File size: 60,620 Bytes
fbff59d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
<!DOCTYPE html>
<html lang="en"><head>
<script src="llm_conf_files/libs/clipboard/clipboard.min.js"></script>
<script src="llm_conf_files/libs/quarto-html/tabby.min.js"></script>
<script src="llm_conf_files/libs/quarto-html/popper.min.js"></script>
<script src="llm_conf_files/libs/quarto-html/tippy.umd.min.js"></script>
<link href="llm_conf_files/libs/quarto-html/tippy.css" rel="stylesheet">
<link href="llm_conf_files/libs/quarto-html/light-border.css" rel="stylesheet">
<link href="llm_conf_files/libs/quarto-html/quarto-html.min.css" rel="stylesheet" data-mode="light">
<link href="llm_conf_files/libs/quarto-html/quarto-syntax-highlighting-dark.css" rel="stylesheet" id="quarto-text-highlighting-styles"><meta charset="utf-8">
  <meta name="generator" content="quarto-99.9.9">

  <title>Scaling Model Training with More Compute, How Do They Do It?</title>
  <meta name="apple-mobile-web-app-capable" content="yes">
  <meta name="apple-mobile-web-app-status-bar-style" content="black-translucent">
  <meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no, minimal-ui">
  <link rel="stylesheet" href="llm_conf_files/libs/revealjs/dist/reset.css">
  <link rel="stylesheet" href="llm_conf_files/libs/revealjs/dist/reveal.css">
  <style>
    code{white-space: pre-wrap;}
    span.smallcaps{font-variant: small-caps;}
    div.columns{display: flex; gap: min(4vw, 1.5em);}
    div.column{flex: auto; overflow-x: auto;}
    div.hanging-indent{margin-left: 1.5em; text-indent: -1.5em;}
    ul.task-list{list-style: none;}
    ul.task-list li input[type="checkbox"] {
      width: 0.8em;
      margin: 0 0.8em 0.2em -1em; /* quarto-specific, see https://github.com/quarto-dev/quarto-cli/issues/4556 */ 
      vertical-align: middle;
    }
    /* CSS for syntax highlighting */
    pre > code.sourceCode { white-space: pre; position: relative; }
    pre > code.sourceCode > span { line-height: 1.25; }
    pre > code.sourceCode > span:empty { height: 1.2em; }
    .sourceCode { overflow: visible; }
    code.sourceCode > span { color: inherit; text-decoration: inherit; }
    div.sourceCode { margin: 1em 0; }
    pre.sourceCode { margin: 0; }
    @media screen {
    div.sourceCode { overflow: auto; }
    }
    @media print {
    pre > code.sourceCode { white-space: pre-wrap; }
    pre > code.sourceCode > span { text-indent: -5em; padding-left: 5em; }
    }
    pre.numberSource code
      { counter-reset: source-line 0; }
    pre.numberSource code > span
      { position: relative; left: -4em; counter-increment: source-line; }
    pre.numberSource code > span > a:first-child::before
      { content: counter(source-line);
        position: relative; left: -1em; text-align: right; vertical-align: baseline;
        border: none; display: inline-block;
        -webkit-touch-callout: none; -webkit-user-select: none;
        -khtml-user-select: none; -moz-user-select: none;
        -ms-user-select: none; user-select: none;
        padding: 0 4px; width: 4em;
      }
    pre.numberSource { margin-left: 3em;  padding-left: 4px; }
    div.sourceCode
      { color: #f8f8f2;  }
    @media screen {
    pre > code.sourceCode > span > a:first-child::before { text-decoration: underline; }
    }
    code span { color: #f8f8f2; } /* Normal */
    code span.al { color: #f07178; background-color: #2a0f15; font-weight: bold; } /* Alert */
    code span.an { color: #d4d0ab; } /* Annotation */
    code span.at { color: #00e0e0; } /* Attribute */
    code span.bn { color: #d4d0ab; } /* BaseN */
    code span.bu { color: #abe338; } /* BuiltIn */
    code span.cf { color: #ffa07a; font-weight: bold; } /* ControlFlow */
    code span.ch { color: #abe338; } /* Char */
    code span.cn { color: #ffd700; } /* Constant */
    code span.co { color: #f8f8f2; font-style: italic; } /* Comment */
    code span.cv { color: #ffd700; } /* CommentVar */
    code span.do { color: #f8f8f2; } /* Documentation */
    code span.dt { color: #ffa07a; } /* DataType */
    code span.dv { color: #d4d0ab; } /* DecVal */
    code span.er { color: #f07178; text-decoration: underline; } /* Error */
    code span.ex { color: #00e0e0; font-weight: bold; } /* Extension */
    code span.fl { color: #d4d0ab; } /* Float */
    code span.fu { color: #ffa07a; } /* Function */
    code span.im { color: #abe338; } /* Import */
    code span.in { color: #d4d0ab; } /* Information */
    code span.kw { color: #ffa07a; font-weight: bold; } /* Keyword */
    code span.op { color: #ffa07a; } /* Operator */
    code span.ot { color: #00e0e0; } /* Other */
    code span.pp { color: #dcc6e0; } /* Preprocessor */
    code span.re { color: #00e0e0; background-color: #f8f8f2; } /* RegionMarker */
    code span.sc { color: #abe338; } /* SpecialChar */
    code span.ss { color: #abe338; } /* SpecialString */
    code span.st { color: #abe338; } /* String */
    code span.va { color: #00e0e0; } /* Variable */
    code span.vs { color: #abe338; } /* VerbatimString */
    code span.wa { color: #dcc6e0; } /* Warning */
  </style>
  <link rel="stylesheet" href="llm_conf_files/libs/revealjs/dist/theme/quarto.css">
  <link href="llm_conf_files/libs/revealjs/plugin/quarto-line-highlight/line-highlight.css" rel="stylesheet">
  <link href="llm_conf_files/libs/revealjs/plugin/reveal-menu/menu.css" rel="stylesheet">
  <link href="llm_conf_files/libs/revealjs/plugin/reveal-menu/quarto-menu.css" rel="stylesheet">
  <link href="llm_conf_files/libs/revealjs/plugin/quarto-support/footer.css" rel="stylesheet">
  <style type="text/css">

  .callout {
    margin-top: 1em;
    margin-bottom: 1em;  
    border-radius: .25rem;
  }

  .callout.callout-style-simple { 
    padding: 0em 0.5em;
    border-left: solid #acacac .3rem;
    border-right: solid 1px silver;
    border-top: solid 1px silver;
    border-bottom: solid 1px silver;
    display: flex;
  }

  .callout.callout-style-default {
    border-left: solid #acacac .3rem;
    border-right: solid 1px silver;
    border-top: solid 1px silver;
    border-bottom: solid 1px silver;
  }

  .callout .callout-body-container {
    flex-grow: 1;
  }

  .callout.callout-style-simple .callout-body {
    font-size: 1rem;
    font-weight: 400;
  }

  .callout.callout-style-default .callout-body {
    font-size: 0.9rem;
    font-weight: 400;
  }

  .callout.callout-titled.callout-style-simple .callout-body {
    margin-top: 0.2em;
  }

  .callout:not(.callout-titled) .callout-body {
      display: flex;
  }

  .callout:not(.no-icon).callout-titled.callout-style-simple .callout-content {
    padding-left: 1.6em;
  }

  .callout.callout-titled .callout-header {
    padding-top: 0.2em;
    margin-bottom: -0.2em;
  }

  .callout.callout-titled .callout-title  p {
    margin-top: 0.5em;
    margin-bottom: 0.5em;
  }
    
  .callout.callout-titled.callout-style-simple .callout-content  p {
    margin-top: 0;
  }

  .callout.callout-titled.callout-style-default .callout-content  p {
    margin-top: 0.7em;
  }

  .callout.callout-style-simple div.callout-title {
    border-bottom: none;
    font-size: .9rem;
    font-weight: 600;
    opacity: 75%;
  }

  .callout.callout-style-default  div.callout-title {
    border-bottom: none;
    font-weight: 600;
    opacity: 85%;
    font-size: 0.9rem;
    padding-left: 0.5em;
    padding-right: 0.5em;
  }

  .callout.callout-style-default div.callout-content {
    padding-left: 0.5em;
    padding-right: 0.5em;
  }

  .callout.callout-style-simple .callout-icon::before {
    height: 1rem;
    width: 1rem;
    display: inline-block;
    content: "";
    background-repeat: no-repeat;
    background-size: 1rem 1rem;
  }

  .callout.callout-style-default .callout-icon::before {
    height: 0.9rem;
    width: 0.9rem;
    display: inline-block;
    content: "";
    background-repeat: no-repeat;
    background-size: 0.9rem 0.9rem;
  }

  .callout-title {
    display: flex
  }
    
  .callout-icon::before {
    margin-top: 1rem;
    padding-right: .5rem;
  }

  .callout.no-icon::before {
    display: none !important;
  }

  .callout.callout-titled .callout-body > .callout-content > :last-child {
    padding-bottom: 0.5rem;
    margin-bottom: 0;
  }

  .callout.callout-titled .callout-icon::before {
    margin-top: .5rem;
    padding-right: .5rem;
  }

  .callout:not(.callout-titled) .callout-icon::before {
    margin-top: 1rem;
    padding-right: .5rem;
  }

  /* Callout Types */

  div.callout-note {
    border-left-color: #4582ec !important;
  }

  div.callout-note .callout-icon::before {
    background-image: url('data:image/png;base64,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');
  }

  div.callout-note.callout-style-default .callout-title {
    background-color: #dae6fb
  }

  div.callout-important {
    border-left-color: #d9534f !important;
  }

  div.callout-important .callout-icon::before {
    background-image: url('data:image/png;base64,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');
  }

  div.callout-important.callout-style-default .callout-title {
    background-color: #f7dddc
  }

  div.callout-warning {
    border-left-color: #f0ad4e !important;
  }

  div.callout-warning .callout-icon::before {
    background-image: url('data:image/png;base64,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');
  }

  div.callout-warning.callout-style-default .callout-title {
    background-color: #fcefdc
  }

  div.callout-tip {
    border-left-color: #02b875 !important;
  }

  div.callout-tip .callout-icon::before {
    background-image: url('data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAACAAAAAgCAYAAABzenr0AAAAAXNSR0IArs4c6QAAAERlWElmTU0AKgAAAAgAAYdpAAQAAAABAAAAGgAAAAAAA6ABAAMAAAABAAEAAKACAAQAAAABAAAAIKADAAQAAAABAAAAIAAAAACshmLzAAADr0lEQVRYCe1XTWgTQRj9ZjZV8a9SPIkKgj8I1bMHsUWrqYLVg4Ue6v9BwZOxSYsIerFao7UiUryIqJcqgtpimhbBXoSCVxUFe9CTiogUrUp2Pt+3aUI2u5vdNh4dmMzOzHvvezuz8xNFM0mjnbXaNu1MvFWRXkXEyE6aYOYJpdW4IXuA4r0fo8qqSMDBU0v1HJUgVieAXxzCsdE/YJTdFcVIZQNMyhruOMJKXYFoLfIfIvVIMWdsrd+Rpd86ZmyzzjJmLStqRn0v8lzkb4rVIXvnpScOJuAn2ACC65FkPzEdEy4TPWRLJ2h7z4cArXzzaOdKlbOvKKX25Wl00jSnrwVxAg3o4dRxhO13RBSdNvH0xSARv3adTXbBdTf64IWO2vH0LT+cv4GR1DJt+DUItaQogeBX/chhbTBxEiZ6gftlDNXTrvT7co4ub5A6gp9HIcHvzTa46OS5fBeP87Qm0fQkr4FsYgVQ7Qg+ZayaDg9jhg1GkWj8RG6lkeSacrrHgDaxdoBiZPg+NXV/KifMuB6//JmYH4CntVEHy/keA6x4h4CU5oFy8GzrBS18cLJMXcljAKB6INjWsRcuZBWVaS3GDrqB7rdapVIeA+isQ57Eev9eCqzqOa81CY05VLd6SamW2wA2H3SiTbnbSxmzfp7WtKZkqy4mdyAlGx7ennghYf8voqp9cLSgKdqNfa6RdRsAAkPwRuJZNbpByn+RrJi1RXTwdi8RQF6ymDwGMAtZ6TVE+4uoKh+MYkcLsT0Hk8eAienbiGdjJHZTpmNjlbFJNKDVAp2fJlYju6IreQxQ08UJDNYdoLSl6AadO+fFuCQqVMB1NJwPm69T04Wv5WhfcWyfXQB+wXRs1pt+nCknRa0LVzSA/2B+a9+zQJadb7IyyV24YAxKp2Jqs3emZTuNnKxsah+uabKbMk7CbTgJx/zIgQYErIeTKRQ9yD9wxVof5YolPHqaWo7TD6tJlh7jQnK5z2n3+fGdggIOx2kaa2YI9QWarc5Ce1ipNWMKeSG4DysFF52KBmTNMmn5HqCFkwy34rDg05gDwgH3bBi+sgFhN/e8QvRn8kbamCOhgrZ9GJhFDgfcMHzFb6BAtjKpFhzTjwv1KCVuxHvCbsSiEz4CANnj84cwHdFXAbAOJ4LTSAawGWFn5tDhLMYz6nWeU2wJfIhmIJBefcd/A5FWQWGgrWzyORZ3Q6HuV+Jf0Bj+BTX69fm1zWgK7By1YTXchFDORywnfQ7GpzOo6S+qECrsx2ifVQAAAABJRU5ErkJggg==');
  }

  div.callout-tip.callout-style-default .callout-title {
    background-color: #ccf1e3
  }

  div.callout-caution {
    border-left-color: #fd7e14 !important;
  }

  div.callout-caution .callout-icon::before {
    background-image: url('data:image/png;base64,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');
  }

  div.callout-caution.callout-style-default .callout-title {
    background-color: #ffe5d0
  }

  </style>
  <style type="text/css">
    .reveal div.sourceCode {
      margin: 0;
      overflow: auto;
    }
    .reveal div.hanging-indent {
      margin-left: 1em;
      text-indent: -1em;
    }
    .reveal .slide:not(.center) {
      height: 100%;
    }
    .reveal .slide.scrollable {
      overflow-y: auto;
    }
    .reveal .footnotes {
      height: 100%;
      overflow-y: auto;
    }
    .reveal .slide .absolute {
      position: absolute;
      display: block;
    }
    .reveal .footnotes ol {
      counter-reset: ol;
      list-style-type: none; 
      margin-left: 0;
    }
    .reveal .footnotes ol li:before {
      counter-increment: ol;
      content: counter(ol) ". "; 
    }
    .reveal .footnotes ol li > p:first-child {
      display: inline-block;
    }
    .reveal .slide ul,
    .reveal .slide ol {
      margin-bottom: 0.5em;
    }
    .reveal .slide ul li,
    .reveal .slide ol li {
      margin-top: 0.4em;
      margin-bottom: 0.2em;
    }
    .reveal .slide ul[role="tablist"] li {
      margin-bottom: 0;
    }
    .reveal .slide ul li > *:first-child,
    .reveal .slide ol li > *:first-child {
      margin-block-start: 0;
    }
    .reveal .slide ul li > *:last-child,
    .reveal .slide ol li > *:last-child {
      margin-block-end: 0;
    }
    .reveal .slide .columns:nth-child(3) {
      margin-block-start: 0.8em;
    }
    .reveal blockquote {
      box-shadow: none;
    }
    .reveal .tippy-content>* {
      margin-top: 0.2em;
      margin-bottom: 0.7em;
    }
    .reveal .tippy-content>*:last-child {
      margin-bottom: 0.2em;
    }
    .reveal .slide > img.stretch.quarto-figure-center,
    .reveal .slide > img.r-stretch.quarto-figure-center {
      display: block;
      margin-left: auto;
      margin-right: auto; 
    }
    .reveal .slide > img.stretch.quarto-figure-left,
    .reveal .slide > img.r-stretch.quarto-figure-left  {
      display: block;
      margin-left: 0;
      margin-right: auto; 
    }
    .reveal .slide > img.stretch.quarto-figure-right,
    .reveal .slide > img.r-stretch.quarto-figure-right  {
      display: block;
      margin-left: auto;
      margin-right: 0; 
    }
  </style>
  <script src="llm_conf_files/libs/quarto-diagram/mermaid.min.js"></script>
  <script src="llm_conf_files/libs/quarto-diagram/mermaid-init.js"></script>
  <link href="llm_conf_files/libs/quarto-diagram/mermaid.css" rel="stylesheet">
</head>
<body class="quarto-dark">
  <div class="reveal">
    <div class="slides">

<section id="title-slide" class="quarto-title-block center">
  <h1 class="title">Scaling Model Training with More Compute, How Do They Do It?</h1>

<div class="quarto-title-authors">
</div>

</section>
<section id="who-am-i" class="slide level2">
<h2>Who am I?</h2>
<ul>
<li>Zachary Mueller</li>
<li>Technical Lead for the 🤗 Accelerate project</li>
<li>API design geek</li>
</ul>
</section>
<section id="understanding-gpu-usage" class="slide level2">
<h2>Understanding GPU Usage</h2>
<ul>
<li>We can somewhat estimate the memory usage in vanilla full-fine-tuning of models</li>
<li>Requires certain assumptions (that I’ll be covering):
<ul>
<li>Adam optimizer</li>
<li>Batch size of 1</li>
</ul></li>
</ul>
</section>
<section id="understanding-gpu-usage-1" class="slide level2">
<h2>Understanding GPU Usage</h2>
<p>General estimate (<code>bert-base-cased</code>, 108M params):</p>
<ul>
<li>Each parameter is 4 bytes</li>
<li>Backward ~= 2x the model size</li>
<li>The optimizer step ~= 4x the model size (1x model, 1x gradients, 2x optimizer):</li>
</ul>
<div style="font-size: 50%;">
<table>
<thead>
<tr class="header">
<th>dtype</th>
<th style="text-align: left;">Model</th>
<th style="text-align: center;">Gradients</th>
<th style="text-align: center;">Backward pass</th>
<th style="text-align: center;">Optimizer step</th>
<th style="text-align: center;">Highest</th>
</tr>
</thead>
<tbody>
<tr class="odd">
<td>float32</td>
<td style="text-align: left;">413.18 MB</td>
<td style="text-align: center;">413.18 MB</td>
<td style="text-align: center;">826.36 MB</td>
<td style="text-align: center;">1.61 GB</td>
<td style="text-align: center;">1.61 GB</td>
</tr>
<tr class="even">
<td>float16</td>
<td style="text-align: left;">413.18 MB*</td>
<td style="text-align: center;">619.77 MB</td>
<td style="text-align: center;">826.36 MB</td>
<td style="text-align: center;">826.36 MB</td>
<td style="text-align: center;">826.36 MB</td>
</tr>
</tbody>
</table>
<p>*All estimations were based off the <a href="https://huggingface.co/spaces/hf-accelerate/model-memory-usage">Model Estimator Tool</a></p>
</div>
</section>
<section id="understanding-gpu-usage-2" class="slide level2">
<h2>Understanding GPU Usage</h2>
<p>This works fine for small models, we have cards with anywhere from 12-24GB of GPU memory (on the GPU-poor side).</p>
<p>But what happens as we scale?</p>
<p>Here’s <code>llama-3-8B</code> (8.03B parameters)</p>
<div style="font-size: 50%;">
<table>
<thead>
<tr class="header">
<th>dtype</th>
<th style="text-align: left;">Model</th>
<th style="text-align: center;">Gradients</th>
<th style="text-align: center;">Backward pass</th>
<th style="text-align: center;">Optimizer step</th>
<th style="text-align: center;">Highest</th>
</tr>
</thead>
<tbody>
<tr class="odd">
<td>float32</td>
<td style="text-align: left;">28.21 GB</td>
<td style="text-align: center;">28.21 GB</td>
<td style="text-align: center;">56.43 GB</td>
<td style="text-align: center;">112.84 GB</td>
<td style="text-align: center;">112.84 GB</td>
</tr>
<tr class="even">
<td>float16</td>
<td style="text-align: left;">28.21 GB*</td>
<td style="text-align: center;">42.32 GB</td>
<td style="text-align: center;">56.43 GB</td>
<td style="text-align: center;">56.43 GB</td>
<td style="text-align: center;">56.43 GB</td>
</tr>
</tbody>
</table>
</div>
<p>Well, <em>I</em> don’t have 56GB of GPU memory in a single card, let alone 112GB.</p>
<p>What can we do?</p>
</section>
<section>
<section id="distributed-training" class="title-slide slide level1 center">
<h1>Distributed Training</h1>

</section>
<section id="kinds-of-training" class="slide level2">
<h2>Kinds of Training</h2>
<ul>
<li>Single GPU:
<ul>
<li>No distributed techniques at play</li>
</ul></li>
<li>DDP:
<ul>
<li>A full copy of the model exists on each device, but data is chunked between each GPU</li>
</ul></li>
<li>FSDP &amp; DeepSpeed:
<ul>
<li>Split chunks of the model and optimizer states across GPUs, allowing for training bigger models on smaller (multiple) GPUs</li>
</ul></li>
</ul>
</section></section>
<section>
<section id="fully-sharded-data-parallelism" class="title-slide slide level1 center">
<h1>Fully Sharded Data Parallelism</h1>

</section>
<section id="fully-sharded-data-parallelism-1" class="slide level2">
<h2>Fully Sharded Data Parallelism</h2>

<img data-src="fsdp.png" id="fig-539a35d47e664c97a50115a146a7f1bd-1" class="r-stretch quarto-figure-center"><aside class="notes">
<ul>
<li>Take the model and split it across <code>n</code> GPUs</li>
<li>Each GPU computes the shard’s gradients</li>
<li>At the end, all gradients are synchronized and the final full model gradient is calculated</li>
<li>The backward pass can then be performed</li>
</ul>
<style type="text/css">
        span.MJX_Assistive_MathML {
          position:absolute!important;
          clip: rect(1px, 1px, 1px, 1px);
          padding: 1px 0 0 0!important;
          border: 0!important;
          height: 1px!important;
          width: 1px!important;
          overflow: hidden!important;
          display:block!important;
      }</style></aside>
</section>
<section id="fsdp-getting-parameter-specific" class="slide level2">
<h2>FSDP: Getting parameter specific</h2>
<ul>
<li>Different parameters can dicatate how much memory is needed for total GPU training across multiple GPUs</li>
<li>These include how model weights are sharded, gradients, and more.</li>
<li>I’ll cover some important ones I needed when doing a Full-Fine-Tune of Llama-3-8B <em>without PEFT</em> on 2x4090’s</li>
</ul>
</section>
<section id="sharding_strategy" class="slide level2">
<h2><code>sharding_strategy</code></h2>
<ul>
<li>Dictates the level of divving resources to perform
<ul>
<li><code>FULL_SHARD</code>: Includes optimizer states, gradients, and parameters</li>
<li><code>SHARD_GRAD_OP</code>: Includes optimizer states and gradients</li>
<li><code>NO_SHARD</code>: Normal DDP</li>
<li><code>HYBRID_SHARD</code>: Includes optimizer states, gradients, and parameters but each node has the full model</li>
</ul>
<aside class="notes">
<pre><code>FULL_SHARD:
  Parameters, Gradients, Optimizer States: All are sharded.
  Parameters Handling: Unshard before forward pass, reshard after forward pass, unshard before backward pass, reshard after backward pass.
  Gradients Handling: Synchronize and shard after backward pass.
  Optimizer States: Updated locally per rank.</code></pre>
<p>SHARD_GRAD_OP: Gradients and Optimizer States: Sharded during computation. Parameters: Unshard before forward pass, remain unsharded during forward pass, reshard after backward pass. Inside no_sync(): Parameters are not resharded after backward computation. Optimizer States: Updated locally per rank.</p>
<p>NO_SHARD: Parameters, Gradients, Optimizer States: Not sharded, replicated across ranks. Gradients Handling: Synchronized via all-reduce after backward pass. Optimizer States: Updated locally per rank.</p>
<p>HYBRID_SHARD: Parameters, Gradients, Optimizer States: Combines FULL_SHARD within a node and replicates parameters across nodes. Communication: Expensive operations like all-gathers and reduce-scatters are limited to within a node, enhancing performance for medium-sized models.</p>
<style type="text/css">
        span.MJX_Assistive_MathML {
          position:absolute!important;
          clip: rect(1px, 1px, 1px, 1px);
          padding: 1px 0 0 0!important;
          border: 0!important;
          height: 1px!important;
          width: 1px!important;
          overflow: hidden!important;
          display:block!important;
      }</style></aside></li>
</ul>
</section>
<section id="auto_wrap_policy" class="slide level2">
<h2><code>auto_wrap_policy</code>:</h2>
<ul>
<li>How the model should be split</li>
<li>Can be either <code>TRANSFORMER_BASED_WRAP</code> or <code>SIZE_BASED_WRAP</code></li>
<li><code>TRANSFORMER</code>/<code>fsdp_transformers_layer_cls_to_wrap</code>:
<ul>
<li>Need to declare the layer</li>
<li>Generally <code>transformers</code> has good defaults</li>
</ul></li>
<li><code>SIZE</code>/<code>fsdp_min_num_param</code>:
<ul>
<li>Number of total parameters in a shard</li>
</ul></li>
</ul>
</section>
<section id="offload_params" class="slide level2">
<h2><code>offload_params</code>:</h2>
<ul>
<li>Offloads the parameters and gradients to the CPU if they can’t fit into memory</li>
<li>Allows you to train much larger models locally, but will be much slower</li>
</ul>
<blockquote>
<p>Case: FFT of Llama-3-8B with <code>fsdp_offload_params</code> on 2x4090 GPUs was 72hrs, vs ~an hour or two when using 1xH100</p>
</blockquote>
</section>
<section id="cpu_ram_efficient_loading-and-sync_module_states" class="slide level2">
<h2><code>cpu_ram_efficient_loading</code> and <code>sync_module_states</code></h2>
<ul>
<li>Uses the idea behind big model inference/the <code>meta</code> device to load in the model to the GPU in a low-ram scenario</li>
<li>Rather than needing <code>model_size</code> * <code>n_gpus</code> RAM, we can load the model on a single node and then send the weights directly to each shard when the time is right via <code>sync_module_states</code></li>
</ul>
</section></section>
<section>
<section id="tying-this-to-accelerate" class="title-slide slide level1 center">
<h1>Tying this to 🤗 Accelerate</h1>

</section>
<section id="tying-this-to-accelerate-1" class="slide level2">
<h2>Tying this to 🤗 Accelerate</h2>
<ul>
<li>So far we’ve covered the theory, but how do we put it into practice</li>
<li>By using a library that’s at the heart of the entire open-source ecosystem</li>
</ul>
<div style="font-size: 60%;padding-left:10%;padding-top:0%;">
<ul>
<li>Nearly all of 🤗</li>
<li><code>axolotl</code></li>
<li><code>fastai</code></li>
<li><code>FastChat</code></li>
<li><code>lucidrains</code></li>
<li><code>kornia</code></li>
</ul>
</div>
<p>Are you using it and you don’t even know?</p>
</section>
<section id="what-is-accelerate" class="slide level2">
<h2>What is 🤗 Accelerate?</h2>
<div class="cell" data-reveal="true" data-fig-height="6">
<div class="cell-output-display">
<div>
<div>
<pre class="mermaid mermaid-js">graph LR
    A(("🤗 Accelerate#32;"))
    A --&gt; B["CLI Interface#32;"]
    A --&gt; C["Training Library#32;"]
    A --&gt; D["Big Model&lt;br&gt;Inference#32;"]
</pre>
</div>
</div>
</div>
</div>
</section>
<section id="a-cli-interface" class="slide level2">
<h2>A CLI Interface</h2>
<ul>
<li><code>accelerate config</code>
<ul>
<li>Configure the environment</li>
</ul></li>
<li><code>accelerate estimate-memory</code>
<ul>
<li>How to guess vRAM requirements</li>
</ul></li>
<li><code>accelerate launch</code>
<ul>
<li>How to run your script</li>
</ul></li>
</ul>
</section>
<section id="launching-distributed-training-is-hard" class="slide level2">
<h2>Launching distributed training is hard</h2>
<ul>
<li><div class="sourceCode" id="cb2"><pre class="sourceCode numberSource bash number-lines code-with-copy"><code class="sourceCode bash"><span id="cb2-1"><a href="#cb2-1"></a><span class="ex">python</span> script.py</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div></li>
<li><div class="sourceCode" id="cb3"><pre class="sourceCode numberSource bash number-lines code-with-copy"><code class="sourceCode bash"><span id="cb3-1"><a href="#cb3-1"></a><span class="ex">torchrun</span> <span class="at">--nnodes</span><span class="op">=</span>1 <span class="at">--nproc_per_node</span><span class="op">=</span>2 script.py</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div></li>
<li><div class="sourceCode" id="cb4"><pre class="sourceCode numberSource bash number-lines code-with-copy"><code class="sourceCode bash"><span id="cb4-1"><a href="#cb4-1"></a><span class="ex">deepspeed</span> <span class="at">--num_gpus</span><span class="op">=</span>2 script.py</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div></li>
</ul>
<p>How can we make this better?</p>
</section>
<section id="accelerate-launch" class="slide level2">
<h2><code>accelerate launch</code></h2>
<div class="sourceCode" id="cb5"><pre class="sourceCode numberSource bash number-lines code-with-copy"><code class="sourceCode bash"><span id="cb5-1"><a href="#cb5-1"></a><span class="ex">accelerate</span> launch script.py</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</section>
<section id="accelerate-config" class="slide level2">
<h2><code>accelerate config</code></h2>
<ul>
<li>Rely on <code>config.yaml</code> files</li>
<li>Choose to either running <code>accelerate config</code> or write your own:</li>
</ul>
<div class="columns" style="font-size: 50%;padding-left:10%;">
<div class="column" style="width:40%;">
<div class="code-with-filename">
<div class="code-with-filename-file">
<pre><strong>ddp_config.yaml</strong></pre>
</div>
<div class="sourceCode" id="cb6" data-filename="ddp_config.yaml"><pre class="sourceCode numberSource yaml number-lines code-with-copy"><code class="sourceCode yaml"><span id="cb6-1"><a href="#cb6-1"></a><span class="fu">compute_environment</span><span class="kw">:</span><span class="at"> LOCAL_MACHINE</span></span>
<span id="cb6-2"><a href="#cb6-2"></a><span class="fu">distributed_type</span><span class="kw">:</span><span class="at"> MULTI_GPU</span></span>
<span id="cb6-3"><a href="#cb6-3"></a><span class="fu">main_training_function</span><span class="kw">:</span><span class="at"> main</span></span>
<span id="cb6-4"><a href="#cb6-4"></a><span class="fu">mixed_precision</span><span class="kw">:</span><span class="at"> bf16</span></span>
<span id="cb6-5"><a href="#cb6-5"></a><span class="fu">num_machines</span><span class="kw">:</span><span class="at"> </span><span class="dv">1</span></span>
<span id="cb6-6"><a href="#cb6-6"></a><span class="fu">num_processes</span><span class="kw">:</span><span class="at"> </span><span class="dv">8</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
</div><div class="column" style="width:40%;">
<div class="code-with-filename">
<div class="code-with-filename-file">
<pre><strong>fsdp_config.yaml</strong></pre>
</div>
<div class="sourceCode" id="cb7" data-filename="fsdp_config.yaml"><pre class="sourceCode numberSource yaml number-lines code-with-copy"><code class="sourceCode yaml"><span id="cb7-1"><a href="#cb7-1"></a><span class="fu">compute_environment</span><span class="kw">:</span><span class="at"> LOCAL_MACHINE</span></span>
<span id="cb7-2"><a href="#cb7-2"></a><span class="fu">distributed_type</span><span class="kw">:</span><span class="at"> FSDP</span></span>
<span id="cb7-3"><a href="#cb7-3"></a><span class="fu">fsdp_config</span><span class="kw">:</span></span>
<span id="cb7-4"><a href="#cb7-4"></a><span class="at">  </span><span class="fu">fsdp_auto_wrap_policy</span><span class="kw">:</span><span class="at"> TRANSFORMER_BASED_WRAP</span></span>
<span id="cb7-5"><a href="#cb7-5"></a><span class="at">  </span><span class="fu">fsdp_backward_prefetch</span><span class="kw">:</span><span class="at"> BACKWARD_PRE</span></span>
<span id="cb7-6"><a href="#cb7-6"></a><span class="at">  </span><span class="fu">fsdp_cpu_ram_efficient_loading</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
<span id="cb7-7"><a href="#cb7-7"></a><span class="at">  </span><span class="fu">fsdp_forward_prefetch</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
<span id="cb7-8"><a href="#cb7-8"></a><span class="at">  </span><span class="fu">fsdp_offload_params</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
<span id="cb7-9"><a href="#cb7-9"></a><span class="at">  </span><span class="fu">fsdp_sharding_strategy</span><span class="kw">:</span><span class="at"> FULL_SHARD</span></span>
<span id="cb7-10"><a href="#cb7-10"></a><span class="at">  </span><span class="fu">fsdp_state_dict_type</span><span class="kw">:</span><span class="at"> SHARDED_STATE_DICT</span></span>
<span id="cb7-11"><a href="#cb7-11"></a><span class="at">  </span><span class="fu">fsdp_sync_module_states</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
<span id="cb7-12"><a href="#cb7-12"></a><span class="at">  </span><span class="fu">fsdp_use_orig_params</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
<span id="cb7-13"><a href="#cb7-13"></a><span class="fu">main_training_function</span><span class="kw">:</span><span class="at"> main</span></span>
<span id="cb7-14"><a href="#cb7-14"></a><span class="fu">mixed_precision</span><span class="kw">:</span><span class="at"> bf16</span></span>
<span id="cb7-15"><a href="#cb7-15"></a><span class="fu">num_machines</span><span class="kw">:</span><span class="at"> </span><span class="dv">1</span></span>
<span id="cb7-16"><a href="#cb7-16"></a><span class="fu">num_processes</span><span class="kw">:</span><span class="at"> </span><span class="dv">8</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
</div>
</div>
</section></section>
<section>
<section id="a-training-library" class="title-slide slide level1 center">
<h1>A Training Library</h1>

</section>
<section id="a-training-library-the-code" class="slide level2">
<h2>A Training Library: The Code</h2>
<div class="columns" style="font-size: 50%;">
<div class="column">
<p><br><br><br></p>
<div class="sourceCode" id="cb8" data-code-line-numbers="5-6,9"><pre class="sourceCode numberSource python number-lines code-with-copy"><code class="sourceCode python"><span id="cb8-1"><a href="#cb8-1"></a><span class="co"># For alignment purposes</span></span>
<span id="cb8-2"><a href="#cb8-2"></a><span class="cf">for</span> batch <span class="kw">in</span> dataloader:</span>
<span id="cb8-3"><a href="#cb8-3"></a>    optimizer.zero_grad()</span>
<span id="cb8-4"><a href="#cb8-4"></a>    inputs, targets <span class="op">=</span> batch</span>
<span id="cb8-5"><a href="#cb8-5"></a>    inputs <span class="op">=</span> inputs.to(device)</span>
<span id="cb8-6"><a href="#cb8-6"></a>    targets <span class="op">=</span> targets.to(device)</span>
<span id="cb8-7"><a href="#cb8-7"></a>    outputs <span class="op">=</span> model(inputs)</span>
<span id="cb8-8"><a href="#cb8-8"></a>    loss <span class="op">=</span> loss_function(outputs, targets)</span>
<span id="cb8-9"><a href="#cb8-9"></a>    loss.backward()</span>
<span id="cb8-10"><a href="#cb8-10"></a>    optimizer.step()</span>
<span id="cb8-11"><a href="#cb8-11"></a>    scheduler.step()</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div><div class="column">
<div class="sourceCode" id="cb9" data-code-line-numbers="1-7,12-13,16"><pre class="sourceCode numberSource python number-lines code-with-copy"><code class="sourceCode python"><span id="cb9-1"><a href="#cb9-1"></a><span class="im">from</span> accelerate <span class="im">import</span> Accelerator</span>
<span id="cb9-2"><a href="#cb9-2"></a>accelerator <span class="op">=</span> Accelerator()</span>
<span id="cb9-3"><a href="#cb9-3"></a>dataloader, model, optimizer scheduler <span class="op">=</span> (</span>
<span id="cb9-4"><a href="#cb9-4"></a>    accelerator.prepare(</span>
<span id="cb9-5"><a href="#cb9-5"></a>        dataloader, model, optimizer, scheduler</span>
<span id="cb9-6"><a href="#cb9-6"></a>    )</span>
<span id="cb9-7"><a href="#cb9-7"></a>)</span>
<span id="cb9-8"><a href="#cb9-8"></a></span>
<span id="cb9-9"><a href="#cb9-9"></a><span class="cf">for</span> batch <span class="kw">in</span> dataloader:</span>
<span id="cb9-10"><a href="#cb9-10"></a>    optimizer.zero_grad()</span>
<span id="cb9-11"><a href="#cb9-11"></a>    inputs, targets <span class="op">=</span> batch</span>
<span id="cb9-12"><a href="#cb9-12"></a>    <span class="co"># inputs = inputs.to(device)</span></span>
<span id="cb9-13"><a href="#cb9-13"></a>    <span class="co"># targets = targets.to(device)</span></span>
<span id="cb9-14"><a href="#cb9-14"></a>    outputs <span class="op">=</span> model(inputs)</span>
<span id="cb9-15"><a href="#cb9-15"></a>    loss <span class="op">=</span> loss_function(outputs, targets)</span>
<span id="cb9-16"><a href="#cb9-16"></a>    accelerator.backward(loss) <span class="co"># loss.backward()</span></span>
<span id="cb9-17"><a href="#cb9-17"></a>    optimizer.step()</span>
<span id="cb9-18"><a href="#cb9-18"></a>    scheduler.step()</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
</div>
</section>
<section id="a-training-library-how-scaling-works" class="slide level2">
<h2>A Training Library: How Scaling Works</h2>
<ul>
<li>Accelerate’s DataLoaders and schedulers work off of a sharding mindset</li>
<li>Rather than repeating the same data across <code>n</code> nodes, we instead split it</li>
<li>Speeds up training linearly</li>
<li>Given a batch size of 16 on a single GPU, to recreate this across 8 GPUs you would use a batch size of 2</li>
<li>This also means the scheduler will be stepped <code>n</code> GPUs at a time per “global step”</li>
</ul>
</section>
<section id="a-training-library-mixed-precision" class="slide level2">
<h2>A Training Library: Mixed Precision</h2>
<ul>
<li>This may be a bit different than your “normal” idea of mixed precision.</li>
<li>We do <strong>not</strong> convert the model weights to BF16/FP16</li>
<li>Instead we <strong>wrap the forward pass</strong> with <code>autocast</code> to convert the gradients automatically</li>
<li>This preserves the original precision of the weights, which leads to stable training and better fine-tuning later on.</li>
<li><strong>If you use <code>.bf16()</code> weights, you are STUCK in bf16 perminantly</strong></li>
</ul>
</section>
<section id="a-training-library-mixed-precision-1" class="slide level2">
<h2>A Training Library: Mixed Precision</h2>
<ul>
<li>Let’s tie that back up to the model estimator with neat tools like NVIDIA’s TransformerEngine</li>
</ul>
<div style="font-size: 60%;">
<table style="width:100%;">
<colgroup>
<col style="width: 14%">
<col style="width: 14%">
<col style="width: 14%">
<col style="width: 14%">
<col style="width: 14%">
<col style="width: 14%">
<col style="width: 14%">
</colgroup>
<thead>
<tr class="header">
<th>Optimization Level</th>
<th>Computation (GEMM)</th>
<th>Comm</th>
<th>Weight</th>
<th>Master Weight</th>
<th>Weight Gradient</th>
<th>Optimizer States</th>
</tr>
</thead>
<tbody>
<tr class="odd">
<td>FP16 AMP</td>
<td>FP16</td>
<td>FP32</td>
<td>FP32</td>
<td>N/A</td>
<td>FP32</td>
<td>FP32+FP32</td>
</tr>
<tr class="even">
<td>Nvidia TE</td>
<td>FP8</td>
<td>FP32</td>
<td>FP32</td>
<td>N/A</td>
<td>FP32</td>
<td>FP32+FP32</td>
</tr>
<tr class="odd">
<td>MS-AMP O1</td>
<td>FP8</td>
<td>FP8</td>
<td>FP16</td>
<td>N/A</td>
<td>FP8</td>
<td>FP32+FP32</td>
</tr>
<tr class="even">
<td>MS-AMP O2</td>
<td>FP8</td>
<td>FP8</td>
<td>FP16</td>
<td>N/A</td>
<td>FP8</td>
<td>FP8+FP16</td>
</tr>
<tr class="odd">
<td>MS-AMP O3</td>
<td>FP8</td>
<td>FP8</td>
<td>FP8</td>
<td>FP16</td>
<td>FP8</td>
<td>FP8+FP16</td>
</tr>
</tbody>
</table>
</div>
<aside class="notes">
<p>What is actually happening: * Linear Layers and other certain compatible layers are wrapped in a special version that allows for FP8 computation * The general forward pass is wrapped around BF16 * This means that the most memory saved is done during the gradients of the model, <em>not</em> the model itself. * With tools like <code>MS-AMP</code> we can convert more chunks into lower precision, but again like before stable training occurs when the models weights are in full precision and the backprop happens in full precision too.</p>
<style type="text/css">
        span.MJX_Assistive_MathML {
          position:absolute!important;
          clip: rect(1px, 1px, 1px, 1px);
          padding: 1px 0 0 0!important;
          border: 0!important;
          height: 1px!important;
          width: 1px!important;
          overflow: hidden!important;
          display:block!important;
      }</style></aside>
</section>
<section id="deepspeed-vs-fully-sharded-data-parallelism" class="slide level2">
<h2>DeepSpeed vs Fully Sharded Data Parallelism</h2>
<ul>
<li>Extremely similar, however mostly used different naming conventions for items and slight tweaks in the implementation</li>
</ul>
<div style="font-size: 50%;">
<table style="width:100%;">
<colgroup>
<col style="width: 16%">
<col style="width: 16%">
<col style="width: 16%">
<col style="width: 16%">
<col style="width: 16%">
<col style="width: 16%">
</colgroup>
<thead>
<tr class="header">
<th>Framework</th>
<th>Model Loading (<code>torch_dtype</code>)</th>
<th>Mixed Precision</th>
<th>Preparation (Local)</th>
<th>Training</th>
<th>Optimizer (Local)</th>
</tr>
</thead>
<tbody>
<tr class="odd">
<td>FSDP</td>
<td>bf16</td>
<td>default (none)</td>
<td>bf16</td>
<td>bf16</td>
<td>bf16</td>
</tr>
<tr class="even">
<td>FSDP</td>
<td>bf16</td>
<td>bf16</td>
<td>fp32</td>
<td>bf16</td>
<td>fp32</td>
</tr>
<tr class="odd">
<td>DeepSpeed</td>
<td>bf16</td>
<td>bf16</td>
<td>fp32</td>
<td>bf16</td>
<td>fp32</td>
</tr>
</tbody>
</table>
</div>
<p>To learn more, check out the <a href="https://huggingface.co/docs/accelerate/concept_guides/fsdp_and_deepspeed">documentation</a> or join my office hours</p>
</section>
<section id="key-takeaways" class="slide level2">
<h2>Key Takeaways:</h2>
<ul>
<li>You can scale out training with <code>accelerate</code>, FSDP, and DeepSpeed across multiple GPUs to train bigger models</li>
<li>Techniques like <code>FP8</code> can help speed up training some and reduce computational overhead</li>
<li>Comes at a cost of end-precision and locking model weights for futher fine-tunes if not careful</li>
</ul>
</section>
<section id="some-handy-resources" class="slide level2">
<h2>Some Handy Resources</h2>
<ul>
<li><a href="https://hf.co/docs/accelerate">🤗 Accelerate documentation</a></li>
<li><a href="https://huggingface.co/docs/accelerate/basic_tutorials/launch">Launching distributed code</a></li>
<li><a href="https://huggingface.co/docs/accelerate/basic_tutorials/notebook">Distributed code and Jupyter Notebooks</a></li>
<li><a href="https://huggingface.co/docs/accelerate/basic_tutorials/migration">Migrating to 🤗 Accelerate easily</a></li>
<li><a href="https://huggingface.co/docs/accelerate/usage_guides/big_modeling">Big Model Inference tutorial</a></li>
<li><a href="https://huggingface.co/docs/accelerate/usage_guides/deepspeed">DeepSpeed and 🤗 Accelerate</a></li>
<li><a href="https://huggingface.co/docs/accelerate/usage_guides/fsdp">Fully Sharded Data Parallelism and 🤗 Accelerate</a></li>
<li><a href="https://huggingface.co/docs/accelerate/concept_guides/fsdp_and_deepspeed">FSDP vs DeepSpeed In-Depth</a></li>
</ul>
<div class="footer footer-default">

</div>
</section></section>
    </div>
  </div>

  <script>window.backupDefine = window.define; window.define = undefined;</script>
  <script src="llm_conf_files/libs/revealjs/dist/reveal.js"></script>
  <!-- reveal.js plugins -->
  <script src="llm_conf_files/libs/revealjs/plugin/quarto-line-highlight/line-highlight.js"></script>
  <script src="llm_conf_files/libs/revealjs/plugin/pdf-export/pdfexport.js"></script>
  <script src="llm_conf_files/libs/revealjs/plugin/reveal-menu/menu.js"></script>
  <script src="llm_conf_files/libs/revealjs/plugin/reveal-menu/quarto-menu.js"></script>
  <script src="llm_conf_files/libs/revealjs/plugin/quarto-support/support.js"></script>
  

  <script src="llm_conf_files/libs/revealjs/plugin/notes/notes.js"></script>
  <script src="llm_conf_files/libs/revealjs/plugin/search/search.js"></script>
  <script src="llm_conf_files/libs/revealjs/plugin/zoom/zoom.js"></script>
  <script src="llm_conf_files/libs/revealjs/plugin/math/math.js"></script>
  <script>window.define = window.backupDefine; window.backupDefine = undefined;</script>

  <script>

      // Full list of configuration options available at:
      // https://revealjs.com/config/
      Reveal.initialize({
'controlsAuto': true,
'previewLinksAuto': false,
'pdfSeparateFragments': false,
'autoAnimateEasing': "ease",
'autoAnimateDuration': 1,
'autoAnimateUnmatched': true,
'menu': {"side":"left","useTextContentForMissingTitles":true,"markers":false,"loadIcons":false,"custom":[{"title":"Tools","icon":"<i class=\"fas fa-gear\"></i>","content":"<ul class=\"slide-menu-items\">\n<li class=\"slide-tool-item active\" data-item=\"0\"><a href=\"#\" onclick=\"RevealMenuToolHandlers.fullscreen(event)\"><kbd>f</kbd> Fullscreen</a></li>\n<li class=\"slide-tool-item\" data-item=\"1\"><a href=\"#\" onclick=\"RevealMenuToolHandlers.speakerMode(event)\"><kbd>s</kbd> Speaker View</a></li>\n<li class=\"slide-tool-item\" data-item=\"2\"><a href=\"#\" onclick=\"RevealMenuToolHandlers.overview(event)\"><kbd>o</kbd> Slide Overview</a></li>\n<li class=\"slide-tool-item\" data-item=\"3\"><a href=\"#\" onclick=\"RevealMenuToolHandlers.togglePdfExport(event)\"><kbd>e</kbd> PDF Export Mode</a></li>\n<li class=\"slide-tool-item\" data-item=\"4\"><a href=\"#\" onclick=\"RevealMenuToolHandlers.keyboardHelp(event)\"><kbd>?</kbd> Keyboard Help</a></li>\n</ul>"}],"openButton":true},
'smaller': false,
 
        // Display controls in the bottom right corner
        controls: false,

        // Help the user learn the controls by providing hints, for example by
        // bouncing the down arrow when they first encounter a vertical slide
        controlsTutorial: false,

        // Determines where controls appear, "edges" or "bottom-right"
        controlsLayout: 'edges',

        // Visibility rule for backwards navigation arrows; "faded", "hidden"
        // or "visible"
        controlsBackArrows: 'faded',

        // Display a presentation progress bar
        progress: true,

        // Display the page number of the current slide
        slideNumber: false,

        // 'all', 'print', or 'speaker'
        showSlideNumber: 'all',

        // Add the current slide number to the URL hash so that reloading the
        // page/copying the URL will return you to the same slide
        hash: true,

        // Start with 1 for the hash rather than 0
        hashOneBasedIndex: false,

        // Flags if we should monitor the hash and change slides accordingly
        respondToHashChanges: true,

        // Push each slide change to the browser history
        history: true,

        // Enable keyboard shortcuts for navigation
        keyboard: true,

        // Enable the slide overview mode
        overview: true,

        // Disables the default reveal.js slide layout (scaling and centering)
        // so that you can use custom CSS layout
        disableLayout: false,

        // Vertical centering of slides
        center: false,

        // Enables touch navigation on devices with touch input
        touch: true,

        // Loop the presentation
        loop: false,

        // Change the presentation direction to be RTL
        rtl: false,

        // see https://revealjs.com/vertical-slides/#navigation-mode
        navigationMode: 'linear',

        // Randomizes the order of slides each time the presentation loads
        shuffle: false,

        // Turns fragments on and off globally
        fragments: true,

        // Flags whether to include the current fragment in the URL,
        // so that reloading brings you to the same fragment position
        fragmentInURL: false,

        // Flags if the presentation is running in an embedded mode,
        // i.e. contained within a limited portion of the screen
        embedded: false,

        // Flags if we should show a help overlay when the questionmark
        // key is pressed
        help: true,

        // Flags if it should be possible to pause the presentation (blackout)
        pause: true,

        // Flags if speaker notes should be visible to all viewers
        showNotes: false,

        // Global override for autoplaying embedded media (null/true/false)
        autoPlayMedia: null,

        // Global override for preloading lazy-loaded iframes (null/true/false)
        preloadIframes: null,

        // Number of milliseconds between automatically proceeding to the
        // next slide, disabled when set to 0, this value can be overwritten
        // by using a data-autoslide attribute on your slides
        autoSlide: 0,

        // Stop auto-sliding after user input
        autoSlideStoppable: true,

        // Use this method for navigation when auto-sliding
        autoSlideMethod: null,

        // Specify the average time in seconds that you think you will spend
        // presenting each slide. This is used to show a pacing timer in the
        // speaker view
        defaultTiming: null,

        // Enable slide navigation via mouse wheel
        mouseWheel: false,

        // The display mode that will be used to show slides
        display: 'block',

        // Hide cursor if inactive
        hideInactiveCursor: true,

        // Time before the cursor is hidden (in ms)
        hideCursorTime: 5000,

        // Opens links in an iframe preview overlay
        previewLinks: false,

        // Transition style (none/fade/slide/convex/concave/zoom)
        transition: 'none',

        // Transition speed (default/fast/slow)
        transitionSpeed: 'default',

        // Transition style for full page slide backgrounds
        // (none/fade/slide/convex/concave/zoom)
        backgroundTransition: 'none',

        // Number of slides away from the current that are visible
        viewDistance: 3,

        // Number of slides away from the current that are visible on mobile
        // devices. It is advisable to set this to a lower number than
        // viewDistance in order to save resources.
        mobileViewDistance: 2,

        // The "normal" size of the presentation, aspect ratio will be preserved
        // when the presentation is scaled to fit different resolutions. Can be
        // specified using percentage units.
        width: 1050,

        height: 700,

        // Factor of the display size that should remain empty around the content
        margin: 0.1,

        math: {
          mathjax: 'https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.0/MathJax.js',
          config: 'TeX-AMS_HTML-full',
          tex2jax: {
            inlineMath: [['\\(','\\)']],
            displayMath: [['\\[','\\]']],
            balanceBraces: true,
            processEscapes: false,
            processRefs: true,
            processEnvironments: true,
            preview: 'TeX',
            skipTags: ['script','noscript','style','textarea','pre','code'],
            ignoreClass: 'tex2jax_ignore',
            processClass: 'tex2jax_process'
          },
        },

        // reveal.js plugins
        plugins: [QuartoLineHighlight, PdfExport, RevealMenu, QuartoSupport,

          RevealMath,
          RevealNotes,
          RevealSearch,
          RevealZoom
        ]
      });
    </script>
    <script id="quarto-html-after-body" type="application/javascript">
    window.document.addEventListener("DOMContentLoaded", function (event) {
      const toggleBodyColorMode = (bsSheetEl) => {
        const mode = bsSheetEl.getAttribute("data-mode");
        const bodyEl = window.document.querySelector("body");
        if (mode === "dark") {
          bodyEl.classList.add("quarto-dark");
          bodyEl.classList.remove("quarto-light");
        } else {
          bodyEl.classList.add("quarto-light");
          bodyEl.classList.remove("quarto-dark");
        }
      }
      const toggleBodyColorPrimary = () => {
        const bsSheetEl = window.document.querySelector("link#quarto-bootstrap");
        if (bsSheetEl) {
          toggleBodyColorMode(bsSheetEl);
        }
      }
      toggleBodyColorPrimary();  
      const tabsets =  window.document.querySelectorAll(".panel-tabset-tabby")
      tabsets.forEach(function(tabset) {
        const tabby = new Tabby('#' + tabset.id);
      });
      const isCodeAnnotation = (el) => {
        for (const clz of el.classList) {
          if (clz.startsWith('code-annotation-')) {                     
            return true;
          }
        }
        return false;
      }
      const clipboard = new window.ClipboardJS('.code-copy-button', {
        text: function(trigger) {
          const codeEl = trigger.previousElementSibling.cloneNode(true);
          for (const childEl of codeEl.children) {
            if (isCodeAnnotation(childEl)) {
              childEl.remove();
            }
          }
          return codeEl.innerText;
        }
      });
      clipboard.on('success', function(e) {
        // button target
        const button = e.trigger;
        // don't keep focus
        button.blur();
        // flash "checked"
        button.classList.add('code-copy-button-checked');
        var currentTitle = button.getAttribute("title");
        button.setAttribute("title", "Copied!");
        let tooltip;
        if (window.bootstrap) {
          button.setAttribute("data-bs-toggle", "tooltip");
          button.setAttribute("data-bs-placement", "left");
          button.setAttribute("data-bs-title", "Copied!");
          tooltip = new bootstrap.Tooltip(button, 
            { trigger: "manual", 
              customClass: "code-copy-button-tooltip",
              offset: [0, -8]});
          tooltip.show();    
        }
        setTimeout(function() {
          if (tooltip) {
            tooltip.hide();
            button.removeAttribute("data-bs-title");
            button.removeAttribute("data-bs-toggle");
            button.removeAttribute("data-bs-placement");
          }
          button.setAttribute("title", currentTitle);
          button.classList.remove('code-copy-button-checked');
        }, 1000);
        // clear code selection
        e.clearSelection();
      });
      function tippyHover(el, contentFn, onTriggerFn, onUntriggerFn) {
        const config = {
          allowHTML: true,
          maxWidth: 500,
          delay: 100,
          arrow: false,
          appendTo: function(el) {
              return el.closest('section.slide') || el.parentElement;
          },
          interactive: true,
          interactiveBorder: 10,
          theme: 'light-border',
          placement: 'bottom-start',
        };
        if (contentFn) {
          config.content = contentFn;
        }
        if (onTriggerFn) {
          config.onTrigger = onTriggerFn;
        }
        if (onUntriggerFn) {
          config.onUntrigger = onUntriggerFn;
        }
          config['offset'] = [0,0];
          config['maxWidth'] = 700;
        window.tippy(el, config); 
      }
      const noterefs = window.document.querySelectorAll('a[role="doc-noteref"]');
      for (var i=0; i<noterefs.length; i++) {
        const ref = noterefs[i];
        tippyHover(ref, function() {
          // use id or data attribute instead here
          let href = ref.getAttribute('data-footnote-href') || ref.getAttribute('href');
          try { href = new URL(href).hash; } catch {}
          const id = href.replace(/^#\/?/, "");
          const note = window.document.getElementById(id);
          return note.innerHTML;
        });
      }
      const findCites = (el) => {
        const parentEl = el.parentElement;
        if (parentEl) {
          const cites = parentEl.dataset.cites;
          if (cites) {
            return {
              el,
              cites: cites.split(' ')
            };
          } else {
            return findCites(el.parentElement)
          }
        } else {
          return undefined;
        }
      };
      var bibliorefs = window.document.querySelectorAll('a[role="doc-biblioref"]');
      for (var i=0; i<bibliorefs.length; i++) {
        const ref = bibliorefs[i];
        const citeInfo = findCites(ref);
        if (citeInfo) {
          tippyHover(citeInfo.el, function() {
            var popup = window.document.createElement('div');
            citeInfo.cites.forEach(function(cite) {
              var citeDiv = window.document.createElement('div');
              citeDiv.classList.add('hanging-indent');
              citeDiv.classList.add('csl-entry');
              var biblioDiv = window.document.getElementById('ref-' + cite);
              if (biblioDiv) {
                citeDiv.innerHTML = biblioDiv.innerHTML;
              }
              popup.appendChild(citeDiv);
            });
            return popup.innerHTML;
          });
        }
      }
    });
    </script>
    

</body></html>