code>
layout, the contiguous chunk of memory is stored in the <code class="docutils literal notranslate"><span class="pre">values</span></code> component, with the <code class="docutils literal notranslate"><span class="pre">offsets</span></code>
component delineating batch item boundaries for the ragged dimension.</p>
<img alt="_images/njt_visual.png" src="_images/njt_visual.png" />
<p>It’s possible to directly access the underlying NJT components when necessary.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">50</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span> <span class="c1"># text 1</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">32</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span> <span class="c1"># text 2</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">values</span><span class="p">()</span><span class="o">.</span><span class="n">shape</span> <span class="c1"># note the "packing" of the ragged dimension; no padding needed</span>
<span class="go">torch.Size([82, 128])</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">offsets</span><span class="p">()</span>
<span class="go">tensor([ 0, 50, 82])</span>
</pre></div>
</div>
<p>It can also be useful to construct an NJT from the jagged <code class="docutils literal notranslate"><span class="pre">values</span></code> and <code class="docutils literal notranslate"><span class="pre">offsets</span></code>
constituents directly; the <code class="docutils literal notranslate"><span class="pre">torch.nested.nested_tensor_from_jagged()</span></code> constructor serves
this purpose.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">values</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">82</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">offsets</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mi">0</span><span class="p">,</span> <span class="mi">50</span><span class="p">,</span> <span class="mi">82</span><span class="p">],</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">int64</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor_from_jagged</span><span class="p">(</span><span class="n">values</span><span class="o">=</span><span class="n">values</span><span class="p">,</span> <span class="n">offsets</span><span class="o">=</span><span class="n">offsets</span><span class="p">)</span>
</pre></div>
</div>
<p>An NJT has a well-defined shape with dimensionality 1 greater than that of its components. The
underlying structure of the ragged dimension is represented by a symbolic value (<code class="docutils literal notranslate"><span class="pre">j1</span></code> in the
example below).</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">50</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">32</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">dim</span><span class="p">()</span>
<span class="go">3</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, j1, 128])</span>
</pre></div>
</div>
<p>NJTs must have the same ragged structure to be compatible with each other. For example, to run a
binary operation involving two NJTs, the ragged structures must match (i.e. they must have the
same ragged shape symbol in their shapes). In the details, each symbol corresponds with an exact
<code class="docutils literal notranslate"><span class="pre">offsets</span></code> tensor, so both NJTs must have the same <code class="docutils literal notranslate"><span class="pre">offsets</span></code> tensor to be compatible with
each other.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">50</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">32</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt1</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt2</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt1</span><span class="o">.</span><span class="n">offsets</span><span class="p">()</span> <span class="ow">is</span> <span class="n">nt2</span><span class="o">.</span><span class="n">offsets</span><span class="p">()</span>
<span class="go">False</span>
<span class="gp">>>> </span><span class="n">nt3</span> <span class="o">=</span> <span class="n">nt1</span> <span class="o">+</span> <span class="n">nt2</span>
<span class="go">RuntimeError: cannot call binary pointwise function add.Tensor with inputs of shapes (2, j2, 128) and (2, j3, 128)</span>
</pre></div>
</div>
<p>In the above example, even though the conceptual shapes of the two NJTs are the same, they don’t
share a reference to the same <code class="docutils literal notranslate"><span class="pre">offsets</span></code> tensor, so their shapes differ, and they are not
compatible. We recognize that this behavior is unintuitive and are working hard to relax this
restriction for the beta release of nested tensors. For a workaround, see the
<a class="reference internal" href="#ragged-structure-incompatibility"><span class="std std-ref">Troubleshooting</span></a> section of this document.</p>
<p>In addition to the <code class="docutils literal notranslate"><span class="pre">offsets</span></code> metadata, NJTs can also compute and cache the minimum and maximum
sequence lengths for its components, which can be useful for invoking particular kernels (e.g. SDPA).
There are currently no public APIs for accessing these, but this will change for the beta release.</p>
</div>
<div class="section" id="supported-operations">
<span id="id2"></span><h2>Supported Operations<a class="headerlink" href="#supported-operations" title="Permalink to this heading">¶</a></h2>
<p>This section contains a list of common operations over nested tensors that you may find useful.
It is not comprehensive, as there are on the order of a couple thousand ops within PyTorch. While
a sizeable subset of these are supported for nested tensors today, full support is a large task.
The ideal state for nested tensors is full support of all PyTorch operations that are available
for non-nested tensors. To help us accomplish this, please consider:</p>
<ul class="simple">
<li><p>Requesting particular ops needed for your use case
<a class="reference external" href="https://github.com/pytorch/pytorch/issues/118107">here</a> to help us prioritize.</p></li>
<li><p>Contributing! It’s not too hard to add nested tensor support for a given PyTorch op; see
the <a class="reference external" href="contributions">Contributions</a> section below for details.</p></li>
</ul>
<div class="section" id="viewing-nested-tensor-constituents">
<h3>Viewing nested tensor constituents<a class="headerlink" href="#viewing-nested-tensor-constituents" title="Permalink to this heading">¶</a></h3>
<p><code class="docutils literal notranslate"><span class="pre">unbind()</span></code> allows you to retrieve a view of the nested tensor’s constituents.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch</span>
<span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">unbind</span><span class="p">()</span>
<span class="go">(tensor([[-0.9916, -0.3363, -0.2799],</span>
<span class="go"> [-2.3520, -0.5896, -0.4374]]), tensor([[-2.0969, -1.0104, 1.4841],</span>
<span class="go"> [ 2.0952, 0.2973, 0.2516],</span>
<span class="go"> [ 0.9035, 1.3623, 0.2026]]))</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">unbind</span><span class="p">()[</span><span class="mi">0</span><span class="p">]</span> <span class="ow">is</span> <span class="ow">not</span> <span class="n">a</span>
<span class="go">True</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">unbind</span><span class="p">()[</span><span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">mul_</span><span class="p">(</span><span class="mi">3</span><span class="p">)</span>
<span class="go">tensor([[ 3.6858, -3.7030, -4.4525],</span>
<span class="go"> [-2.3481, 2.0236, 0.1975]])</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">unbind</span><span class="p">()</span>
<span class="go">(tensor([[-2.9747, -1.0089, -0.8396],</span>
<span class="go"> [-7.0561, -1.7688, -1.3122]]), tensor([[-2.0969, -1.0104, 1.4841],</span>
<span class="go"> [ 2.0952, 0.2973, 0.2516],</span>
<span class="go"> [ 0.9035, 1.3623, 0.2026]]))</span>
</pre></div>
</div>
<p>Note that <code class="docutils literal notranslate"><span class="pre">nt.unbind()[0]</span></code> is not a copy, but rather a slice of the underlying memory, which
represents the first entry or constituent of the nested tensor.</p>
</div>
<div class="section" id="conversions-to-from-padded">
<h3>Conversions to / from padded<a class="headerlink" href="#conversions-to-from-padded" title="Permalink to this heading">¶</a></h3>
<p><code class="docutils literal notranslate"><span class="pre">torch.nested.to_padded_tensor()</span></code> converts an NJT to a padded dense tensor with the specified
padding value. The ragged dimension will be padded out to the size of the maximum sequence length.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch</span>
<span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">6</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">padded</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">to_padded_tensor</span><span class="p">(</span><span class="n">nt</span><span class="p">,</span> <span class="n">padding</span><span class="o">=</span><span class="mf">4.2</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">padded</span>
<span class="go">tensor([[[ 1.6107, 0.5723, 0.3913],</span>
<span class="go"> [ 0.0700, -0.4954, 1.8663],</span>
<span class="go"> [ 4.2000, 4.2000, 4.2000],</span>
<span class="go"> [ 4.2000, 4.2000, 4.2000],</span>
<span class="go"> [ 4.2000, 4.2000, 4.2000],</span>
<span class="go"> [ 4.2000, 4.2000, 4.2000]],</span>
<span class="go"> [[-0.0479, -0.7610, -0.3484],</span>
<span class="go"> [ 1.1345, 1.0556, 0.3634],</span>
<span class="go"> [-1.7122, -0.5921, 0.0540],</span>
<span class="go"> [-0.5506, 0.7608, 2.0606],</span>
<span class="go"> [ 1.5658, -1.1934, 0.3041],</span>
<span class="go"> [ 0.1483, -1.1284, 0.6957]]])</span>
</pre></div>
</div>
<p>This can be useful as an escape hatch to work around NJT support gaps, but ideally such
conversions should be avoided when possible for optimal memory usage and performance, as the
more efficient nested tensor layout does not materialize padding.</p>
<p>The reverse conversion can be accomplished using <code class="docutils literal notranslate"><span class="pre">torch.nested.narrow()</span></code>, which applies
ragged structure to a given dense tensor to produce an NJT. Note that by default, this operation
does not copy the underlying data, and thus the output NJT is generally non-contiguous. It may be
useful to explicitly call <code class="docutils literal notranslate"><span class="pre">contiguous()</span></code> here if a contiguous NJT is desired.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">padded</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">5</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">seq_lens</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mi">3</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="mi">5</span><span class="p">],</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">int64</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">narrow</span><span class="p">(</span><span class="n">padded</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="n">length</span><span class="o">=</span><span class="n">seq_lens</span><span class="p">,</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([3, j1, 4])</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">nt</span><span class="o">.</span><span class="n">contiguous</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([3, j2, 4])</span>
</pre></div>
</div>
</div>
<div class="section" id="shape-manipulations">
<h3>Shape manipulations<a class="headerlink" href="#shape-manipulations" title="Permalink to this heading">¶</a></h3>
<p>Nested tensors support a wide array of operations for shape manipulation, including views.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">6</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">6</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, j1, 6])</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, j1, 6, 1])</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">unflatten</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="p">[</span><span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">])</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, j1, 2, 3])</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">([</span><span class="n">nt</span><span class="p">,</span> <span class="n">nt</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, j1, 12])</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">([</span><span class="n">nt</span><span class="p">,</span> <span class="n">nt</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, j1, 2, 6])</span>
<span class="gp">>>> </span><span class="n">nt</span><span class="o">.</span><span class="n">transpose</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="o">-</span><span class="mi">2</span><span class="p">)</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, 6, j1])</span>
</pre></div>
</div>
</div>
<div class="section" id="attention-mechanisms">
<h3>Attention mechanisms<a class="headerlink" href="#attention-mechanisms" title="Permalink to this heading">¶</a></h3>
<p>As variable-length sequences are common inputs to attention mechanisms, nested tensors support
important attention operators
<a class="reference external" href="https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html">Scaled Dot Product Attention (SDPA)</a> and
<a class="reference external" href="https://pytorch.org/docs/stable/nn.attention.flex_attention.html#module-torch.nn.attention.flex_attention">FlexAttention</a>.
See
<a class="reference external" href="https://pytorch.org/tutorials/intermediate/transformer_building_blocks.html#multiheadattention">here</a>
for usage examples of NJT with SDPA and
<a class="reference external" href="https://pytorch.org/tutorials/intermediate/transformer_building_blocks.html#flexattention-njt">here</a>
for usage examples of NJT with FlexAttention.</p>
</div>
</div>
<div class="section" id="usage-with-torch-compile">
<span id="id3"></span><h2>Usage with torch.compile<a class="headerlink" href="#usage-with-torch-compile" title="Permalink to this heading">¶</a></h2>
<p>NJTs are designed to be used with <code class="docutils literal notranslate"><span class="pre">torch.compile()</span></code> for optimal performance, and we always
recommend utilizing <code class="docutils literal notranslate"><span class="pre">torch.compile()</span></code> with NJTs when possible. NJTs work out-of-the-box and
graph-break-free both when passed as inputs to a compiled function or module OR when
instantiated in-line within the function.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>If you’re not able to utilize <code class="docutils literal notranslate"><span class="pre">torch.compile()</span></code> for your use case, performance and memory
usage may still benefit from the use of NJTs, but it’s not as clear-cut whether this will be
the case. It is important that the tensors being operated on are large enough so the
performance gains are not outweighed by the overhead of python tensor subclasses.</p>
</div>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">import</span> <span class="nn">torch</span>
<span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">>>> </span><span class="k">def</span> <span class="nf">f</span><span class="p">(</span><span class="n">x</span><span class="p">):</span> <span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">sin</span><span class="p">()</span> <span class="o">+</span> <span class="mi">1</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="n">compiled_f</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">f</span><span class="p">,</span> <span class="n">fullgraph</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">output</span> <span class="o">=</span> <span class="n">compiled_f</span><span class="p">(</span><span class="n">nt</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">output</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, j1, 3])</span>
<span class="gp">>>> </span><span class="k">def</span> <span class="nf">g</span><span class="p">(</span><span class="n">values</span><span class="p">,</span> <span class="n">offsets</span><span class="p">):</span> <span class="k">return</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor_from_jagged</span><span class="p">(</span><span class="n">values</span><span class="p">,</span> <span class="n">offsets</span><span class="p">)</span> <span class="o">*</span> <span class="mf">2.</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="n">compiled_g</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">g</span><span class="p">,</span> <span class="n">fullgraph</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">output2</span> <span class="o">=</span> <span class="n">compiled_g</span><span class="p">(</span><span class="n">nt</span><span class="o">.</span><span class="n">values</span><span class="p">(),</span> <span class="n">nt</span><span class="o">.</span><span class="n">offsets</span><span class="p">())</span>
<span class="gp">>>> </span><span class="n">output2</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, j1, 3])</span>
</pre></div>
</div>
<p>Note that NJTs support
<a class="reference external" href="https://pytorch.org/docs/stable/torch.compiler_dynamic_shapes.html">Dynamic Shapes</a>
to avoid unnecessary recompiles with changing ragged structure.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">c</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">5</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">d</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">6</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt1</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt2</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">c</span><span class="p">,</span> <span class="n">d</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">)</span>
<span class="gp">>>> </span><span class="k">def</span> <span class="nf">f</span><span class="p">(</span><span class="n">x</span><span class="p">):</span> <span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">sin</span><span class="p">()</span> <span class="o">+</span> <span class="mi">1</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="n">compiled_f</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">f</span><span class="p">,</span> <span class="n">fullgraph</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">output1</span> <span class="o">=</span> <span class="n">compiled_f</span><span class="p">(</span><span class="n">nt1</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">output2</span> <span class="o">=</span> <span class="n">compiled_f</span><span class="p">(</span><span class="n">nt2</span><span class="p">)</span> <span class="c1"># NB: No recompile needed even though ragged structure differs</span>
</pre></div>
</div>
<p>If you run into problems or arcane errors when utilizing NJT + <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code>, please file a
PyTorch issue. Full subclass support within <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code> is a long-term effort and there may
be some rough edges at this time.</p>
</div>
<div class="section" id="troubleshooting">
<span id="id4"></span><h2>Troubleshooting<a class="headerlink" href="#troubleshooting" title="Permalink to this heading">¶</a></h2>
<p>This section contains common errors that you may run into when utilizing nested tensors, alongside
the reason for these errors and suggestions for how to address them.</p>
<div class="section" id="unimplemented-ops">
<span id="unimplemented-op"></span><h3>Unimplemented ops<a class="headerlink" href="#unimplemented-ops" title="Permalink to this heading">¶</a></h3>
<p>This error is becoming rarer as nested tensor op support grows, but it’s still possible to hit it
today given that there are a couple thousand ops within PyTorch.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="ne">NotImplementedError</span><span class="p">:</span> <span class="n">aten</span><span class="o">.</span><span class="n">view_as_real</span><span class="o">.</span><span class="n">default</span>
</pre></div>
</div>
<p>The error is straightforward; we haven’t gotten around to adding op support for this particular op
yet. If you’d like, you can <a class="reference external" href="contributions">contribute</a> an implementation yourself OR simply
<a class="reference external" href="https://github.com/pytorch/pytorch/issues/118107">request</a> that we add support for this op
in a future PyTorch release.</p>
</div>
<div class="section" id="ragged-structure-incompatibility">
<span id="id5"></span><h3>Ragged structure incompatibility<a class="headerlink" href="#ragged-structure-incompatibility" title="Permalink to this heading">¶</a></h3>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="ne">RuntimeError</span><span class="p">:</span> <span class="n">cannot</span> <span class="n">call</span> <span class="n">binary</span> <span class="n">pointwise</span> <span class="n">function</span> <span class="n">add</span><span class="o">.</span><span class="n">Tensor</span> <span class="k">with</span> <span class="n">inputs</span> <span class="n">of</span> <span class="n">shapes</span> <span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="n">j2</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span> <span class="ow">and</span> <span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="n">j3</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
</pre></div>
</div>
<p>This error occurs when calling an op that operates over multiple NJTs with incompatible ragged
structures. Currently, it is required that input NJTs have the exact same <code class="docutils literal notranslate"><span class="pre">offsets</span></code> constituent
in order to have the same symbolic ragged structure symbol (e.g. <code class="docutils literal notranslate"><span class="pre">j1</span></code>).</p>
<p>As a workaround for this situation, it is possible to construct NJTs from the <code class="docutils literal notranslate"><span class="pre">values</span></code> and
<code class="docutils literal notranslate"><span class="pre">offsets</span></code> components directly. With both NJTs referencing the same <code class="docutils literal notranslate"><span class="pre">offsets</span></code> components, they
are considered to have the same ragged structure and are thus compatible.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">50</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">32</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt1</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt2</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor_from_jagged</span><span class="p">(</span><span class="n">values</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">82</span><span class="p">,</span> <span class="mi">128</span><span class="p">),</span> <span class="n">offsets</span><span class="o">=</span><span class="n">nt1</span><span class="o">.</span><span class="n">offsets</span><span class="p">())</span>
<span class="gp">>>> </span><span class="n">nt3</span> <span class="o">=</span> <span class="n">nt1</span> <span class="o">+</span> <span class="n">nt2</span>
<span class="gp">>>> </span><span class="n">nt3</span><span class="o">.</span><span class="n">shape</span>
<span class="go">torch.Size([2, j1, 128])</span>
</pre></div>
</div>
</div>
<div class="section" id="data-dependent-operation-within-torch-compile">
<h3>Data dependent operation within torch.compile<a class="headerlink" href="#data-dependent-operation-within-torch-compile" title="Permalink to this heading">¶</a></h3>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">torch</span><span class="o">.</span><span class="n">_dynamo</span><span class="o">.</span><span class="n">exc</span><span class="o">.</span><span class="n">Unsupported</span><span class="p">:</span> <span class="n">data</span> <span class="n">dependent</span> <span class="n">operator</span><span class="p">:</span> <span class="n">aten</span><span class="o">.</span><span class="n">_local_scalar_dense</span><span class="o">.</span><span class="n">default</span><span class="p">;</span> <span class="n">to</span> <span class="n">enable</span><span class="p">,</span> <span class="nb">set</span> <span class="n">torch</span><span class="o">.</span><span class="n">_dynamo</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">capture_scalar_outputs</span> <span class="o">=</span> <span class="kc">True</span>
</pre></div>
</div>
<p>This error occurs when calling an op that does data-dependent operation within torch.compile; this
commonly occurs for ops that need to examine the values of the NJT’s <code class="docutils literal notranslate"><span class="pre">offsets</span></code> to determine the
output shape. For example:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">50</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">32</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">nt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nested</span><span class="o">.</span><span class="n">nested_tensor</span><span class="p">([</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">],</span> <span class="n">layout</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">jagged</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span>
<span class="gp">>>> </span><span class="k">def</span> <span class="nf">f</span><span class="p">(</span><span class="n">nt</span><span class="p">):</span> <span class="k">return</span> <span class="n">nt</span><span class="o">.</span><span class="n">chunk</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)[</span><span class="mi">0</span><span class="p">]</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="n">compiled_f</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">compile</span><span class="p">(</span><span class="n">f</span><span class="p">,</span> <span class="n">fullgraph</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">output</span> <span class="o">=</span> <span class="n">compiled_f</span><span class="p">(</span><span class="n">nt</span><span class="p">)</span>
</pre></div>
</div>
<p>In this example, calling <code class="docutils literal notranslate"><span class="pre">chunk()</span></code> on the batch dimension of the NJT requires examination of the
NJT’s <code class="docutils literal notranslate"><span class="pre">offsets</span></code> data to delineate batch item boundaries within the packed ragged dimension. As a
workaround, there are a couple torch.compile flags that can be set:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">_dynamo</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">capture_dynamic_output_shape_ops</span> <span class="o">=</span> <span class="kc">True</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">_dynamo</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">capture_scalar_outputs</span> <span class="o">=</span> <span class="kc">True</span>
</pre></div>
</div>
<p>If, after setting these, you still see data-dependent operator errors, please file an issue with
PyTorch. This area of <code class="docutils literal notranslate"><span class="pre">torch.compile()</span></code> is still in heavy development and certain aspects of
NJT support may be incomplete.</p>
</div>