File size: 125,381 Bytes
eeb7ca1
1e8c453
eeb7ca1
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
1e8c453
eeb7ca1
1e8c453
 
 
 
 
 
 
eeb7ca1
 
 
 
 
1e8c453
 
eeb7ca1
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
eeb7ca1
 
1e8c453
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
1e8c453
eeb7ca1
1e8c453
 
eeb7ca1
1e8c453
 
eeb7ca1
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
eeb7ca1
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
eeb7ca1
 
1e8c453
 
 
 
eeb7ca1
 
1e8c453
eeb7ca1
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
eeb7ca1
 
 
1e8c453
eeb7ca1
1e8c453
eeb7ca1
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
1e6e9f4
eeb7ca1
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
1e8c453
eeb7ca1
 
1e8c453
eeb7ca1
 
 
 
1e8c453
 
eeb7ca1
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
1e8c453
 
 
 
 
eeb7ca1
 
 
 
 
1e6e9f4
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
1e8c453
eeb7ca1
 
 
1e8c453
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
eeb7ca1
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
1e8c453
 
 
 
eeb7ca1
 
1e8c453
eeb7ca1
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
1e8c453
eeb7ca1
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
1e8c453
eeb7ca1
1e8c453
eeb7ca1
 
 
1e8c453
eeb7ca1
 
 
1e8c453
eeb7ca1
 
 
 
1e8c453
eeb7ca1
 
1e8c453
eeb7ca1
 
1e8c453
eeb7ca1
 
0539589
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
1e8c453
eeb7ca1
 
0539589
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
1e8c453
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0539589
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
0539589
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
1e8c453
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
eeb7ca1
 
 
 
 
1e8c453
 
0539589
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
1e8c453
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
1e8c453
 
eeb7ca1
1e8c453
 
eeb7ca1
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
1e8c453
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
1e8c453
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
eeb7ca1
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0539589
1e8c453
0539589
eeb7ca1
 
 
 
 
1e8c453
 
 
 
 
 
 
eeb7ca1
 
 
 
1e8c453
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
eeb7ca1
 
 
 
 
1e8c453
 
eeb7ca1
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
1e8c453
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
 
 
 
eeb7ca1
1e8c453
 
 
 
 
 
 
 
 
eeb7ca1
1e8c453
eeb7ca1
1e8c453
eeb7ca1
1e8c453
eeb7ca1
1e8c453
 
eeb7ca1
 
1e8c453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eeb7ca1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e8c453
 
 
 
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
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
import ast
import copy
import functools
import glob
import inspect
import queue
import sys
import os
import time
import traceback
import types
import typing
import warnings
from datetime import datetime
import filelock
import requests
import psutil
from requests import ConnectTimeout, JSONDecodeError
from urllib3.exceptions import ConnectTimeoutError, MaxRetryError, ConnectionError
from requests.exceptions import ConnectionError as ConnectionError2
from requests.exceptions import ReadTimeout as ReadTimeout2

if os.path.dirname(os.path.abspath(__file__)) not in sys.path:
    sys.path.append(os.path.dirname(os.path.abspath(__file__)))

os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1'
os.environ['BITSANDBYTES_NOWELCOME'] = '1'
warnings.filterwarnings('ignore', category=UserWarning, message='TypedStorage is deprecated')

from enums import DocumentChoices, LangChainMode, no_lora_str, model_token_mapping, no_model_str, source_prefix, \
    source_postfix
from loaders import get_loaders
from utils import set_seed, clear_torch_cache, save_generate_output, NullContext, wrapped_partial, EThread, get_githash, \
    import_matplotlib, get_device, makedirs, get_kwargs, start_faulthandler, get_hf_server, FakeTokenizer, remove

start_faulthandler()
import_matplotlib()

SEED = 1236
set_seed(SEED)

from typing import Union

import fire
import torch
from transformers import GenerationConfig, AutoModel, TextIteratorStreamer

from prompter import Prompter, inv_prompt_type_to_model_lower, non_hf_types, PromptType, get_prompt, generate_prompt
from stopping import get_stopping

eval_extra_columns = ['prompt', 'response', 'score']

langchain_modes = [x.value for x in list(LangChainMode)]

scratch_base_dir = '/tmp/'


def main(
        load_8bit: bool = False,
        load_4bit: bool = False,
        load_half: bool = True,
        infer_devices: bool = True,
        base_model: str = '',
        tokenizer_base_model: str = '',
        lora_weights: str = "",
        gpu_id: int = 0,
        compile_model: bool = True,
        use_cache: bool = None,
        inference_server: str = "",
        prompt_type: Union[int, str] = None,
        prompt_dict: typing.Dict = None,

        model_lock: typing.List[typing.Dict[str, str]] = None,
        model_lock_columns: int = None,
        fail_if_cannot_connect: bool = False,

        # input to generation
        temperature: float = None,
        top_p: float = None,
        top_k: int = None,
        num_beams: int = None,
        repetition_penalty: float = None,
        num_return_sequences: int = None,
        do_sample: bool = None,
        max_new_tokens: int = None,
        min_new_tokens: int = None,
        early_stopping: Union[bool, str] = None,
        max_time: float = None,

        memory_restriction_level: int = None,
        debug: bool = False,
        save_dir: str = None,
        share: bool = True,
        local_files_only: bool = False,
        resume_download: bool = True,
        use_auth_token: Union[str, bool] = False,
        trust_remote_code: Union[str, bool] = True,
        offload_folder: str = "offline_folder",

        src_lang: str = "English",
        tgt_lang: str = "Russian",

        cli: bool = False,
        cli_loop: bool = True,
        gradio: bool = True,
        gradio_offline_level: int = 0,
        chat: bool = True,
        chat_context: bool = False,
        stream_output: bool = True,
        show_examples: bool = None,
        verbose: bool = False,
        h2ocolors: bool = True,
        height: int = 600,
        show_lora: bool = True,
        login_mode_if_model0: bool = False,
        block_gradio_exit: bool = True,
        concurrency_count: int = 1,
        api_open: bool = False,
        allow_api: bool = True,
        input_lines: int = 1,
        gradio_size: str = None,
        auth: typing.List[typing.Tuple[str, str]] = None,
        max_max_time=None,
        max_max_new_tokens=None,

        sanitize_user_prompt: bool = False,
        sanitize_bot_response: bool = False,

        extra_model_options: typing.List[str] = [],
        extra_lora_options: typing.List[str] = [],
        extra_server_options: typing.List[str] = [],

        score_model: str = 'OpenAssistant/reward-model-deberta-v3-large-v2',

        eval_filename: str = None,
        eval_prompts_only_num: int = 0,
        eval_prompts_only_seed: int = 1234,
        eval_as_output: bool = False,

        langchain_mode: str = 'Disabled',
        force_langchain_evaluate: bool = False,
        visible_langchain_modes: list = ['UserData', 'MyData'],
        document_choice: list = [DocumentChoices.All_Relevant.name],
        user_path: str = None,
        detect_user_path_changes_every_query: bool = False,
        load_db_if_exists: bool = True,
        keep_sources_in_context: bool = False,
        db_type: str = 'chroma',
        use_openai_embedding: bool = False,
        use_openai_model: bool = False,
        hf_embedding_model: str = None,
        allow_upload_to_user_data: bool = True,
        allow_upload_to_my_data: bool = True,
        enable_url_upload: bool = True,
        enable_text_upload: bool = True,
        enable_sources_list: bool = True,
        chunk: bool = True,
        chunk_size: int = 512,
        top_k_docs: int = None,
        reverse_docs: bool = True,
        auto_reduce_chunks: bool = True,
        max_chunks: int = 100,
        n_jobs: int = -1,
        enable_captions: bool = True,
        captions_model: str = "Salesforce/blip-image-captioning-base",
        pre_load_caption_model: bool = False,
        caption_gpu: bool = True,
        enable_ocr: bool = False,
):
    """

    :param load_8bit: load model in 8-bit using bitsandbytes
    :param load_4bit: load model in 4-bit using bitsandbytes
    :param load_half: load model in float16
    :param infer_devices: whether to control devices with gpu_id.  If False, then spread across GPUs
    :param base_model: model HF-type name.  If use --base_model to preload model, cannot unload in gradio in models tab
    :param tokenizer_base_model: tokenizer HF-type name.  Usually not required, inferred from base_model.
    :param lora_weights: LORA weights path/HF link
    :param gpu_id: if infer_devices, then use gpu_id for cuda device ID, or auto mode if gpu_id != -1
    :param compile_model Whether to compile the model
    :param use_cache: Whether to use caching in model (some models fail when multiple threads use)
    :param inference_server: Consume base_model as type of model at this address
                             Address can be text-generation-server hosting that base_model
                             e.g. python generate.py --inference_server="http://192.168.1.46:6112" --base_model=h2oai/h2ogpt-oasst1-512-12b
                             Or Address can be "openai_chat" or "openai" for OpenAI API
                             e.g. python generate.py --inference_server="openai_chat" --base_model=gpt-3.5-turbo
                             e.g. python generate.py --inference_server="openai" --base_model=text-davinci-003
    :param prompt_type: type of prompt, usually matched to fine-tuned model or plain for foundational model
    :param prompt_dict: If prompt_type=custom, then expects (some) items returned by get_prompt(..., return_dict=True)
    :param model_lock: Lock models to specific combinations, for ease of use and extending to many models
           Only used if gradio = True
           List of dicts, each dict has base_model, tokenizer_base_model, lora_weights, inference_server, prompt_type, and prompt_dict
           If all models have same prompt_type, and prompt_dict, can still specify that once in CLI outside model_lock as default for dict
           Can specify model_lock instead of those items on CLI
           As with CLI itself, base_model can infer prompt_type and prompt_dict if in prompter.py.
             Also, tokenizer_base_model and lora_weights are optional.
             Also, inference_server is optional if loading model from local system.
           All models provided will automatically appear in compare model mode
           Model loading-unloading and related choices will be disabled.  Model/lora/server adding will be disabled
    :param model_lock_columns: How many columns to show if locking models (and so showing all at once)
           If None, then defaults to up to 3
           if -1, then all goes into 1 row
           Maximum value is 4 due to non-dynamic gradio rendering elements
    :param fail_if_cannot_connect: if doing model locking (e.g. with many models), fail if True.  Otherwise ignore.
           Useful when many endpoints and want to just see what works, but still have to wait for timeout.
    :param temperature: generation temperature
    :param top_p: generation top_p
    :param top_k: generation top_k
    :param num_beams: generation number of beams
    :param repetition_penalty: generation repetition penalty
    :param num_return_sequences: generation number of sequences (1 forced for chat)
    :param do_sample: generation sample
    :param max_new_tokens: generation max new tokens
    :param min_new_tokens: generation min tokens
    :param early_stopping: generation early stopping
    :param max_time: maximum time to allow for generation
    :param memory_restriction_level: 0 = no restriction to tokens or model, 1 = some restrictions on token 2 = HF like restriction 3 = very low memory case
    :param debug: enable debug mode
    :param save_dir: directory chat data is saved to
    :param share: whether to share the gradio app with sharable URL
    :param local_files_only: whether to only use local files instead of doing to HF for models
    :param resume_download: whether to resume downloads from HF for models
    :param use_auth_token: whether to use HF auth token (requires CLI did huggingface-cli login before)
    :param trust_remote_code: whether to use trust any code needed for HF model
    :param offload_folder: path for spilling model onto disk
    :param src_lang: source languages to include if doing translation (None = all)
    :param tgt_lang: target languages to include if doing translation (None = all)
    :param cli: whether to use CLI (non-gradio) interface.
    :param cli_loop: whether to loop for CLI (False usually only for testing)
    :param gradio: whether to enable gradio, or to enable benchmark mode
    :param gradio_offline_level: > 0, then change fonts so full offline
           == 1 means backend won't need internet for fonts, but front-end UI might if font not cached
           == 2 means backend and frontend don't need internet to download any fonts.
           Note: Some things always disabled include HF telemetry, gradio telemetry, chromadb posthog that involve uploading.
           This option further disables google fonts for downloading, which is less intrusive than uploading,
           but still required in air-gapped case.  The fonts don't look as nice as google fonts, but ensure full offline behavior.
           Also set --share=False to avoid sharing a gradio live link.
    :param chat: whether to enable chat mode with chat history
    :param chat_context: whether to use extra helpful context if human_bot
    :param stream_output: whether to stream output from generate
    :param show_examples: whether to show clickable examples in gradio
    :param verbose: whether to show verbose prints
    :param h2ocolors: whether to use H2O.ai theme
    :param height: height of chat window
    :param show_lora: whether to show LORA options in UI (expert so can be hard to understand)
    :param login_mode_if_model0: set to True to load --base_model after client logs in, to be able to free GPU memory when model is swapped
    :param block_gradio_exit: whether to block gradio exit (used for testing)
    :param concurrency_count: gradio concurrency count (1 is optimal for LLMs)
    :param api_open: If False, don't let API calls skip gradio queue
    :param allow_api: whether to allow API calls at all to gradio server
    :param input_lines: how many input lines to show for chat box (>1 forces shift-enter for submit, else enter is submit)
    :param gradio_size: Overall size of text and spaces: "xsmall", "small", "medium", "large".
           Small useful for many chatbots in model_lock mode
    :param auth: gradio auth for launcher in form [(user1, pass1), (user2, pass2), ...]
                 e.g. --auth=[('jon','password')] with no spaces
    :param max_max_time: Maximum max_time for gradio slider
    :param max_max_new_tokens: Maximum max_new_tokens for gradio slider
    :param sanitize_user_prompt: whether to remove profanity from user input (slows down input processing)
    :param sanitize_bot_response: whether to remove profanity and repeat lines from bot output (about 2x slower generation for long streaming cases due to better_profanity being slow)
    :param extra_model_options: extra models to show in list in gradio
    :param extra_lora_options: extra LORA to show in list in gradio
    :param extra_server_options: extra servers to show in list in gradio
    :param score_model: which model to score responses (None means no scoring)
    :param eval_filename: json file to use for evaluation, if None is sharegpt
    :param eval_prompts_only_num: for no gradio benchmark, if using eval_filename prompts for eval instead of examples
    :param eval_prompts_only_seed: for no gradio benchmark, seed for eval_filename sampling
    :param eval_as_output: for no gradio benchmark, whether to test eval_filename output itself
    :param langchain_mode: Data source to include.  Choose "UserData" to only consume files from make_db.py.
           WARNING: wiki_full requires extra data processing via read_wiki_full.py and requires really good workstation to generate db, unless already present.
    :param force_langchain_evaluate: Whether to force langchain LLM use even if not doing langchain, mostly for testing.
    :param user_path: user path to glob from to generate db for vector search, for 'UserData' langchain mode.
           If already have db, any new/changed files are added automatically if path set, does not have to be same path used for prior db sources
    :param detect_user_path_changes_every_query: whether to detect if any files changed or added every similarity search (by file hashes).
           Expensive for large number of files, so not done by default.  By default only detect changes during db loading.
    :param visible_langchain_modes: dbs to generate at launch to be ready for LLM
           Can be up to ['wiki', 'wiki_full', 'UserData', 'MyData', 'github h2oGPT', 'DriverlessAI docs']
           But wiki_full is expensive and requires preparation
           To allow scratch space only live in session, add 'MyData' to list
           Default: If only want to consume local files, e.g. prepared by make_db.py, only include ['UserData']
           FIXME: Avoid 'All' for now, not implemented
    :param document_choice: Default document choice when taking subset of collection
    :param load_db_if_exists: Whether to load chroma db if exists or re-generate db
    :param keep_sources_in_context: Whether to keep url sources in context, not helpful usually
    :param db_type: 'faiss' for in-memory or 'chroma' or 'weaviate' for persisted on disk
    :param use_openai_embedding: Whether to use OpenAI embeddings for vector db
    :param use_openai_model: Whether to use OpenAI model for use with vector db
    :param hf_embedding_model: Which HF embedding model to use for vector db
           Default is instructor-large with 768 parameters per embedding if have GPUs, else all-MiniLM-L6-v1 if no GPUs
           Can also choose simpler model with 384 parameters per embedding: "sentence-transformers/all-MiniLM-L6-v2"
           Can also choose even better embedding with 1024 parameters: 'hkunlp/instructor-xl'
           We support automatically changing of embeddings for chroma, with a backup of db made if this is done
    :param allow_upload_to_user_data: Whether to allow file uploads to update shared vector db
    :param allow_upload_to_my_data: Whether to allow file uploads to update scratch vector db
    :param enable_url_upload: Whether to allow upload from URL
    :param enable_text_upload: Whether to allow upload of text
    :param enable_sources_list: Whether to allow list (or download for non-shared db) of list of sources for chosen db
    :param chunk: Whether to chunk data (True unless know data is already optimally chunked)
    :param chunk_size: Size of chunks, with typically top-4 passed to LLM, so neesd to be in context length
    :param top_k_docs: number of chunks to give LLM
    :param reverse_docs: whether to reverse docs order so most relevant is closest to question.
           Best choice for sufficiently smart model, and truncation occurs for oldest context, so best then too.
           But smaller 6_9 models fail to use newest context and can get stuck on old information.
    :param auto_reduce_chunks: Whether to automatically reduce top_k_docs to fit context given prompt
    :param max_chunks: If top_k_docs=-1, maximum number of chunks to allow
    :param n_jobs: Number of processors to use when consuming documents (-1 = all, is default)
    :param enable_captions: Whether to support captions using BLIP for image files as documents, then preloads that model
    :param captions_model: Which model to use for captions.
           captions_model: str = "Salesforce/blip-image-captioning-base",  # continue capable
           captions_model: str = "Salesforce/blip2-flan-t5-xl",   # question/answer capable, 16GB state
           captions_model: str = "Salesforce/blip2-flan-t5-xxl",  # question/answer capable, 60GB state
           Note: opt-based blip2 are not permissive license due to opt and Meta license restrictions
    :param pre_load_caption_model: Whether to preload caption model, or load after forking parallel doc loader
           parallel loading disabled if preload and have images, to prevent deadlocking on cuda context
           Recommended if using larger caption model
    :param caption_gpu: If support caption, then use GPU if exists
    :param enable_ocr: Whether to support OCR on images
    :return:
    """
    if base_model is None:
        base_model = ''
    if tokenizer_base_model is None:
        tokenizer_base_model = ''
    if lora_weights is None:
        lora_weights = ''
    if inference_server is None:
        inference_server = ''

    # listen to env if set
    model_lock = os.getenv('model_lock', str(model_lock))
    model_lock = ast.literal_eval(model_lock)

    if model_lock:
        assert gradio, "model_lock only supported for gradio=True"
        if len(model_lock) > 1:
            assert chat, "model_lock only works for multiple models for chat=True"
        assert not cli, "model_lock only supported for cli=False"
        assert not (not cli and not gradio), "model_lock only supported for eval (cli=gradio=False)"
        assert not base_model, "Don't specify model_lock and base_model"
        assert not tokenizer_base_model, "Don't specify model_lock and tokenizer_base_model"
        assert not lora_weights, "Don't specify model_lock and lora_weights"
        assert not inference_server, "Don't specify model_lock and inference_server"
        # assert not prompt_type, "Don't specify model_lock and prompt_type"
        # assert not prompt_dict, "Don't specify model_lock and prompt_dict"

    is_hf = bool(int(os.getenv("HUGGINGFACE_SPACES", '0')))
    is_gpth2oai = bool(int(os.getenv("GPT_H2O_AI", '0')))
    is_public = is_hf or is_gpth2oai  # multi-user case with fixed model and disclaimer
    if memory_restriction_level is None:
        memory_restriction_level = 2 if is_hf else 0  # 2 assumes run on 24GB consumer GPU
    else:
        assert 0 <= memory_restriction_level <= 3, "Bad memory_restriction_level=%s" % memory_restriction_level
    admin_pass = os.getenv("ADMIN_PASS")
    # will sometimes appear in UI or sometimes actual generation, but maybe better than empty result
    # but becomes unrecoverable sometimes if raise, so just be silent for now
    raise_generate_gpu_exceptions = True

    # allow set token directly
    use_auth_token = os.environ.get("HUGGINGFACE_API_TOKEN", use_auth_token)
    allow_upload_to_user_data = bool(
        int(os.environ.get("allow_upload_to_user_data", str(int(allow_upload_to_user_data)))))
    allow_upload_to_my_data = bool(int(os.environ.get("allow_upload_to_my_data", str(int(allow_upload_to_my_data)))))
    height = int(os.environ.get("HEIGHT", height))
    h2ocolors = bool(int(os.getenv('h2ocolors', h2ocolors)))

    # allow enabling langchain via ENV
    # FIRST PLACE where LangChain referenced, but no imports related to it
    langchain_mode = os.environ.get("LANGCHAIN_MODE", langchain_mode)
    assert langchain_mode in langchain_modes, "Invalid langchain_mode %s" % langchain_mode
    visible_langchain_modes = ast.literal_eval(os.environ.get("visible_langchain_modes", str(visible_langchain_modes)))
    if langchain_mode not in visible_langchain_modes and langchain_mode in langchain_modes:
        visible_langchain_modes += [langchain_mode]

    # if specifically chose not to show My or User Data, disable upload, so gradio elements are simpler
    if LangChainMode.MY_DATA.value not in visible_langchain_modes:
        allow_upload_to_my_data = False
    if LangChainMode.USER_DATA.value not in visible_langchain_modes:
        allow_upload_to_user_data = False

    if is_public:
        allow_upload_to_user_data = False
        input_lines = 1  # ensure set, for ease of use
        temperature = 0.2 if temperature is None else temperature
        top_p = 0.85 if top_p is None else top_p
        top_k = 70 if top_k is None else top_k
        if is_hf:
            do_sample = True if do_sample is None else do_sample
            top_k_docs = 3 if top_k_docs is None else top_k_docs
        else:
            # by default don't sample, too chatty
            do_sample = False if do_sample is None else do_sample
            top_k_docs = 4 if top_k_docs is None else top_k_docs

        if memory_restriction_level == 2:
            if not base_model and not inference_server:
                base_model = 'h2oai/h2ogpt-oasst1-512-12b'
                # don't set load_8bit if passed base_model, doesn't always work so can't just override
                load_8bit = True
                load_4bit = False  # FIXME - consider using 4-bit instead of 8-bit
        elif not inference_server:
            top_k_docs = 10 if top_k_docs is None else top_k_docs
    if memory_restriction_level >= 2:
        load_8bit = True
        load_4bit = False  # FIXME - consider using 4-bit instead of 8-bit
        if hf_embedding_model is None:
            hf_embedding_model = "sentence-transformers/all-MiniLM-L6-v2"
        top_k_docs = 3 if top_k_docs is None else top_k_docs
    if top_k_docs is None:
        top_k_docs = 3
    if is_public:
        if not max_time:
            max_time = 60 * 2
        if not max_max_time:
            max_max_time = max_time
        if not max_new_tokens:
            max_new_tokens = 256
        if not max_max_new_tokens:
            max_max_new_tokens = 256
    else:
        if not max_max_time:
            max_max_time = 60 * 20
        if not max_max_new_tokens:
            max_max_new_tokens = 512
    if is_hf:
        # must override share if in spaces
        share = False
        if not max_time:
            max_time = 60 * 1
        if not max_max_time:
            max_max_time = max_time
        # HF accounted for later in get_max_max_new_tokens()
    save_dir = os.getenv('SAVE_DIR', save_dir)
    score_model = os.getenv('SCORE_MODEL', score_model)
    if score_model == 'None' or score_model is None:
        score_model = ''
    concurrency_count = int(os.getenv('CONCURRENCY_COUNT', concurrency_count))
    api_open = bool(int(os.getenv('API_OPEN', str(int(api_open)))))
    allow_api = bool(int(os.getenv('ALLOW_API', str(int(allow_api)))))

    n_gpus = torch.cuda.device_count() if torch.cuda.is_available else 0
    if n_gpus == 0:
        gpu_id = None
        load_8bit = False
        load_4bit = False
        load_half = False
        infer_devices = False
        torch.backends.cudnn.benchmark = True
        torch.backends.cudnn.enabled = False
        torch.set_default_dtype(torch.float32)
        if psutil.virtual_memory().available < 94 * 1024 ** 3 and not inference_server:
            # 12B uses ~94GB
            # 6.9B uses ~47GB
            base_model = 'h2oai/h2ogpt-oig-oasst1-512-6_9b' if not base_model else base_model
        if hf_embedding_model is None:
            # if no GPUs, use simpler embedding model to avoid cost in time
            hf_embedding_model = "sentence-transformers/all-MiniLM-L6-v2"
    else:
        if hf_embedding_model is None:
            # if still None, then set default
            hf_embedding_model = 'hkunlp/instructor-large'

    # get defaults
    model_lower = base_model.lower()
    if not gradio:
        # force, else not single response like want to look at
        stream_output = False
        # else prompt removal can mess up output
        chat = False
    # hard-coded defaults
    first_para = False
    text_limit = None

    if offload_folder:
        makedirs(offload_folder)
    if user_path:
        makedirs(user_path)

    placeholder_instruction, placeholder_input, \
        stream_output, show_examples, \
        prompt_type, prompt_dict, \
        temperature, top_p, top_k, num_beams, \
        max_new_tokens, min_new_tokens, early_stopping, max_time, \
        repetition_penalty, num_return_sequences, \
        do_sample, \
        src_lang, tgt_lang, \
        examples, \
        task_info = \
        get_generate_params(model_lower, chat,
                            stream_output, show_examples,
                            prompt_type, prompt_dict,
                            temperature, top_p, top_k, num_beams,
                            max_new_tokens, min_new_tokens, early_stopping, max_time,
                            repetition_penalty, num_return_sequences,
                            do_sample,
                            top_k_docs,
                            chunk,
                            chunk_size,
                            verbose,
                            )

    git_hash = get_githash()
    locals_dict = locals()
    locals_print = '\n'.join(['%s: %s' % (k, v) for k, v in locals_dict.items()])
    if verbose:
        print(f"Generating model with params:\n{locals_print}", flush=True)
        print("Command: %s\nHash: %s" % (str(' '.join(sys.argv)), git_hash), flush=True)

    if langchain_mode != "Disabled":
        # SECOND PLACE where LangChain referenced, but all imports are kept local so not required
        from gpt_langchain import prep_langchain, get_some_dbs_from_hf
        if is_hf:
            get_some_dbs_from_hf()
        dbs = {}
        for langchain_mode1 in visible_langchain_modes:
            if langchain_mode1 in ['MyData']:
                # don't use what is on disk, remove it instead
                for gpath1 in glob.glob(os.path.join(scratch_base_dir, 'db_dir_%s*' % langchain_mode1)):
                    if os.path.isdir(gpath1):
                        print("Removing old MyData: %s" % gpath1, flush=True)
                        remove(gpath1)
                continue
            if langchain_mode1 in ['All']:
                # FIXME: All should be avoided until scans over each db, shouldn't be separate db
                continue
            persist_directory1 = 'db_dir_%s' % langchain_mode1  # single place, no special names for each case
            try:
                db = prep_langchain(persist_directory1,
                                    load_db_if_exists,
                                    db_type, use_openai_embedding,
                                    langchain_mode1, user_path,
                                    hf_embedding_model,
                                    kwargs_make_db=locals())
            finally:
                # in case updated embeddings or created new embeddings
                clear_torch_cache()
            dbs[langchain_mode1] = db
        # remove None db's so can just rely upon k in dbs for if hav db
        dbs = {k: v for k, v in dbs.items() if v is not None}
    else:
        dbs = {}
        # import control
        if os.environ.get("TEST_LANGCHAIN_IMPORT"):
            assert 'gpt_langchain' not in sys.modules, "Dev bug, import of langchain when should not have"
            assert 'langchain' not in sys.modules, "Dev bug, import of langchain when should not have"

    model_state_none = dict(model=None, tokenizer=None, device=None,
                            base_model=None, tokenizer_base_model=None, lora_weights=None,
                            inference_server=None, prompt_type=None, prompt_dict=None)

    if cli:
        from cli import run_cli
        return run_cli(**get_kwargs(run_cli, exclude_names=['model_state0'], **locals()))
    elif not gradio:
        from eval import run_eval
        return run_eval(**get_kwargs(run_eval, exclude_names=['model_state0'], **locals()))
    elif gradio:
        # imported here so don't require gradio to run generate
        from gradio_runner import go_gradio

        # get default model
        model_states = []
        model_list = [dict(base_model=base_model, tokenizer_base_model=tokenizer_base_model, lora_weights=lora_weights,
                           inference_server=inference_server, prompt_type=prompt_type, prompt_dict=prompt_dict)]
        model_list0 = copy.deepcopy(model_list)  # just strings, safe to deepcopy
        model_state0 = model_state_none.copy()
        assert len(model_state_none) == len(model_state0)
        if model_lock:
            model_list = model_lock
        for model_dict in reversed(model_list):
            # do reverse, so first is default base_model etc., so some logic works in go_gradio() more easily
            # handles defaults user didn't have to pass
            model_dict['base_model'] = base_model = model_dict.get('base_model', '')
            model_dict['tokenizer_base_model'] = tokenizer_base_model = model_dict.get('tokenizer_base_model', '')
            model_dict['lora_weights'] = lora_weights = model_dict.get('lora_weights', '')
            model_dict['inference_server'] = inference_server = model_dict.get('inference_server', '')
            prompt_type = model_dict.get('prompt_type', model_list0[0]['prompt_type'])  # don't use mutated value
            # try to infer, ignore empty initial state leading to get_generate_params -> 'plain'
            if model_dict.get('prompt_type') is None:
                model_lower = base_model.lower()
                if model_lower in inv_prompt_type_to_model_lower:
                    prompt_type = inv_prompt_type_to_model_lower[model_lower]
                    prompt_dict, error0 = get_prompt(prompt_type, '',
                                                     chat=False, context='', reduced=False, making_context=False,
                                                     return_dict=True)
            model_dict['prompt_type'] = prompt_type
            model_dict['prompt_dict'] = prompt_dict = model_dict.get('prompt_dict', prompt_dict)
            all_kwargs = locals().copy()
            if base_model and not login_mode_if_model0:
                model0, tokenizer0, device = get_model(reward_type=False,
                                                       **get_kwargs(get_model, exclude_names=['reward_type'],
                                                                    **all_kwargs))
            else:
                # if empty model, then don't load anything, just get gradio up
                model0, tokenizer0, device = None, None, None
            if model0 is None:
                if fail_if_cannot_connect:
                    raise RuntimeError("Could not connect, see logs")
                # skip
                if isinstance(model_lock, list):
                    model_lock.remove(model_dict)
                continue
            model_state_trial = dict(model=model0, tokenizer=tokenizer0, device=device)
            model_state_trial.update(model_dict)
            assert len(model_state_none) == len(model_state_trial)
            print("Model %s" % model_dict, flush=True)
            if model_lock:
                # last in iteration will be first
                model_states.insert(0, model_state_trial)
                # fill model_state0 so go_gradio() easier, manage model_states separately
                model_state0 = model_state_trial.copy()
            else:
                model_state0 = model_state_trial.copy()
            assert len(model_state_none) == len(model_state0)

        # get score model
        all_kwargs = locals().copy()
        smodel, stokenizer, sdevice = get_score_model(reward_type=True,
                                                      **get_kwargs(get_score_model, exclude_names=['reward_type'],
                                                                   **all_kwargs))
        score_model_state0 = dict(model=smodel, tokenizer=stokenizer, device=sdevice,
                                  base_model=score_model, tokenizer_base_model='', lora_weights='',
                                  inference_server='', prompt_type='', prompt_dict='')

        if enable_captions:
            if pre_load_caption_model:
                from image_captions import H2OImageCaptionLoader
                caption_loader = H2OImageCaptionLoader(caption_gpu=caption_gpu).load_model()
            else:
                caption_loader = 'gpu' if caption_gpu else 'cpu'
        else:
            caption_loader = False

        # assume gradio needs everything
        go_gradio(**locals())


def get_config(base_model,
               use_auth_token=False,
               trust_remote_code=True,
               offload_folder=None,
               triton_attn=False,
               long_sequence=True,
               return_model=False,
               raise_exception=False,
               ):
    from accelerate import init_empty_weights
    with init_empty_weights():
        from transformers import AutoConfig
        try:
            config = AutoConfig.from_pretrained(base_model, use_auth_token=use_auth_token,
                                                trust_remote_code=trust_remote_code,
                                                offload_folder=offload_folder)
        except OSError as e:
            if raise_exception:
                raise
            if 'not a local folder and is not a valid model identifier listed on' in str(
                    e) or '404 Client Error' in str(e):
                # e.g. llama, gpjt, etc.
                # e.g. HF TGI but not model on HF or private etc.
                # HF TGI server only should really require prompt_type, not HF model state
                return None, None
            else:
                raise
        if triton_attn and 'mpt-' in base_model.lower():
            config.attn_config['attn_impl'] = 'triton'
        if long_sequence:
            if 'mpt-7b-storywriter' in base_model.lower():
                config.update({"max_seq_len": 83968})
            if 'mosaicml/mpt-7b-chat' in base_model.lower():
                config.update({"max_seq_len": 4096})
            if 'mpt-30b' in base_model.lower():
                config.update({"max_seq_len": 2 * 8192})
        if return_model and \
                issubclass(config.__class__, tuple(AutoModel._model_mapping.keys())):
            model = AutoModel.from_config(
                config,
                trust_remote_code=trust_remote_code,
            )
        else:
            # can't infer
            model = None
    if 'falcon' in base_model.lower():
        config.use_cache = False

    return config, model


def get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type,
                       config, model,
                       gpu_id=0,
                       ):
    """
    Ensure model gets on correct device
    """

    if model is not None:
        # NOTE: Can specify max_memory={0: max_mem, 1: max_mem}, to shard model
        # NOTE: Some models require avoiding sharding some layers,
        # then would pass no_split_module_classes and give list of those layers.
        from accelerate import infer_auto_device_map
        device_map = infer_auto_device_map(
            model,
            dtype=torch.float16 if load_half else torch.float32,
        )
        if hasattr(model, 'model'):
            device_map_model = infer_auto_device_map(
                model.model,
                dtype=torch.float16 if load_half else torch.float32,
            )
            device_map.update(device_map_model)
    else:
        device_map = "auto"

    n_gpus = torch.cuda.device_count() if torch.cuda.is_available else 0

    if n_gpus > 0:
        if gpu_id >= 0:
            # FIXME: If really distributes model, tend to get things like: ValueError: gpt_neox.embed_in.weight doesn't have any device set.
            # So avoid for now, just put on first GPU, unless score_model, put on last
            if reward_type:
                device_map = {'': n_gpus - 1}
            else:
                device_map = {'': min(n_gpus - 1, gpu_id)}
        if gpu_id == -1:
            device_map = {'': 'cuda'}
    else:
        device_map = {'': 'cpu'}
        model_kwargs['load_in_8bit'] = False
        model_kwargs['load_in_4bit'] = False
    print('device_map: %s' % device_map, flush=True)

    load_in_8bit = model_kwargs.get('load_in_8bit', False)
    load_in_4bit = model_kwargs.get('load_in_4bit', False)
    model_kwargs['device_map'] = device_map
    pop_unused_model_kwargs(model_kwargs)

    if load_in_8bit or load_in_4bit or not load_half:
        model = model_loader.from_pretrained(
            base_model,
            config=config,
            **model_kwargs,
        )
    else:
        model = model_loader.from_pretrained(
            base_model,
            config=config,
            **model_kwargs,
        ).half()
    return model


def get_client_from_inference_server(inference_server, raise_connection_exception=False):
    inference_server, headers = get_hf_server(inference_server)
    # preload client since slow for gradio case especially
    from gradio_utils.grclient import GradioClient
    gr_client = None
    hf_client = None
    if headers is None:
        try:
            print("GR Client Begin: %s" % inference_server, flush=True)
            # first do sanity check if alive, else gradio client takes too long by default
            requests.get(inference_server, timeout=int(os.getenv('REQUEST_TIMEOUT', '30')))
            gr_client = GradioClient(inference_server)
            print("GR Client End: %s" % inference_server, flush=True)
        except (OSError, ValueError) as e:
            # Occurs when wrong endpoint and should have been HF client, so don't hard raise, just move to HF
            gr_client = None
            print("GR Client Failed %s: %s" % (inference_server, str(e)), flush=True)
        except (ConnectTimeoutError, ConnectTimeout, MaxRetryError, ConnectionError, ConnectionError2,
                JSONDecodeError, ReadTimeout2, KeyError) as e:
            t, v, tb = sys.exc_info()
            ex = ''.join(traceback.format_exception(t, v, tb))
            print("GR Client Failed %s: %s" % (inference_server, str(ex)), flush=True)
            if raise_connection_exception:
                raise

    if gr_client is None:
        res = None
        from text_generation import Client as HFClient
        print("HF Client Begin: %s" % inference_server)
        try:
            hf_client = HFClient(inference_server, headers=headers, timeout=int(os.getenv('REQUEST_TIMEOUT', '30')))
            # quick check valid TGI endpoint
            res = hf_client.generate('What?', max_new_tokens=1)
            hf_client = HFClient(inference_server, headers=headers, timeout=300)
        except (ConnectTimeoutError, ConnectTimeout, MaxRetryError, ConnectionError, ConnectionError2,
                JSONDecodeError, ReadTimeout2, KeyError) as e:
            hf_client = None
            t, v, tb = sys.exc_info()
            ex = ''.join(traceback.format_exception(t, v, tb))
            print("HF Client Failed %s: %s" % (inference_server, str(ex)))
            if raise_connection_exception:
                raise
        print("HF Client End: %s %s" % (inference_server, res))
    return inference_server, gr_client, hf_client


def get_model(
        load_8bit: bool = False,
        load_4bit: bool = False,
        load_half: bool = True,
        infer_devices: bool = True,
        base_model: str = '',
        inference_server: str = "",
        tokenizer_base_model: str = '',
        lora_weights: str = "",
        gpu_id: int = 0,

        reward_type: bool = None,
        local_files_only: bool = False,
        resume_download: bool = True,
        use_auth_token: Union[str, bool] = False,
        trust_remote_code: bool = True,
        offload_folder: str = None,
        compile_model: bool = True,

        verbose: bool = False,
):
    """

    :param load_8bit: load model in 8-bit, not supported by all models
    :param load_4bit: load model in 4-bit, not supported by all models
    :param load_half: load model in 16-bit
    :param infer_devices: Use torch infer of optimal placement of layers on devices (for non-lora case)
           For non-LORA case, False will spread shards across multiple GPUs, but this can lead to cuda:x cuda:y mismatches
           So it is not the default
    :param base_model: name/path of base model
    :param inference_server: whether base_model is hosted locally ('') or via http (url)
    :param tokenizer_base_model: name/path of tokenizer
    :param lora_weights: name/path
    :param gpu_id: which GPU (0..n_gpus-1) or allow all GPUs if relevant (-1)
    :param reward_type: reward type model for sequence classification
    :param local_files_only: use local files instead of from HF
    :param resume_download: resume downloads from HF
    :param use_auth_token: assumes user did on CLI `huggingface-cli login` to access private repo
    :param trust_remote_code: trust code needed by model
    :param offload_folder: offload folder
    :param compile_model: whether to compile torch model
    :param verbose:
    :return:
    """
    if verbose:
        print("Get %s model" % base_model, flush=True)

    triton_attn = False
    long_sequence = True
    config_kwargs = dict(use_auth_token=use_auth_token,
                         trust_remote_code=trust_remote_code,
                         offload_folder=offload_folder,
                         triton_attn=triton_attn,
                         long_sequence=long_sequence)
    config, _ = get_config(base_model, **config_kwargs, raise_exception=False)

    if base_model in non_hf_types:
        assert config is None, "Expected config None for %s" % base_model

    llama_type_from_config = 'llama' in str(config).lower()
    llama_type_from_name = "llama" in base_model.lower()
    llama_type = llama_type_from_config or llama_type_from_name
    if "xgen" in base_model.lower():
        llama_type = False
    if llama_type:
        if verbose:
            print("Detected as llama type from"
                  " config (%s) or name (%s)" % (llama_type_from_config, llama_type_from_name), flush=True)

    model_loader, tokenizer_loader = get_loaders(model_name=base_model, reward_type=reward_type, llama_type=llama_type)

    tokenizer_kwargs = dict(local_files_only=local_files_only,
                            resume_download=resume_download,
                            use_auth_token=use_auth_token,
                            trust_remote_code=trust_remote_code,
                            offload_folder=offload_folder,
                            padding_side='left',
                            config=config,
                            )
    if not tokenizer_base_model:
        tokenizer_base_model = base_model

    if config is not None and tokenizer_loader is not None and not isinstance(tokenizer_loader, str):
        tokenizer = tokenizer_loader.from_pretrained(tokenizer_base_model, **tokenizer_kwargs)
        # sets raw (no cushion) limit
        set_model_max_len(config, tokenizer, verbose=False)
        # if using fake tokenizer, not really accurate when lots of numbers, give a bit of buffer, else get:
        # Generation Failed: Input validation error: `inputs` must have less than 2048 tokens. Given: 2233
        tokenizer.model_max_length = tokenizer.model_max_length - 50
    else:
        tokenizer = FakeTokenizer()

    if isinstance(inference_server, str) and inference_server.startswith("http"):
        inference_server, gr_client, hf_client = get_client_from_inference_server(inference_server)
        client = gr_client or hf_client
        # Don't return None, None for model, tokenizer so triggers
        return client, tokenizer, 'http'
    if isinstance(inference_server, str) and inference_server.startswith('openai'):
        assert os.getenv('OPENAI_API_KEY'), "Set environment for OPENAI_API_KEY"
        # Don't return None, None for model, tokenizer so triggers
        # include small token cushion
        tokenizer = FakeTokenizer(model_max_length=model_token_mapping[base_model] - 50)
        return inference_server, tokenizer, inference_server
    assert not inference_server, "Malformed inference_server=%s" % inference_server
    if base_model in non_hf_types:
        from gpt4all_llm import get_model_tokenizer_gpt4all
        model, tokenizer, device = get_model_tokenizer_gpt4all(base_model)
        return model, tokenizer, device

    # get local torch-HF model
    return get_hf_model(load_8bit=load_8bit,
                        load_4bit=load_4bit,
                        load_half=load_half,
                        infer_devices=infer_devices,
                        base_model=base_model,
                        tokenizer_base_model=tokenizer_base_model,
                        lora_weights=lora_weights,
                        gpu_id=gpu_id,

                        reward_type=reward_type,
                        local_files_only=local_files_only,
                        resume_download=resume_download,
                        use_auth_token=use_auth_token,
                        trust_remote_code=trust_remote_code,
                        offload_folder=offload_folder,
                        compile_model=compile_model,

                        llama_type=llama_type,
                        config_kwargs=config_kwargs,
                        tokenizer_kwargs=tokenizer_kwargs,

                        verbose=verbose)


def get_hf_model(load_8bit: bool = False,
                 load_4bit: bool = False,
                 load_half: bool = True,
                 infer_devices: bool = True,
                 base_model: str = '',
                 tokenizer_base_model: str = '',
                 lora_weights: str = "",
                 gpu_id: int = 0,

                 reward_type: bool = None,
                 local_files_only: bool = False,
                 resume_download: bool = True,
                 use_auth_token: Union[str, bool] = False,
                 trust_remote_code: bool = True,
                 offload_folder: str = None,
                 compile_model: bool = True,

                 llama_type: bool = False,
                 config_kwargs=None,
                 tokenizer_kwargs=None,

                 verbose: bool = False,
                 ):
    assert config_kwargs is not None
    assert tokenizer_kwargs is not None

    if lora_weights is not None and lora_weights.strip():
        if verbose:
            print("Get %s lora weights" % lora_weights, flush=True)
    device = get_device()

    if 'gpt2' in base_model.lower():
        # RuntimeError: where expected condition to be a boolean tensor, but got a tensor with dtype Half
        load_8bit = False
        load_4bit = False

    assert base_model.strip(), (
        "Please choose a base model with --base_model (CLI) or load one from Models Tab (gradio)"
    )

    model_loader, tokenizer_loader = get_loaders(model_name=base_model, reward_type=reward_type, llama_type=llama_type)

    config, _ = get_config(base_model, return_model=False, raise_exception=True, **config_kwargs)

    if tokenizer_loader is not None and not isinstance(tokenizer_loader, str):
        tokenizer = tokenizer_loader.from_pretrained(tokenizer_base_model,
                                                     **tokenizer_kwargs)
    else:
        tokenizer = tokenizer_loader

    if isinstance(tokenizer, str):
        # already a pipeline, tokenizer_loader is string for task
        model = model_loader(tokenizer,
                             model=base_model,
                             device=0 if device == "cuda" else -1,
                             torch_dtype=torch.float16 if device == 'cuda' else torch.float32)
    else:
        assert device in ["cuda", "cpu"], "Unsupported device %s" % device
        model_kwargs = dict(local_files_only=local_files_only,
                            torch_dtype=torch.float16 if device == 'cuda' else torch.float32,
                            resume_download=resume_download,
                            use_auth_token=use_auth_token,
                            trust_remote_code=trust_remote_code,
                            offload_folder=offload_folder,
                            )
        if 'mbart-' not in base_model.lower() and 'mpt-' not in base_model.lower():
            model_kwargs.update(dict(load_in_8bit=load_8bit,
                                     load_in_4bit=load_4bit,
                                     device_map={"": 0} if (load_8bit or load_4bit) and device == 'cuda' else "auto",
                                     ))
        if 'mpt-' in base_model.lower() and gpu_id is not None and gpu_id >= 0:
            model_kwargs.update(dict(device_map={"": gpu_id} if device == 'cuda' else "cpu"))

        if 'OpenAssistant/reward-model'.lower() in base_model.lower():
            # FIXME: could put on other GPUs
            model_kwargs['device_map'] = {"": 0} if device == 'cuda' else {"": 'cpu'}
            model_kwargs.pop('torch_dtype', None)
        pop_unused_model_kwargs(model_kwargs)

        if not lora_weights:
            with torch.device(device):

                if infer_devices:
                    config, model = get_config(base_model, return_model=True, raise_exception=True, **config_kwargs)
                    model = get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type,
                                               config, model,
                                               gpu_id=gpu_id,
                                               )
                else:
                    config, _ = get_config(base_model, **config_kwargs)
                    if load_half and not (load_8bit or load_4bit):
                        model = model_loader.from_pretrained(
                            base_model,
                            config=config,
                            **model_kwargs).half()
                    else:
                        model = model_loader.from_pretrained(
                            base_model,
                            config=config,
                            **model_kwargs)
        elif load_8bit or load_4bit:
            config, _ = get_config(base_model, **config_kwargs)
            model = model_loader.from_pretrained(
                base_model,
                config=config,
                **model_kwargs
            )
            from peft import PeftModel  # loads cuda, so avoid in global scope
            model = PeftModel.from_pretrained(
                model,
                lora_weights,
                torch_dtype=torch.float16 if device == 'cuda' else torch.float32,
                local_files_only=local_files_only,
                resume_download=resume_download,
                use_auth_token=use_auth_token,
                trust_remote_code=trust_remote_code,
                offload_folder=offload_folder,
                device_map={"": 0} if device == 'cuda' else {"": 'cpu'},  # seems to be required
            )
        else:
            with torch.device(device):
                config, _ = get_config(base_model, raise_exception=True, **config_kwargs)
                model = model_loader.from_pretrained(
                    base_model,
                    config=config,
                    **model_kwargs
                )
                from peft import PeftModel  # loads cuda, so avoid in global scope
                model = PeftModel.from_pretrained(
                    model,
                    lora_weights,
                    torch_dtype=torch.float16 if device == 'cuda' else torch.float32,
                    local_files_only=local_files_only,
                    resume_download=resume_download,
                    use_auth_token=use_auth_token,
                    trust_remote_code=trust_remote_code,
                    offload_folder=offload_folder,
                    device_map="auto",
                )
                if load_half:
                    model.half()

    # unwind broken decapoda-research config
    if llama_type:
        model.config.pad_token_id = tokenizer.pad_token_id = 0  # unk
        model.config.bos_token_id = 1
        model.config.eos_token_id = 2
    if 'gpt2' in base_model.lower():
        # add special tokens that otherwise all share the same id
        tokenizer.add_special_tokens({'bos_token': '<bos>',
                                      'eos_token': '<eos>',
                                      'pad_token': '<pad>'})

    if not isinstance(tokenizer, str):
        model.eval()
        if torch.__version__ >= "2" and sys.platform != "win32" and compile_model:
            model = torch.compile(model)

    set_model_max_len(config, tokenizer, verbose=False, reward_type=reward_type)

    return model, tokenizer, device


def set_model_max_len(config, tokenizer, verbose=False, reward_type=False):
    if reward_type:
        # limit deberta, else uses too much memory and not worth response score
        tokenizer.model_max_length = 512
    if hasattr(config, 'max_seq_len') and isinstance(config.max_seq_len, int):
        tokenizer.model_max_length = config.max_seq_len
    elif hasattr(config, 'max_position_embeddings') and isinstance(config.max_position_embeddings, int):
        # help automatically limit inputs to generate
        tokenizer.model_max_length = config.max_position_embeddings
    else:
        if verbose:
            print("Could not determine model_max_length, setting to 2048", flush=True)
        tokenizer.model_max_length = 2048
    # for bug in HF transformers
    if tokenizer.model_max_length > 100000000:
        tokenizer.model_max_length = 2048


def pop_unused_model_kwargs(model_kwargs):
    """
    in-place pop unused kwargs that are not dependency-upgrade friendly
    no point passing in False, is default, and helps avoid needing to update requirements for new deps
    :param model_kwargs:
    :return:
    """
    check_list = ['load_in_8bit', 'load_in_4bit']
    for k in check_list:
        if k in model_kwargs and not model_kwargs[k]:
            model_kwargs.pop(k)


def get_score_model(score_model: str = None,
                    load_8bit: bool = False,
                    load_4bit: bool = False,
                    load_half: bool = True,
                    infer_devices: bool = True,
                    base_model: str = '',
                    inference_server: str = '',
                    tokenizer_base_model: str = '',
                    lora_weights: str = "",
                    gpu_id: int = 0,

                    reward_type: bool = None,
                    local_files_only: bool = False,
                    resume_download: bool = True,
                    use_auth_token: Union[str, bool] = False,
                    trust_remote_code: bool = True,
                    offload_folder: str = None,
                    compile_model: bool = True,

                    verbose: bool = False,
                    ):
    if score_model is not None and score_model.strip():
        load_8bit = False
        load_4bit = False
        load_half = False
        base_model = score_model.strip()
        tokenizer_base_model = ''
        lora_weights = ''
        inference_server = ''
        llama_type = False
        compile_model = False
        smodel, stokenizer, sdevice = get_model(reward_type=True,
                                                **get_kwargs(get_model, exclude_names=['reward_type'], **locals()))
    else:
        smodel, stokenizer, sdevice = None, None, None
    return smodel, stokenizer, sdevice


no_default_param_names = [
    'instruction',
    'iinput',
    'context',
    'instruction_nochat',
    'iinput_nochat',
]

gen_hyper = ['temperature',
             'top_p',
             'top_k',
             'num_beams',
             'max_new_tokens',
             'min_new_tokens',
             'early_stopping',
             'max_time',
             'repetition_penalty',
             'num_return_sequences',
             'do_sample',
             ]

eval_func_param_names = ['instruction',
                         'iinput',
                         'context',
                         'stream_output',
                         'prompt_type',
                         'prompt_dict'] + \
                        gen_hyper + \
                        ['chat',
                         'instruction_nochat',
                         'iinput_nochat',
                         'langchain_mode',
                         'top_k_docs',
                         'chunk',
                         'chunk_size',
                         'document_choice',
                         ]

# form evaluate defaults for submit_nochat_api
eval_func_param_names_defaults = eval_func_param_names.copy()
for k in no_default_param_names:
    if k in eval_func_param_names_defaults:
        eval_func_param_names_defaults.remove(k)


def evaluate_from_str(
        model_state,
        my_db_state,
        # START NOTE: Examples must have same order of parameters
        user_kwargs,
        # END NOTE: Examples must have same order of parameters
        default_kwargs=None,
        src_lang=None,
        tgt_lang=None,
        debug=False,
        concurrency_count=None,
        save_dir=None,
        sanitize_bot_response=False,
        model_state0=None,
        memory_restriction_level=None,
        max_max_new_tokens=None,
        is_public=None,
        max_max_time=None,
        raise_generate_gpu_exceptions=None,
        chat_context=None,
        lora_weights=None,
        load_db_if_exists=True,
        dbs=None,
        user_path=None,
        detect_user_path_changes_every_query=None,
        use_openai_embedding=None,
        use_openai_model=None,
        hf_embedding_model=None,
        db_type=None,
        n_jobs=None,
        first_para=None,
        text_limit=None,
        verbose=False,
        cli=False,
        reverse_docs=True,
        use_cache=None,
        auto_reduce_chunks=None,
        max_chunks=None,
        model_lock=None,
        force_langchain_evaluate=None,
        model_state_none=None,
):
    if isinstance(user_kwargs, str):
        user_kwargs = ast.literal_eval(user_kwargs)
    # only used for submit_nochat_api
    user_kwargs['chat'] = False
    if 'stream_output' not in user_kwargs:
        user_kwargs['stream_output'] = False
    if 'langchain_mode' not in user_kwargs:
        # if user doesn't specify, then assume disabled, not use default
        user_kwargs['langchain_mode'] = 'Disabled'

    assert set(list(default_kwargs.keys())) == set(eval_func_param_names)
    # correct ordering.  Note some things may not be in default_kwargs, so can't be default of user_kwargs.get()
    args_list = [user_kwargs[k] if k in user_kwargs else default_kwargs[k] for k in eval_func_param_names]

    ret = evaluate(
        model_state,
        my_db_state,
        # START NOTE: Examples must have same order of parameters
        *tuple(args_list),
        # END NOTE: Examples must have same order of parameters
        src_lang=src_lang,
        tgt_lang=tgt_lang,
        debug=debug,
        concurrency_count=concurrency_count,
        save_dir=save_dir,
        sanitize_bot_response=sanitize_bot_response,
        model_state0=model_state0,
        memory_restriction_level=memory_restriction_level,
        max_max_new_tokens=max_max_new_tokens,
        is_public=is_public,
        max_max_time=max_max_time,
        raise_generate_gpu_exceptions=raise_generate_gpu_exceptions,
        chat_context=chat_context,
        lora_weights=lora_weights,
        load_db_if_exists=load_db_if_exists,
        dbs=dbs,
        user_path=user_path,
        detect_user_path_changes_every_query=detect_user_path_changes_every_query,
        use_openai_embedding=use_openai_embedding,
        use_openai_model=use_openai_model,
        hf_embedding_model=hf_embedding_model,
        db_type=db_type,
        n_jobs=n_jobs,
        first_para=first_para,
        text_limit=text_limit,
        verbose=verbose,
        cli=cli,
        reverse_docs=reverse_docs,
        use_cache=use_cache,
        auto_reduce_chunks=auto_reduce_chunks,
        max_chunks=max_chunks,
        model_lock=model_lock,
        force_langchain_evaluate=force_langchain_evaluate,
        model_state_none=model_state_none,
    )
    try:
        for ret1 in ret:
            yield ret1
    finally:
        # clear before return, in finally in case GPU OOM exception
        clear_torch_cache()


def evaluate(
        model_state,
        my_db_state,
        # START NOTE: Examples must have same order of parameters
        instruction,
        iinput,
        context,
        stream_output,
        prompt_type,
        prompt_dict,
        temperature,
        top_p,
        top_k,
        num_beams,
        max_new_tokens,
        min_new_tokens,
        early_stopping,
        max_time,
        repetition_penalty,
        num_return_sequences,
        do_sample,
        chat,
        instruction_nochat,
        iinput_nochat,
        langchain_mode,
        top_k_docs,
        chunk,
        chunk_size,
        document_choice,
        # END NOTE: Examples must have same order of parameters
        src_lang=None,
        tgt_lang=None,
        debug=False,
        concurrency_count=None,
        save_dir=None,
        sanitize_bot_response=False,
        model_state0=None,
        memory_restriction_level=None,
        max_max_new_tokens=None,
        is_public=None,
        max_max_time=None,
        raise_generate_gpu_exceptions=None,
        chat_context=None,
        lora_weights=None,
        load_db_if_exists=True,
        dbs=None,
        user_path=None,
        detect_user_path_changes_every_query=None,
        use_openai_embedding=None,
        use_openai_model=None,
        hf_embedding_model=None,
        db_type=None,
        n_jobs=None,
        first_para=None,
        text_limit=None,
        verbose=False,
        cli=False,
        reverse_docs=True,
        use_cache=None,
        auto_reduce_chunks=None,
        max_chunks=None,
        model_lock=None,
        force_langchain_evaluate=None,
        model_state_none=None,
):
    # ensure passed these
    assert concurrency_count is not None
    assert memory_restriction_level is not None
    assert raise_generate_gpu_exceptions is not None
    assert chat_context is not None
    assert use_openai_embedding is not None
    assert use_openai_model is not None
    assert hf_embedding_model is not None
    assert db_type is not None
    assert top_k_docs is not None and isinstance(top_k_docs, int)
    assert chunk is not None and isinstance(chunk, bool)
    assert chunk_size is not None and isinstance(chunk_size, int)
    assert n_jobs is not None
    assert first_para is not None

    if debug:
        locals_dict = locals().copy()
        locals_dict.pop('model_state', None)
        locals_dict.pop('model_state0', None)
        locals_dict.pop('model_states', None)
        print(locals_dict)

    no_model_msg = "Please choose a base model with --base_model (CLI) or load in Models Tab (gradio).\n" \
                   "Then start New Conversation"

    if model_state is None:
        model_state = model_state_none.copy()
    if model_state0 is None:
        # e.g. for no gradio case, set dummy value, else should be set
        model_state0 = model_state_none.copy()

    # model_state['model] is only 'model' if should use model_state0
    # model could also be None
    have_model_lock = model_lock is not None
    have_fresh_model = model_state['model'] not in [None, 'model', no_model_str]
    # for gradio UI control, expect model_state and model_state0 to match, so if have_model_lock=True, then should have_fresh_model=True
    # but gradio API control will only use nochat api etc. and won't use fresh model, so can't assert in general
    # if have_model_lock:
    #    assert have_fresh_model, "Expected model_state and model_state0 to match if have_model_lock"
    have_cli_model = model_state0['model'] not in [None, 'model', no_model_str]

    if have_fresh_model:
        # USE FRESH MODEL
        if not have_model_lock:
            # model_state0 is just one of model_state if model_lock, so don't nuke
            # try to free-up original model (i.e. list was passed as reference)
            if model_state0['model'] and hasattr(model_state0['model'], 'cpu'):
                model_state0['model'].cpu()
                model_state0['model'] = None
            # try to free-up original tokenizer (i.e. list was passed as reference)
            if model_state0['tokenizer']:
                model_state0['tokenizer'] = None
            clear_torch_cache()
        chosen_model_state = model_state
    elif have_cli_model:
        # USE MODEL SETUP AT CLI
        assert isinstance(model_state['model'], str)  # expect no fresh model
        chosen_model_state = model_state0
    else:
        raise AssertionError(no_model_msg)
    # get variables
    model = chosen_model_state['model']
    tokenizer = chosen_model_state['tokenizer']
    device = chosen_model_state['device']
    base_model = chosen_model_state['base_model']
    tokenizer_base_model = chosen_model_state['tokenizer_base_model']
    lora_weights = chosen_model_state['lora_weights']
    inference_server = chosen_model_state['inference_server']
    # prefer use input from API over model state
    prompt_type = prompt_type or chosen_model_state['prompt_type']
    prompt_dict = prompt_dict or chosen_model_state['prompt_dict']

    if base_model is None:
        raise AssertionError(no_model_msg)

    assert base_model.strip(), no_model_msg
    assert model, "Model is missing"
    assert tokenizer, "Tokenizer is missing"

    # choose chat or non-chat mode
    if not chat:
        instruction = instruction_nochat
        iinput = iinput_nochat

    # in some cases, like lean nochat API, don't want to force sending prompt_type, allow default choice
    model_lower = base_model.lower()
    if not prompt_type and model_lower in inv_prompt_type_to_model_lower:
        prompt_type = inv_prompt_type_to_model_lower[model_lower]
        if verbose:
            print("Auto-selecting prompt_type=%s for %s" % (prompt_type, model_lower), flush=True)
    assert prompt_type is not None, "prompt_type was None"

    # Control generation hyperparameters
    # adjust for bad inputs, e.g. in case also come from API that doesn't get constrained by gradio sliders
    # below is for TGI server, not required for HF transformers
    # limits are chosen similar to gradio_runner.py sliders/numbers
    top_p = min(max(1e-3, top_p), 1.0 - 1e-3)
    top_k = min(max(1, int(top_k)), 100)
    temperature = min(max(0.01, temperature), 2.0)
    # FIXME: https://github.com/h2oai/h2ogpt/issues/106
    num_beams = 1 if stream_output else num_beams  # See max_beams in gradio_runner
    max_max_new_tokens = get_max_max_new_tokens(chosen_model_state,
                                                memory_restriction_level=memory_restriction_level,
                                                max_new_tokens=max_new_tokens,
                                                max_max_new_tokens=max_max_new_tokens)
    model_max_length = get_model_max_length(chosen_model_state)
    max_new_tokens = min(max(1, int(max_new_tokens)), max_max_new_tokens)
    min_new_tokens = min(max(0, int(min_new_tokens)), max_new_tokens)
    max_time = min(max(0, max_time), max_max_time)
    repetition_penalty = min(max(0.01, repetition_penalty), 3.0)
    num_return_sequences = 1 if chat else min(max(1, int(num_return_sequences)), 10)
    min_top_k_docs, max_top_k_docs, label_top_k_docs = get_minmax_top_k_docs(is_public)
    top_k_docs = min(max(min_top_k_docs, int(top_k_docs)), max_top_k_docs)
    chunk_size = min(max(128, int(chunk_size)), 2048)
    if not context:
        # get hidden context if have one
        context = get_context(chat_context, prompt_type)

    # restrict instruction, typically what has large input
    from h2oai_pipeline import H2OTextGenerationPipeline
    instruction, num_prompt_tokens1 = H2OTextGenerationPipeline.limit_prompt(instruction, tokenizer)
    context, num_prompt_tokens2 = H2OTextGenerationPipeline.limit_prompt(context, tokenizer)
    iinput, num_prompt_tokens3 = H2OTextGenerationPipeline.limit_prompt(iinput, tokenizer)
    num_prompt_tokens = (num_prompt_tokens1 or 0) + (num_prompt_tokens2 or 0) + (num_prompt_tokens3 or 0)

    # get prompt
    prompter = Prompter(prompt_type, prompt_dict, debug=debug, chat=chat, stream_output=stream_output)
    data_point = dict(context=context, instruction=instruction, input=iinput)
    prompt = prompter.generate_prompt(data_point)

    # THIRD PLACE where LangChain referenced, but imports only occur if enabled and have db to use
    assert langchain_mode in langchain_modes, "Invalid langchain_mode %s" % langchain_mode
    if langchain_mode in ['MyData'] and my_db_state is not None and len(my_db_state) > 0 and my_db_state[0] is not None:
        db1 = my_db_state[0]
    elif dbs is not None and langchain_mode in dbs:
        db1 = dbs[langchain_mode]
    else:
        db1 = None
    do_langchain_path = langchain_mode not in [False, 'Disabled', 'ChatLLM', 'LLM'] and \
                        db1 is not None or \
                        base_model in non_hf_types or \
                        force_langchain_evaluate
    if do_langchain_path:
        query = instruction if not iinput else "%s\n%s" % (instruction, iinput)
        outr = ""
        # use smaller cut_distanct for wiki_full since so many matches could be obtained, and often irrelevant unless close
        from gpt_langchain import run_qa_db
        gen_hyper_langchain = dict(do_sample=do_sample,
                                   temperature=temperature,
                                   repetition_penalty=repetition_penalty,
                                   top_k=top_k,
                                   top_p=top_p,
                                   num_beams=num_beams,
                                   min_new_tokens=min_new_tokens,
                                   max_new_tokens=max_new_tokens,
                                   early_stopping=early_stopping,
                                   max_time=max_time,
                                   num_return_sequences=num_return_sequences,
                                   )
        for r in run_qa_db(query=query,
                           model_name=base_model, model=model, tokenizer=tokenizer,
                           inference_server=inference_server,
                           stream_output=stream_output,
                           prompter=prompter,
                           load_db_if_exists=load_db_if_exists,
                           db=db1,
                           user_path=user_path,
                           detect_user_path_changes_every_query=detect_user_path_changes_every_query,
                           cut_distanct=1.1 if langchain_mode in ['wiki_full'] else 1.64,  # FIXME, too arbitrary
                           use_openai_embedding=use_openai_embedding,
                           use_openai_model=use_openai_model,
                           hf_embedding_model=hf_embedding_model,
                           first_para=first_para,
                           text_limit=text_limit,
                           chunk=chunk,
                           chunk_size=chunk_size,
                           langchain_mode=langchain_mode,
                           document_choice=document_choice,
                           db_type=db_type,
                           top_k_docs=top_k_docs,

                           **gen_hyper_langchain,

                           prompt_type=prompt_type,
                           prompt_dict=prompt_dict,
                           n_jobs=n_jobs,
                           verbose=verbose,
                           cli=cli,
                           sanitize_bot_response=sanitize_bot_response,
                           reverse_docs=reverse_docs,

                           lora_weights=lora_weights,

                           auto_reduce_chunks=auto_reduce_chunks,
                           max_chunks=max_chunks,
                           ):
            outr, extra = r  # doesn't accumulate, new answer every yield, so only save that full answer
            yield dict(response=outr, sources=extra)
        if save_dir:
            extra_dict = gen_hyper_langchain.copy()
            extra_dict.update(prompt_type=prompt_type, inference_server=inference_server,
                              langchain_mode=langchain_mode, document_choice=document_choice,
                              num_prompt_tokens=num_prompt_tokens)
            save_generate_output(prompt=query, output=outr, base_model=base_model, save_dir=save_dir,
                                 where_from='run_qa_db',
                                 extra_dict=extra_dict)
            if verbose:
                print(
                    'Post-Generate Langchain: %s decoded_output: %s' % (str(datetime.now()), len(outr) if outr else -1),
                    flush=True)
        if outr or base_model in non_hf_types:
            # if got no response (e.g. not showing sources and got no sources,
            # so nothing to give to LLM), then slip through and ask LLM
            # Or if llama/gptj, then just return since they had no response and can't go down below code path
            # clear before return, since .then() never done if from API
            clear_torch_cache()
            return

    if inference_server.startswith('openai') or inference_server.startswith('http'):
        if inference_server.startswith('openai'):
            import openai
            where_from = "openai_client"

            openai.api_key = os.getenv("OPENAI_API_KEY")
            stop_sequences = list(set(prompter.terminate_response + [prompter.PreResponse]))
            # OpenAI will complain if ask for too many new tokens, takes it as min in some sense, wrongly so.
            max_new_tokens_openai = min(max_new_tokens, model_max_length - num_prompt_tokens)
            gen_server_kwargs = dict(temperature=temperature if do_sample else 0,
                                     max_tokens=max_new_tokens_openai,
                                     top_p=top_p if do_sample else 1,
                                     frequency_penalty=0,
                                     n=num_return_sequences,
                                     presence_penalty=1.07 - repetition_penalty + 0.6,  # so good default
                                     )
            if inference_server == 'openai':
                response = openai.Completion.create(
                    model=base_model,
                    prompt=prompt,
                    **gen_server_kwargs,
                    stop=stop_sequences,
                    stream=stream_output,
                )
                if not stream_output:
                    text = response['choices'][0]['text']
                    yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
                                                              sanitize_bot_response=sanitize_bot_response),
                               sources='')
                else:
                    collected_events = []
                    text = ''
                    for event in response:
                        collected_events.append(event)  # save the event response
                        event_text = event['choices'][0]['text']  # extract the text
                        text += event_text  # append the text
                        yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
                                                                  sanitize_bot_response=sanitize_bot_response),
                                   sources='')
            elif inference_server == 'openai_chat':
                response = openai.ChatCompletion.create(
                    model=base_model,
                    messages=[
                        {"role": "system", "content": "You are a helpful assistant."},
                        {'role': 'user',
                         'content': prompt,
                         }
                    ],
                    stream=stream_output,
                    **gen_server_kwargs,
                )
                if not stream_output:
                    text = response["choices"][0]["message"]["content"]
                    yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
                                                              sanitize_bot_response=sanitize_bot_response),
                               sources='')
                else:
                    text = ""
                    for chunk in response:
                        delta = chunk["choices"][0]["delta"]
                        if 'content' in delta:
                            text += delta['content']
                            yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
                                                                      sanitize_bot_response=sanitize_bot_response),
                                       sources='')
            else:
                raise RuntimeError("No such OpenAI mode: %s" % inference_server)
        elif inference_server.startswith('http'):
            inference_server, headers = get_hf_server(inference_server)
            from gradio_utils.grclient import GradioClient
            from text_generation import Client as HFClient
            if isinstance(model, GradioClient):
                gr_client = model
                hf_client = None
            elif isinstance(model, HFClient):
                gr_client = None
                hf_client = model
            else:
                inference_server, gr_client, hf_client = get_client_from_inference_server(inference_server)

            # quick sanity check to avoid long timeouts, just see if can reach server
            requests.get(inference_server, timeout=int(os.getenv('REQUEST_TIMEOUT_FAST', '10')))

            if gr_client is not None:
                # Note: h2oGPT gradio server could handle input token size issues for prompt,
                # but best to handle here so send less data to server

                chat_client = False
                where_from = "gr_client"
                client_langchain_mode = 'Disabled'
                gen_server_kwargs = dict(temperature=temperature,
                                         top_p=top_p,
                                         top_k=top_k,
                                         num_beams=num_beams,
                                         max_new_tokens=max_new_tokens,
                                         min_new_tokens=min_new_tokens,
                                         early_stopping=early_stopping,
                                         max_time=max_time,
                                         repetition_penalty=repetition_penalty,
                                         num_return_sequences=num_return_sequences,
                                         do_sample=do_sample,
                                         chat=chat_client,
                                         )
                # account for gradio into gradio that handles prompting, avoid duplicating prompter prompt injection
                if prompt_type in [None, '', PromptType.plain.name, PromptType.plain.value,
                                   str(PromptType.plain.value)]:
                    # if our prompt is plain, assume either correct or gradio server knows different prompt type,
                    # so pass empty prompt_Type
                    gr_prompt_type = ''
                    gr_prompt_dict = ''
                    gr_prompt = prompt  # already prepared prompt
                    gr_context = ''
                    gr_iinput = ''
                else:
                    # if already have prompt_type that is not plain, None, or '', then already applied some prompting
                    #  But assume server can handle prompting, and need to avoid double-up.
                    #  Also assume server can do better job of using stopping.py to stop early, so avoid local prompting, let server handle
                    #  So avoid "prompt" and let gradio server reconstruct from prompt_type we passed
                    # Note it's ok that prompter.get_response() has prompt+text, prompt=prompt passed,
                    #  because just means extra processing and removal of prompt, but that has no human-bot prompting doesn't matter
                    #  since those won't appear
                    gr_context = context
                    gr_prompt = instruction
                    gr_iinput = iinput
                    gr_prompt_type = prompt_type
                    gr_prompt_dict = prompt_dict
                client_kwargs = dict(instruction=gr_prompt if chat_client else '',  # only for chat=True
                                     iinput=gr_iinput,  # only for chat=True
                                     context=gr_context,
                                     # streaming output is supported, loops over and outputs each generation in streaming mode
                                     # but leave stream_output=False for simple input/output mode
                                     stream_output=stream_output,

                                     **gen_server_kwargs,

                                     prompt_type=gr_prompt_type,
                                     prompt_dict=gr_prompt_dict,

                                     instruction_nochat=gr_prompt if not chat_client else '',
                                     iinput_nochat=gr_iinput,  # only for chat=False
                                     langchain_mode=client_langchain_mode,
                                     top_k_docs=top_k_docs,
                                     chunk=chunk,
                                     chunk_size=chunk_size,
                                     document_choice=[DocumentChoices.All_Relevant.name],
                                     )
                api_name = '/submit_nochat_api'  # NOTE: like submit_nochat but stable API for string dict passing
                if not stream_output:
                    res = gr_client.predict(str(dict(client_kwargs)), api_name=api_name)
                    res_dict = ast.literal_eval(res)
                    text = res_dict['response']
                    sources = res_dict['sources']
                    yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
                                                              sanitize_bot_response=sanitize_bot_response),
                               sources=sources)
                else:
                    job = gr_client.submit(str(dict(client_kwargs)), api_name=api_name)
                    text = ''
                    sources = ''
                    res_dict = dict(response=text, sources=sources)
                    while not job.done():
                        outputs_list = job.communicator.job.outputs
                        if outputs_list:
                            res = job.communicator.job.outputs[-1]
                            res_dict = ast.literal_eval(res)
                            text = res_dict['response']
                            sources = res_dict['sources']
                            if gr_prompt_type == 'plain':
                                # then gradio server passes back full prompt + text
                                prompt_and_text = text
                            else:
                                prompt_and_text = prompt + text
                            yield dict(response=prompter.get_response(prompt_and_text, prompt=prompt,
                                                                      sanitize_bot_response=sanitize_bot_response),
                                       sources=sources)
                        time.sleep(0.01)
                    # ensure get last output to avoid race
                    res_all = job.outputs()
                    if len(res_all) > 0:
                        res = res_all[-1]
                        res_dict = ast.literal_eval(res)
                        text = res_dict['response']
                        sources = res_dict['sources']
                    else:
                        # go with old text if last call didn't work
                        e = job.future._exception
                        if e is not None:
                            stre = str(e)
                            strex = ''.join(traceback.format_tb(e.__traceback__))
                        else:
                            stre = ''
                            strex = ''

                        print("Bad final response: %s %s %s %s %s: %s %s" % (base_model, inference_server,
                                                                             res_all, prompt, text, stre, strex),
                              flush=True)
                    if gr_prompt_type == 'plain':
                        # then gradio server passes back full prompt + text
                        prompt_and_text = text
                    else:
                        prompt_and_text = prompt + text
                    yield dict(response=prompter.get_response(prompt_and_text, prompt=prompt,
                                                              sanitize_bot_response=sanitize_bot_response),
                               sources=sources)
            elif hf_client:
                # HF inference server needs control over input tokens
                where_from = "hf_client"

                # prompt must include all human-bot like tokens, already added by prompt
                # https://github.com/huggingface/text-generation-inference/tree/main/clients/python#types
                stop_sequences = list(set(prompter.terminate_response + [prompter.PreResponse]))
                gen_server_kwargs = dict(do_sample=do_sample,
                                         max_new_tokens=max_new_tokens,
                                         # best_of=None,
                                         repetition_penalty=repetition_penalty,
                                         return_full_text=True,
                                         seed=SEED,
                                         stop_sequences=stop_sequences,
                                         temperature=temperature,
                                         top_k=top_k,
                                         top_p=top_p,
                                         # truncate=False,  # behaves oddly
                                         # typical_p=top_p,
                                         # watermark=False,
                                         # decoder_input_details=False,
                                         )
                # work-around for timeout at constructor time, will be issue if multi-threading,
                # so just do something reasonable or max_time if larger
                # lower bound because client is re-used if multi-threading
                hf_client.timeout = max(300, max_time)
                if not stream_output:
                    text = hf_client.generate(prompt, **gen_server_kwargs).generated_text
                    yield dict(response=prompter.get_response(text, prompt=prompt,
                                                              sanitize_bot_response=sanitize_bot_response),
                               sources='')
                else:
                    text = ""
                    for response in hf_client.generate_stream(prompt, **gen_server_kwargs):
                        if not response.token.special:
                            # stop_sequences
                            text_chunk = response.token.text
                            text += text_chunk
                            yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
                                                                      sanitize_bot_response=sanitize_bot_response),
                                       sources='')
            else:
                raise RuntimeError("Failed to get client: %s" % inference_server)
        else:
            raise RuntimeError("No such inference_server  %s" % inference_server)

        if save_dir and text:
            # save prompt + new text
            extra_dict = gen_server_kwargs.copy()
            extra_dict.update(dict(inference_server=inference_server, num_prompt_tokens=num_prompt_tokens))
            save_generate_output(prompt=prompt, output=text, base_model=base_model, save_dir=save_dir,
                                 where_from=where_from, extra_dict=extra_dict)
        return
    else:
        assert not inference_server, "inferene_server=%s not supported" % inference_server

    if isinstance(tokenizer, str):
        # pipeline
        if tokenizer == "summarization":
            key = 'summary_text'
        else:
            raise RuntimeError("No such task type %s" % tokenizer)
        # NOTE: uses max_length only
        yield dict(response=model(prompt, max_length=max_new_tokens)[0][key], sources='')

    if 'mbart-' in base_model.lower():
        assert src_lang is not None
        tokenizer.src_lang = languages_covered()[src_lang]

    stopping_criteria = get_stopping(prompt_type, prompt_dict, tokenizer, device,
                                     model_max_length=tokenizer.model_max_length)

    inputs = tokenizer(prompt, return_tensors="pt")
    if debug and len(inputs["input_ids"]) > 0:
        print('input_ids length', len(inputs["input_ids"][0]), flush=True)
    input_ids = inputs["input_ids"].to(device)
    # CRITICAL LIMIT else will fail
    max_max_tokens = tokenizer.model_max_length
    max_input_tokens = max_max_tokens - min_new_tokens
    # NOTE: Don't limit up front due to max_new_tokens, let go up to max or reach max_max_tokens in stopping.py
    input_ids = input_ids[:, -max_input_tokens:]
    # required for falcon if multiple threads or asyncio accesses to model during generation
    if use_cache is None:
        use_cache = False if 'falcon' in base_model else True
    gen_config_kwargs = dict(temperature=float(temperature),
                             top_p=float(top_p),
                             top_k=top_k,
                             num_beams=num_beams,
                             do_sample=do_sample,
                             repetition_penalty=float(repetition_penalty),
                             num_return_sequences=num_return_sequences,
                             renormalize_logits=True,
                             remove_invalid_values=True,
                             use_cache=use_cache,
                             )
    token_ids = ['eos_token_id', 'pad_token_id', 'bos_token_id', 'cls_token_id', 'sep_token_id']
    for token_id in token_ids:
        if hasattr(tokenizer, token_id) and getattr(tokenizer, token_id) is not None:
            gen_config_kwargs.update({token_id: getattr(tokenizer, token_id)})
    generation_config = GenerationConfig(**gen_config_kwargs)

    gen_kwargs = dict(input_ids=input_ids,
                      generation_config=generation_config,
                      return_dict_in_generate=True,
                      output_scores=True,
                      max_new_tokens=max_new_tokens,  # prompt + new
                      min_new_tokens=min_new_tokens,  # prompt + new
                      early_stopping=early_stopping,  # False, True, "never"
                      max_time=max_time,
                      stopping_criteria=stopping_criteria,
                      )
    if 'gpt2' in base_model.lower():
        gen_kwargs.update(dict(bos_token_id=tokenizer.bos_token_id, pad_token_id=tokenizer.eos_token_id))
    elif 'mbart-' in base_model.lower():
        assert tgt_lang is not None
        tgt_lang = languages_covered()[tgt_lang]
        gen_kwargs.update(dict(forced_bos_token_id=tokenizer.lang_code_to_id[tgt_lang]))
    else:
        token_ids = ['eos_token_id', 'bos_token_id', 'pad_token_id']
        for token_id in token_ids:
            if hasattr(tokenizer, token_id) and getattr(tokenizer, token_id) is not None:
                gen_kwargs.update({token_id: getattr(tokenizer, token_id)})

    decoder_kwargs = dict(skip_special_tokens=True,
                          clean_up_tokenization_spaces=True)

    decoder = functools.partial(tokenizer.decode,
                                **decoder_kwargs
                                )
    decoder_raw_kwargs = dict(skip_special_tokens=False,
                              clean_up_tokenization_spaces=True)

    decoder_raw = functools.partial(tokenizer.decode,
                                    **decoder_raw_kwargs
                                    )

    with torch.no_grad():
        have_lora_weights = lora_weights not in [no_lora_str, '', None]
        context_class_cast = NullContext if device == 'cpu' or have_lora_weights else torch.autocast
        with context_class_cast(device):
            # protection for gradio not keeping track of closed users,
            # else hit bitsandbytes lack of thread safety:
            # https://github.com/h2oai/h2ogpt/issues/104
            # but only makes sense if concurrency_count == 1
            context_class = NullContext  # if concurrency_count > 1 else filelock.FileLock
            if verbose:
                print('Pre-Generate: %s' % str(datetime.now()), flush=True)
            decoded_output = None
            with context_class("generate.lock"):
                if verbose:
                    print('Generate: %s' % str(datetime.now()), flush=True)
                # decoded tokenized prompt can deviate from prompt due to special characters
                inputs_decoded = decoder(input_ids[0])
                inputs_decoded_raw = decoder_raw(input_ids[0])
                if inputs_decoded == prompt:
                    # normal
                    pass
                elif inputs_decoded.lstrip() == prompt.lstrip():
                    # sometimes extra space in front, make prompt same for prompt removal
                    prompt = inputs_decoded
                elif inputs_decoded_raw == prompt:
                    # some models specify special tokens that are part of normal prompt, so can't skip them
                    inputs_decoded = prompt = inputs_decoded_raw
                    decoder = decoder_raw
                    decoder_kwargs = decoder_raw_kwargs
                elif inputs_decoded_raw.replace("<unk> ", "").replace("<unk>", "").replace('\n', ' ').replace(' ',
                                                                                                              '') == prompt.replace(
                    '\n', ' ').replace(' ', ''):
                    inputs_decoded = prompt = inputs_decoded_raw
                    decoder = decoder_raw
                    decoder_kwargs = decoder_raw_kwargs
                else:
                    if verbose:
                        print("WARNING: Special characters in prompt", flush=True)
                if stream_output:
                    skip_prompt = False
                    streamer = H2OTextIteratorStreamer(tokenizer, skip_prompt=skip_prompt, block=False,
                                                       **decoder_kwargs)
                    gen_kwargs.update(dict(streamer=streamer))
                    target = wrapped_partial(generate_with_exceptions, model.generate,
                                             prompt=prompt, inputs_decoded=inputs_decoded,
                                             raise_generate_gpu_exceptions=raise_generate_gpu_exceptions,
                                             **gen_kwargs)
                    bucket = queue.Queue()
                    thread = EThread(target=target, streamer=streamer, bucket=bucket)
                    thread.start()
                    outputs = ""
                    try:
                        for new_text in streamer:
                            if bucket.qsize() > 0 or thread.exc:
                                thread.join()
                            outputs += new_text
                            yield dict(response=prompter.get_response(outputs, prompt=inputs_decoded,
                                                                      sanitize_bot_response=sanitize_bot_response),
                                       sources='')
                    except BaseException:
                        # if any exception, raise that exception if was from thread, first
                        if thread.exc:
                            raise thread.exc
                        raise
                    finally:
                        # clear before return, since .then() never done if from API
                        clear_torch_cache()
                        # in case no exception and didn't join with thread yet, then join
                        if not thread.exc:
                            thread.join()
                    # in case raise StopIteration or broke queue loop in streamer, but still have exception
                    if thread.exc:
                        raise thread.exc
                    decoded_output = outputs
                else:
                    try:
                        outputs = model.generate(**gen_kwargs)
                    finally:
                        clear_torch_cache()  # has to be here for API submit_nochat_api since.then() not called
                    outputs = [decoder(s) for s in outputs.sequences]
                    yield dict(response=prompter.get_response(outputs, prompt=inputs_decoded,
                                                              sanitize_bot_response=sanitize_bot_response), sources='')
                    if outputs and len(outputs) >= 1:
                        decoded_output = prompt + outputs[0]
                if save_dir and decoded_output:
                    extra_dict = gen_config_kwargs.copy()
                    extra_dict.update(dict(num_prompt_tokens=num_prompt_tokens))
                    save_generate_output(prompt=prompt, output=decoded_output, base_model=base_model, save_dir=save_dir,
                                         where_from="evaluate_%s" % str(stream_output),
                                         extra_dict=gen_config_kwargs)
            if verbose:
                print('Post-Generate: %s decoded_output: %s' % (
                    str(datetime.now()), len(decoded_output) if decoded_output else -1), flush=True)


inputs_list_names = list(inspect.signature(evaluate).parameters)
state_names = ['model_state', 'my_db_state']
inputs_kwargs_list = [x for x in inputs_list_names if x not in eval_func_param_names + state_names]


def get_cutoffs(memory_restriction_level, for_context=False, model_max_length=2048):
    # help to avoid errors like:
    # RuntimeError: The size of tensor a (2048) must match the size of tensor b (2049) at non-singleton dimension 3
    # RuntimeError: expected scalar type Half but found Float
    # with - 256
    if memory_restriction_level > 0:
        max_length_tokenize = 768 - 256 if memory_restriction_level <= 2 else 512 - 256
    else:
        # at least give room for 1 paragraph output
        max_length_tokenize = model_max_length - 256
    cutoff_len = max_length_tokenize * 4  # if reaches limit, then can't generate new tokens
    output_smallest = 30 * 4
    max_prompt_length = cutoff_len - output_smallest

    if for_context:
        # then lower even more to avoid later chop, since just estimate tokens in context bot
        max_prompt_length = max(64, int(max_prompt_length * 0.8))

    return cutoff_len, output_smallest, max_length_tokenize, max_prompt_length


class H2OTextIteratorStreamer(TextIteratorStreamer):
    """
    normally, timeout required for now to handle exceptions, else get()
    but with H2O version of TextIteratorStreamer, loop over block to handle
    """

    def __init__(self, tokenizer, skip_prompt: bool = False, timeout: typing.Optional[float] = None,
                 block=True, **decode_kwargs):
        super().__init__(tokenizer, skip_prompt, **decode_kwargs)
        self.text_queue = queue.Queue()
        self.stop_signal = None
        self.do_stop = False
        self.timeout = timeout
        self.block = block

    def on_finalized_text(self, text: str, stream_end: bool = False):
        """Put the new text in the queue. If the stream is ending, also put a stop signal in the queue."""
        self.text_queue.put(text, timeout=self.timeout)
        if stream_end:
            self.text_queue.put(self.stop_signal, timeout=self.timeout)

    def __iter__(self):
        return self

    def __next__(self):
        while True:
            try:
                value = self.stop_signal  # value looks unused in pycharm, not true
                if self.do_stop:
                    print("hit stop", flush=True)
                    # could raise or break, maybe best to raise and make parent see if any exception in thread
                    raise StopIteration()
                    # break
                value = self.text_queue.get(block=self.block, timeout=self.timeout)
                break
            except queue.Empty:
                time.sleep(0.01)
        if value == self.stop_signal:
            raise StopIteration()
        else:
            return value


def generate_with_exceptions(func, *args, prompt='', inputs_decoded='', raise_generate_gpu_exceptions=True, **kwargs):
    try:
        func(*args, **kwargs)
    except torch.cuda.OutOfMemoryError as e:
        print("GPU OOM 2: prompt: %s inputs_decoded: %s exception: %s" % (prompt, inputs_decoded, str(e)),
              flush=True)
        if 'input_ids' in kwargs:
            if kwargs['input_ids'] is not None:
                kwargs['input_ids'].cpu()
            kwargs['input_ids'] = None
        traceback.print_exc()
        clear_torch_cache()
        return
    except (Exception, RuntimeError) as e:
        if 'Expected all tensors to be on the same device' in str(e) or \
                'expected scalar type Half but found Float' in str(e) or \
                'probability tensor contains either' in str(e) or \
                'cublasLt ran into an error!' in str(e) or \
                'mat1 and mat2 shapes cannot be multiplied' in str(e):
            print(
                "GPU Error: prompt: %s inputs_decoded: %s exception: %s" % (prompt, inputs_decoded, str(e)),
                flush=True)
            traceback.print_exc()
            clear_torch_cache()
            if raise_generate_gpu_exceptions:
                raise
            return
        else:
            clear_torch_cache()
            if raise_generate_gpu_exceptions:
                raise


def get_generate_params(model_lower, chat,
                        stream_output, show_examples,
                        prompt_type, prompt_dict,
                        temperature, top_p, top_k, num_beams,
                        max_new_tokens, min_new_tokens, early_stopping, max_time,
                        repetition_penalty, num_return_sequences,
                        do_sample,
                        top_k_docs, chunk, chunk_size,
                        verbose):
    use_defaults = False
    use_default_examples = True
    examples = []
    task_info = 'LLM'
    if model_lower:
        print(f"Using Model {model_lower}", flush=True)
    else:
        if verbose:
            print("No model defined yet", flush=True)

    min_new_tokens = min_new_tokens if min_new_tokens is not None else 0
    early_stopping = early_stopping if early_stopping is not None else False
    max_time_defaults = 60 * 3
    max_time = max_time if max_time is not None else max_time_defaults

    if not prompt_type and model_lower in inv_prompt_type_to_model_lower:
        prompt_type = inv_prompt_type_to_model_lower[model_lower]
        if verbose:
            print("Auto-selecting prompt_type=%s for %s" % (prompt_type, model_lower), flush=True)

    # examples at first don't include chat, instruction_nochat, iinput_nochat, added at end
    if show_examples is None:
        if chat:
            show_examples = False
        else:
            show_examples = True

    summarize_example1 = """Jeff: Can I train a ? Transformers model on Amazon SageMaker? 
Philipp: Sure you can use the new Hugging Face Deep Learning Container. 
Jeff: ok.
Jeff: and how can I get started? 
Jeff: where can I find documentation? 
Philipp: ok, ok you can find everything here. https://huggingface.co/blog/the-partnership-amazon-sagemaker-and-hugging-face"""

    use_placeholder_instruction_as_example = False
    if 'bart-large-cnn-samsum' in model_lower or 'flan-t5-base-samsum' in model_lower:
        placeholder_instruction = summarize_example1
        placeholder_input = ""
        use_defaults = True
        use_default_examples = False
        use_placeholder_instruction_as_example = True
        task_info = "Summarization"
    elif 't5-' in model_lower or 't5' == model_lower or 'flan-' in model_lower:
        placeholder_instruction = "The square root of x is the cube root of y. What is y to the power of 2, if x = 4?"
        placeholder_input = ""
        use_defaults = True
        use_default_examples = True
        task_info = "Multi-Task: Q/A, translation, Chain-of-Thought, Logical Reasoning, Summarization, etc.  Best to use task prefix as trained on, e.g. `translate English to German: ` (space after colon)"
    elif 'mbart-' in model_lower:
        placeholder_instruction = "The girl has long hair."
        placeholder_input = ""
        use_defaults = True
        use_default_examples = False
        use_placeholder_instruction_as_example = True
    elif 'gpt2' in model_lower:
        placeholder_instruction = "The sky is"
        placeholder_input = ""
        prompt_type = prompt_type or 'plain'
        use_default_examples = True  # some will be odd "continuations" but can be ok
        use_placeholder_instruction_as_example = True
        task_info = "Auto-complete phrase, code, etc."
        use_defaults = True
    else:
        if chat:
            placeholder_instruction = ""
        else:
            placeholder_instruction = "Give detailed answer for whether Einstein or Newton is smarter."
        placeholder_input = ""
        if model_lower in inv_prompt_type_to_model_lower:
            prompt_type = inv_prompt_type_to_model_lower[model_lower]
        elif model_lower:
            # default is plain, because might rely upon trust_remote_code to handle prompting
            prompt_type = prompt_type or 'plain'
        else:
            prompt_type = ''
        task_info = "No task"
        if prompt_type == 'instruct':
            task_info = "Answer question or follow imperative as instruction with optionally input."
        elif prompt_type == 'plain':
            task_info = "Auto-complete phrase, code, etc."
        elif prompt_type == 'human_bot':
            if chat:
                task_info = "Chat (Shift-Enter to give question/imperative, input concatenated with instruction)"
            else:
                task_info = "Ask question/imperative (input concatenated with instruction)"

    # revert to plain if still nothing
    prompt_type = prompt_type or 'plain'
    if use_defaults:
        temperature = 1.0 if temperature is None else temperature
        top_p = 1.0 if top_p is None else top_p
        top_k = 40 if top_k is None else top_k
        num_beams = num_beams or 1
        max_new_tokens = max_new_tokens or 128
        repetition_penalty = repetition_penalty or 1.07
        num_return_sequences = min(num_beams, num_return_sequences or 1)
        do_sample = False if do_sample is None else do_sample
    else:
        temperature = 0.1 if temperature is None else temperature
        top_p = 0.75 if top_p is None else top_p
        top_k = 40 if top_k is None else top_k
        num_beams = num_beams or 1
        max_new_tokens = max_new_tokens or 256
        repetition_penalty = repetition_penalty or 1.07
        num_return_sequences = min(num_beams, num_return_sequences or 1)
        do_sample = False if do_sample is None else do_sample
    # doesn't include chat, instruction_nochat, iinput_nochat, added later
    params_list = ["",
                   stream_output,
                   prompt_type, prompt_dict,
                   temperature, top_p, top_k, num_beams,
                   max_new_tokens, min_new_tokens,
                   early_stopping, max_time, repetition_penalty, num_return_sequences, do_sample]

    if use_placeholder_instruction_as_example:
        examples += [[placeholder_instruction, ''] + params_list]

    if use_default_examples:
        examples += [
            ["Translate English to French", "Good morning"] + params_list,
            ["Give detailed answer for whether Einstein or Newton is smarter.", ''] + params_list,
            ["Explain in detailed list, all the best practices for coding in python.", ''] + params_list,
            [
                "Create a markdown table with 3 rows for the primary colors, and 2 columns, with color name and hex codes.",
                ''] + params_list,
            ['Translate to German:  My name is Arthur', ''] + params_list,
            ["Please answer to the following question. Who is going to be the next Ballon d'or?", ''] + params_list,
            ['Can Geoffrey Hinton have a conversation with George Washington? Give the rationale before answering.',
             ''] + params_list,
            ['Please answer the following question. What is the boiling point of Nitrogen?', ''] + params_list,
            ['Answer the following yes/no question. Can you write a whole Haiku in a single tweet?', ''] + params_list,
            ["Simplify the following expression: (False or False and True). Explain your answer.", ''] + params_list,
            [
                "Premise: At my age you will probably have learnt one lesson. Hypothesis:  It's not certain how many lessons you'll learn by your thirties. Does the premise entail the hypothesis?",
                ''] + params_list,
            ['The square root of x is the cube root of y. What is y to the power of 2, if x = 4?', ''] + params_list,
            [
                'Answer the following question by reasoning step by step.  The cafeteria had 23 apples. If they used 20 for lunch, and bought 6 more, how many apple do they have?',
                ''] + params_list,
            ["""def area_of_rectangle(a: float, b: float):
    \"\"\"Return the area of the rectangle.\"\"\"""", ''] + params_list,
            ["""# a function in native python:
def mean(a):
    return sum(a)/len(a)

# the same function using numpy:
import numpy as np
def mean(a):""", ''] + params_list,
            ["""X = np.random.randn(100, 100)
y = np.random.randint(0, 1, 100)

# fit random forest classifier with 20 estimators""", ''] + params_list,
        ]
    # add summary example
    examples += [
        [summarize_example1, 'Summarize' if prompt_type not in ['plain', 'instruct_simple'] else ''] + params_list]

    src_lang = "English"
    tgt_lang = "Russian"

    # move to correct position
    for example in examples:
        example += [chat, '', '', 'Disabled', top_k_docs, chunk, chunk_size, [DocumentChoices.All_Relevant.name]]
        # adjust examples if non-chat mode
        if not chat:
            example[eval_func_param_names.index('instruction_nochat')] = example[
                eval_func_param_names.index('instruction')]
            example[eval_func_param_names.index('instruction')] = ''

            example[eval_func_param_names.index('iinput_nochat')] = example[eval_func_param_names.index('iinput')]
            example[eval_func_param_names.index('iinput')] = ''
        assert len(example) == len(eval_func_param_names), "Wrong example: %s %s" % (
            len(example), len(eval_func_param_names))

    if prompt_type == PromptType.custom.name and not prompt_dict:
        raise ValueError("Unexpected to get non-empty prompt_dict=%s for prompt_type=%s" % (prompt_dict, prompt_type))

    # get prompt_dict from prompt_type, so user can see in UI etc., or for custom do nothing except check format
    prompt_dict, error0 = get_prompt(prompt_type, prompt_dict,
                                     chat=False, context='', reduced=False, making_context=False, return_dict=True)
    if error0:
        raise RuntimeError("Prompt wrong: %s" % error0)

    return placeholder_instruction, placeholder_input, \
        stream_output, show_examples, \
        prompt_type, prompt_dict, \
        temperature, top_p, top_k, num_beams, \
        max_new_tokens, min_new_tokens, early_stopping, max_time, \
        repetition_penalty, num_return_sequences, \
        do_sample, \
        src_lang, tgt_lang, \
        examples, \
        task_info


def languages_covered():
    # https://huggingface.co/facebook/mbart-large-50-many-to-many-mmt#languages-covered
    covered = """Arabic (ar_AR), Czech (cs_CZ), German (de_DE), English (en_XX), Spanish (es_XX), Estonian (et_EE), Finnish (fi_FI), French (fr_XX), Gujarati (gu_IN), Hindi (hi_IN), Italian (it_IT), Japanese (ja_XX), Kazakh (kk_KZ), Korean (ko_KR), Lithuanian (lt_LT), Latvian (lv_LV), Burmese (my_MM), Nepali (ne_NP), Dutch (nl_XX), Romanian (ro_RO), Russian (ru_RU), Sinhala (si_LK), Turkish (tr_TR), Vietnamese (vi_VN), Chinese (zh_CN), Afrikaans (af_ZA), Azerbaijani (az_AZ), Bengali (bn_IN), Persian (fa_IR), Hebrew (he_IL), Croatian (hr_HR), Indonesian (id_ID), Georgian (ka_GE), Khmer (km_KH), Macedonian (mk_MK), Malayalam (ml_IN), Mongolian (mn_MN), Marathi (mr_IN), Polish (pl_PL), Pashto (ps_AF), Portuguese (pt_XX), Swedish (sv_SE), Swahili (sw_KE), Tamil (ta_IN), Telugu (te_IN), Thai (th_TH), Tagalog (tl_XX), Ukrainian (uk_UA), Urdu (ur_PK), Xhosa (xh_ZA), Galician (gl_ES), Slovene (sl_SI)"""
    covered = covered.split(', ')
    covered = {x.split(' ')[0]: x.split(' ')[1].replace(')', '').replace('(', '') for x in covered}
    return covered


def get_context(chat_context, prompt_type):
    if chat_context and prompt_type == 'human_bot':
        context0 = """<bot>: I am an intelligent, helpful, truthful, and fair assistant named h2oGPT, who will give accurate, balanced, and reliable responses.  I will not respond with I don't know or I don't understand.
<human>: I am a human person seeking useful assistance and request all questions be answered completely, and typically expect detailed responses.  Give answers in numbered list format if several distinct but related items are being listed."""
    else:
        context0 = ''
    return context0


def score_qa(smodel, stokenizer, max_length_tokenize, question, answer, cutoff_len):
    question = question[-cutoff_len:]
    answer = answer[-cutoff_len:]

    inputs = stokenizer(question, answer,
                        return_tensors="pt",
                        truncation=True,
                        max_length=max_length_tokenize).to(smodel.device)
    try:
        score = torch.sigmoid(smodel(**inputs).logits[0]).cpu().detach().numpy()[0]
    except torch.cuda.OutOfMemoryError as e:
        print("GPU OOM 3: question: %s answer: %s exception: %s" % (question, answer, str(e)), flush=True)
        del inputs
        traceback.print_exc()
        clear_torch_cache()
        return 'Response Score: GPU OOM'
    except (Exception, RuntimeError) as e:
        if 'Expected all tensors to be on the same device' in str(e) or \
                'expected scalar type Half but found Float' in str(e) or \
                'probability tensor contains either' in str(e) or \
                'cublasLt ran into an error!' in str(e) or \
                'device-side assert triggered' in str(e):
            print("GPU Error: question: %s answer: %s exception: %s" % (question, answer, str(e)),
                  flush=True)
            traceback.print_exc()
            clear_torch_cache()
            return 'Response Score: GPU Error'
        else:
            raise
    os.environ['TOKENIZERS_PARALLELISM'] = 'true'
    return score


def check_locals(**kwargs):
    # ensure everything in evaluate is here
    can_skip_because_locally_generated = no_default_param_names + [
        # get_model:
        'reward_type'
    ]
    for k in eval_func_param_names:
        if k in can_skip_because_locally_generated:
            continue
        assert k in kwargs, "Missing %s" % k
    for k in inputs_kwargs_list:
        if k in can_skip_because_locally_generated:
            continue
        assert k in kwargs, "Missing %s" % k

    for k in list(inspect.signature(get_model).parameters):
        if k in can_skip_because_locally_generated:
            continue
        assert k in kwargs, "Missing %s" % k


def get_model_max_length(model_state):
    if not isinstance(model_state['tokenizer'], (str, types.NoneType)):
        return model_state['tokenizer'].model_max_length
    else:
        return 2048


def get_max_max_new_tokens(model_state, **kwargs):
    if not isinstance(model_state['tokenizer'], (str, types.NoneType)):
        max_max_new_tokens = model_state['tokenizer'].model_max_length
    else:
        max_max_new_tokens = None

    if kwargs['max_max_new_tokens'] is not None and max_max_new_tokens is not None:
        return min(max_max_new_tokens, kwargs['max_max_new_tokens'])
    elif kwargs['max_max_new_tokens'] is not None:
        return kwargs['max_max_new_tokens']
    elif kwargs['memory_restriction_level'] == 1:
        return 768
    elif kwargs['memory_restriction_level'] == 2:
        return 512
    elif kwargs['memory_restriction_level'] >= 3:
        return 256
    else:
        # FIXME: Need to update after new model loaded, so user can control with slider
        return 2048


def get_minmax_top_k_docs(is_public):
    if is_public:
        min_top_k_docs = 1
        max_top_k_docs = 3
        label_top_k_docs = "Number of document chunks"
    else:
        min_top_k_docs = -1
        max_top_k_docs = 100
        label_top_k_docs = "Number of document chunks (-1 = auto fill model context)"
    return min_top_k_docs, max_top_k_docs, label_top_k_docs


def history_to_context(history, langchain_mode1, prompt_type1, prompt_dict1, chat1, model_max_length1,
                       memory_restriction_level1, keep_sources_in_context1):
    """
    consumes all history up to (but not including) latest history item that is presumed to be an [instruction, None] pair
    :param history:
    :param langchain_mode1:
    :param prompt_type1:
    :param prompt_dict1:
    :param chat1:
    :param model_max_length1:
    :param memory_restriction_level1:
    :param keep_sources_in_context1:
    :return:
    """
    # ensure output will be unique to models
    _, _, _, max_prompt_length = get_cutoffs(memory_restriction_level1,
                                             for_context=True, model_max_length=model_max_length1)
    context1 = ''
    if max_prompt_length is not None and langchain_mode1 not in ['LLM']:
        context1 = ''
        # - 1 below because current instruction already in history from user()
        for histi in range(0, len(history) - 1):
            data_point = dict(instruction=history[histi][0], input='', output=history[histi][1])
            prompt, pre_response, terminate_response, chat_sep, chat_turn_sep = generate_prompt(data_point,
                                                                                                prompt_type1,
                                                                                                prompt_dict1,
                                                                                                chat1,
                                                                                                reduced=True,
                                                                                                making_context=True)
            # md -> back to text, maybe not super important if model trained enough
            if not keep_sources_in_context1 and langchain_mode1 != 'Disabled' and prompt.find(source_prefix) >= 0:
                # FIXME: This is relatively slow even for small amount of text, like 0.3s each history item
                import re
                prompt = re.sub(f'{re.escape(source_prefix)}.*?{re.escape(source_postfix)}', '', prompt,
                                flags=re.DOTALL)
                if prompt.endswith('\n<p>'):
                    prompt = prompt[:-4]
            prompt = prompt.replace('<br>', chat_turn_sep)
            if not prompt.endswith(chat_turn_sep):
                prompt += chat_turn_sep
            # most recent first, add older if can
            # only include desired chat history
            if len(prompt + context1) > max_prompt_length:
                break
            context1 += prompt

        _, pre_response, terminate_response, chat_sep, chat_turn_sep = generate_prompt({}, prompt_type1, prompt_dict1,
                                                                                       chat1, reduced=True,
                                                                                       making_context=True)
        if context1 and not context1.endswith(chat_turn_sep):
            context1 += chat_turn_sep  # ensure if terminates abruptly, then human continues on next line
    return context1


def entrypoint_main():
    """
    Examples:

    WORLD_SIZE=4 CUDA_VISIBLE_DEVICES="0,1,2,3" torchrun --nproc_per_node=4 --master_port=1234 generate.py --base_model='EleutherAI/gpt-j-6B' --lora_weights=lora-alpaca_6B
    python generate.py --base_model='EleutherAI/gpt-j-6B' --lora_weights='lora-alpaca_6B'
    python generate.py --base_model='EleutherAI/gpt-neox-20b' --lora_weights='lora-alpaca_20B'
    
    # generate without lora weights, no prompt
    python generate.py --base_model='EleutherAI/gpt-neox-20b' --prompt_type='plain'
    python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='dai_faq'

    python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='dai_faq' --lora_weights='lora_20B_daifaq'
    # OpenChatKit settings:
    python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='human_bot --debug=True --num_beams=1 --temperature=0.6 --top_k=40 --top_p=1.0

    python generate.py --base_model='distilgpt2' --prompt_type='plain' --debug=True --num_beams=1 --temperature=0.6 --top_k=40 --top_p=1.0 --share=False
    python generate.py --base_model='t5-large' --prompt_type='simple_instruct'
    python generate.py --base_model='philschmid/bart-large-cnn-samsum'
    python generate.py --base_model='philschmid/flan-t5-base-samsum'
    python generate.py --base_model='facebook/mbart-large-50-many-to-many-mmt'

    python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='human_bot' --lora_weights='GPT-NeoXT-Chat-Base-20B.merged.json.8_epochs.57b2892c53df5b8cefac45f84d019cace803ef26.28'

    must have 4*48GB GPU and run without 8bit in order for sharding to work with infer_devices=False
    can also pass --prompt_type='human_bot' and model can somewhat handle instructions without being instruct tuned
    python generate.py --base_model=decapoda-research/llama-65b-hf --load_8bit=False --infer_devices=False --prompt_type='human_bot'

    python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b
    """
    fire.Fire(main)


if __name__ == "__main__":
    entrypoint_main()