forked from henk717/KoboldAI
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathaiserver.py
10291 lines (9392 loc) · 446 KB
/
aiserver.py
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
#!/usr/bin/python3
#==================================================================#
# KoboldAI
# Version: 1.19.2
# By: The KoboldAI Community
#==================================================================#
# External packages
import eventlet
eventlet.monkey_patch(all=True, thread=False, os=False)
import os
os.system("")
__file__ = os.path.dirname(os.path.realpath(__file__))
os.chdir(__file__)
os.environ['EVENTLET_THREADPOOL_SIZE'] = '1'
os.environ['TOKENIZERS_PARALLELISM'] = 'false'
from eventlet import tpool
import logging
from logger import logger, set_logger_verbosity, quiesce_logger
logging.getLogger("urllib3").setLevel(logging.ERROR)
from os import path, getcwd
import time
import re
import json
import collections
import zipfile
import packaging
import packaging.version
import contextlib
import traceback
import threading
import markdown
import bleach
import itertools
import bisect
import functools
import traceback
import inspect
import warnings
from collections.abc import Iterable
from typing import Any, Callable, TypeVar, Tuple, Union, Dict, Set, List, Optional, Type
import requests
import html
import argparse
import sys
import gc
import lupa
import importlib
# KoboldAI
import fileops
import gensettings
from utils import debounce
import utils
import structures
import torch
from transformers import StoppingCriteria, GPT2Tokenizer, GPT2LMHeadModel, GPTNeoForCausalLM, GPTNeoModel, AutoModelForCausalLM, AutoTokenizer, PreTrainedModel, modeling_utils
from transformers import __version__ as transformers_version
import transformers
try:
from transformers.models.opt.modeling_opt import OPTDecoder
except:
pass
import transformers.generation_utils
global tpu_mtj_backend
if lupa.LUA_VERSION[:2] != (5, 4):
logger.error(f"Please install lupa==1.10. You have lupa {lupa.__version__}.")
patch_causallm_patched = False
# Make sure tqdm progress bars display properly in Colab
from tqdm.auto import tqdm
old_init = tqdm.__init__
def new_init(self, *args, **kwargs):
old_init(self, *args, **kwargs)
if(self.ncols == 0 and kwargs.get("ncols") != 0):
self.ncols = 99
tqdm.__init__ = new_init
# Fix some issues with the OPT tokenizer
from transformers import PreTrainedTokenizerBase
old_pretrainedtokenizerbase_from_pretrained = PreTrainedTokenizerBase.from_pretrained.__func__
@classmethod
def new_pretrainedtokenizerbase_from_pretrained(cls, *args, **kwargs):
tokenizer = old_pretrainedtokenizerbase_from_pretrained(cls, *args, **kwargs)
tokenizer._koboldai_header = tokenizer.encode("")
tokenizer.add_bos_token = False
tokenizer.add_prefix_space = False
return tokenizer
PreTrainedTokenizerBase.from_pretrained = new_pretrainedtokenizerbase_from_pretrained
#==================================================================#
# Variables & Storage
#==================================================================#
# Terminal tags for colored text
class colors:
PURPLE = '\033[95m'
BLUE = '\033[94m'
CYAN = '\033[96m'
GREEN = '\033[92m'
YELLOW = '\033[93m'
RED = '\033[91m'
END = '\033[0m'
UNDERLINE = '\033[4m'
# AI models Menu
# This is a dict of lists where they key is the menu name, and the list is the menu items.
# Each item takes the 4 elements, 1: Text to display, 2: Model Name (var.model) or menu name (Key name for another menu),
# 3: the memory requirement for the model, 4: if the item is a menu or not (True/False)
model_menu = {
'mainmenu': [
["Load a model from its directory", "NeoCustom", "", False],
["Load an old GPT-2 model (eg CloverEdition)", "GPT2Custom", "", False],
["Adventure Models", "adventurelist", "", True],
["Novel Models", "novellist", "", True],
["NSFW Models", "nsfwlist", "", True],
["Untuned OPT", "optlist", "", True],
["Untuned GPT-Neo/J", "gptneolist", "", True],
["Untuned Pythia", "pythialist", "", True],
["Untuned Fairseq Dense", "fsdlist", "", True],
["Untuned Bloom", "bloomlist", "", True],
["Untuned XGLM", "xglmlist", "", True],
["Untuned GPT2", "gpt2list", "", True],
["Online Services", "apilist", "", True],
["Read Only (No AI)", "ReadOnly", "", False]
],
'adventurelist': [
["Skein 20B", "KoboldAI/GPT-NeoX-20B-Skein", "64GB", False],
["Nerys OPT 13B V2 (Hybrid)", "KoboldAI/OPT-13B-Nerys-v2", "32GB", False],
["Nerys FSD 13B V2 (Hybrid)", "KoboldAI/fairseq-dense-13B-Nerys-v2", "32GB", False],
["Nerys FSD 13B (Hybrid)", "KoboldAI/fairseq-dense-13B-Nerys", "32GB", False],
["Skein 6B", "KoboldAI/GPT-J-6B-Skein", "16GB", False],
["OPT Nerys 6B V2 (Hybrid)", "KoboldAI/OPT-6B-nerys-v2", "16GB", False],
["Adventure 6B", "KoboldAI/GPT-J-6B-Adventure", "16GB", False],
["Nerys FSD 2.7B (Hybrid)", "KoboldAI/fairseq-dense-2.7B-Nerys", "8GB", False],
["Adventure 2.7B", "KoboldAI/GPT-Neo-2.7B-AID", "8GB", False],
["Adventure 1.3B", "KoboldAI/GPT-Neo-1.3B-Adventure", "6GB", False],
["Adventure 125M (Mia)", "Merry/AID-Neo-125M", "2GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'novellist': [
["Nerys OPT 13B V2 (Hybrid)", "KoboldAI/OPT-13B-Nerys-v2", "32GB", False],
["Nerys FSD 13B V2 (Hybrid)", "KoboldAI/fairseq-dense-13B-Nerys-v2", "32GB", False],
["Janeway FSD 13B", "KoboldAI/fairseq-dense-13B-Janeway", "32GB", False],
["Nerys FSD 13B (Hybrid)", "KoboldAI/fairseq-dense-13B-Nerys", "32GB", False],
["OPT Nerys 6B V2 (Hybrid)", "KoboldAI/OPT-6B-nerys-v2", "16GB", False],
["Janeway FSD 6.7B", "KoboldAI/fairseq-dense-6.7B-Janeway", "16GB", False],
["Janeway Neo 6B", "KoboldAI/GPT-J-6B-Janeway", "16GB", False],
["Qilin Lit 6B (SFW)", "rexwang8/qilin-lit-6b", "16GB", False],
["Janeway Neo 2.7B", "KoboldAI/GPT-Neo-2.7B-Janeway", "8GB", False],
["Janeway FSD 2.7B", "KoboldAI/fairseq-dense-2.7B-Janeway", "8GB", False],
["Nerys FSD 2.7B (Hybrid)", "KoboldAI/fairseq-dense-2.7B-Nerys", "8GB", False],
["Horni-LN 2.7B", "KoboldAI/GPT-Neo-2.7B-Horni-LN", "8GB", False],
["Picard 2.7B (Older Janeway)", "KoboldAI/GPT-Neo-2.7B-Picard", "8GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'nsfwlist': [
["Erebus 20B (NSFW)", "KoboldAI/GPT-NeoX-20B-Erebus", "64GB", False],
["Erebus 13B (NSFW)", "KoboldAI/OPT-13B-Erebus", "32GB", False],
["Shinen FSD 13B (NSFW)", "KoboldAI/fairseq-dense-13B-Shinen", "32GB", False],
["Erebus 6.7B (NSFW)", "KoboldAI/OPT-6.7B-Erebus", "16GB", False],
["Shinen FSD 6.7B (NSFW)", "KoboldAI/fairseq-dense-6.7B-Shinen", "16GB", False],
["Lit V2 6B (NSFW)", "hakurei/litv2-6B-rev3", "16GB", False],
["Lit 6B (NSFW)", "hakurei/lit-6B", "16GB", False],
["Shinen 6B (NSFW)", "KoboldAI/GPT-J-6B-Shinen", "16GB", False],
["Erebus 2.7B (NSFW)", "KoboldAI/OPT-2.7B-Erebus", "8GB", False],
["Horni 2.7B (NSFW)", "KoboldAI/GPT-Neo-2.7B-Horni", "8GB", False],
["Shinen 2.7B (NSFW)", "KoboldAI/GPT-Neo-2.7B-Shinen", "8GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'chatlist': [
["Convo 6B (Chatbot)", "hitomi-team/convo-6B", "16GB", False],
["C1 6B (Chatbot)", "hakurei/c1-6B", "16GB", False],
["C1 1.3B (Chatbot)", "iokru/c1-1.3B", "6GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'gptneolist': [
["GPT-NeoX 20B", "EleutherAI/gpt-neox-20b", "64GB", False],
["Pythia 13B (NeoX, Same dataset)", "EleutherAI/pythia-13b", "32GB", False],
["GPT-J 6B", "EleutherAI/gpt-j-6B", "16GB", False],
["GPT-Neo 2.7B", "EleutherAI/gpt-neo-2.7B", "8GB", False],
["GPT-Neo 1.3B", "EleutherAI/gpt-neo-1.3B", "6GB", False],
["Pythia 800M (NeoX, Same dataset)", "EleutherAI/pythia-800m", "4GB", False],
["Pythia 350M (NeoX, Same dataset)", "EleutherAI/pythia-350m", "2GB", False],
["GPT-Neo 125M", "EleutherAI/gpt-neo-125M", "2GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'pythialist': [
["Pythia 13B Deduped", "EleutherAI/pythia-13b-deduped", "32GB", False],
["Pythia 13B", "EleutherAI/pythia-13b", "32GB", False],
["Pythia 6.7B Deduped", "EleutherAI/pythia-6.7b-deduped", "16GB", False],
["Pythia 6.7B", "EleutherAI/pythia-6.7b", "16GB", False],
["Pythia 1.3B Deduped", "EleutherAI/pythia-1.3b-deduped", "6GB", False],
["Pythia 1.3B", "EleutherAI/pythia-1.3b", "6GB", False],
["Pythia 800M", "EleutherAI/pythia-800m", "4GB", False],
["Pythia 350M Deduped", "EleutherAI/pythia-350m-deduped", "2GB", False],
["Pythia 350M", "EleutherAI/pythia-350m", "2GB", False],
["Pythia 125M Deduped", "EleutherAI/pythia-125m-deduped", "2GB", False],
["Pythia 125M", "EleutherAI/pythia-125m", "2GB", False],
["Pythia 19M Deduped", "EleutherAI/pythia-19m-deduped", "1GB", False],
["Pythia 19M", "EleutherAI/pythia-19m", "1GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'gpt2list': [
["GPT-2 XL", "gpt2-xl", "6GB", False],
["GPT-2 Large", "gpt2-large", "4GB", False],
["GPT-2 Med", "gpt2-medium", "2GB", False],
["GPT-2", "gpt2", "2GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'bloomlist': [
["Bloom 176B", "bigscience/bloom", "", False],
["Bloom 7.1B", "bigscience/bloom-7b1", "", False],
["Bloom 3B", "bigscience/bloom-3b", "", False],
["Bloom 1.7B", "bigscience/bloom-1b7", "", False],
["Bloom 560M", "bigscience/bloom-560m", "", False],
["Return to Main Menu", "mainmenu", "", True],
],
'optlist': [
["OPT 66B", "facebook/opt-66b", "128GB", False],
["OPT 30B", "facebook/opt-30b", "64GB", False],
["OPT 13B", "facebook/opt-13b", "32GB", False],
["OPT 6.7B", "facebook/opt-6.7b", "16GB", False],
["OPT 2.7B", "facebook/opt-2.7b", "8GB", False],
["OPT 1.3B", "facebook/opt-1.3b", "4GB", False],
["OPT 350M", "facebook/opt-350m", "2GB", False],
["OPT 125M", "facebook/opt-125m", "1GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'fsdlist': [
["Fairseq Dense 13B", "KoboldAI/fairseq-dense-13B", "32GB", False],
["Fairseq Dense 6.7B", "KoboldAI/fairseq-dense-6.7B", "16GB", False],
["Fairseq Dense 2.7B", "KoboldAI/fairseq-dense-2.7B", "8GB", False],
["Fairseq Dense 1.3B", "KoboldAI/fairseq-dense-1.3B", "4GB", False],
["Fairseq Dense 355M", "KoboldAI/fairseq-dense-355M", "2GB", False],
["Fairseq Dense 125M", "KoboldAI/fairseq-dense-125M", "1GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'xglmlist': [
["XGLM 4.5B (Larger Dataset)", "facebook/xglm-4.5B", "12GB", False],
["XGLM 7.5B", "facebook/xglm-7.5B", "18GB", False],
["XGLM 2.9B", "facebook/xglm-2.9B", "10GB", False],
["XGLM 1.7B", "facebook/xglm-1.7B", "6GB", False],
["XGLM 564M", "facebook/xglm-564M", "4GB", False],
["Return to Main Menu", "mainmenu", "", True],
],
'apilist': [
["GooseAI API (requires API key)", "GooseAI", "", False],
["OpenAI API (requires API key)", "OAI", "", False],
["InferKit API (requires API key)", "InferKit", "", False],
# ["KoboldAI Server API (Old Google Colab)", "Colab", "", False],
["KoboldAI API", "API", "", False],
["KoboldAI Horde", "CLUSTER", "", False],
["Return to Main Menu", "mainmenu", "", True],
]
}
class TokenStreamQueue:
def __init__(self):
self.probability_buffer = None
self.queue = []
def add_text(self, text):
self.queue.append({
"decoded": text,
"probabilities": self.probability_buffer
})
self.probability_buffer = None
# Variables
class vars:
lastact = "" # The last action received from the user
submission = "" # Same as above, but after applying input formatting
lastctx = "" # The last context submitted to the generator
model = "ReadOnly" # Model ID string chosen at startup
online_model = "" # Used when Model ID is an online service, and there is a secondary option for the actual model name
model_selected = "" #selected model in UI
model_type = "" # Model Type (Automatically taken from the model config)
noai = False # Runs the script without starting up the transformers pipeline
aibusy = False # Stops submissions while the AI is working
max_length = 1024 # Maximum number of tokens to submit per action
ikmax = 3000 # Maximum number of characters to submit to InferKit
genamt = 80 # Amount of text for each action to generate
ikgen = 200 # Number of characters for InferKit to generate
rep_pen = 1.1 # Default generator repetition_penalty
rep_pen_slope = 0.7 # Default generator repetition penalty slope
rep_pen_range = 1024 # Default generator repetition penalty range
temp = 0.5 # Default generator temperature
top_p = 0.9 # Default generator top_p
top_k = 0 # Default generator top_k
top_a = 0.0 # Default generator top-a
tfs = 1.0 # Default generator tfs (tail-free sampling)
typical = 1.0 # Default generator typical sampling threshold
numseqs = 1 # Number of sequences to ask the generator to create
full_determinism = False # Whether or not full determinism is enabled
seed_specified = False # Whether or not the current RNG seed was specified by the user (in their settings file)
seed = None # The current RNG seed (as an int), or None if unknown
gamestarted = False # Whether the game has started (disables UI elements)
gamesaved = True # Whether or not current game is saved
serverstarted = False # Whether or not the Flask server has started
prompt = "" # Prompt
memory = "" # Text submitted to memory field
authornote = "" # Text submitted to Author's Note field
authornotetemplate = "[Author's note: <|>]" # Author's note template
setauthornotetemplate = authornotetemplate # Saved author's note template in settings
andepth = 3 # How far back in history to append author's note
actions = structures.KoboldStoryRegister() # Actions submitted by user and AI
actions_metadata = {} # List of dictonaries, one dictonary for every action that contains information about the action like alternative options.
# Contains at least the same number of items as actions. Back action will remove an item from actions, but not actions_metadata
# Dictonary keys are:
# Selected Text: (text the user had selected. None when this is a newly generated action)
# Alternative Generated Text: {Text, Pinned, Previous Selection, Edited}
#
worldinfo = [] # List of World Info key/value objects
worldinfo_i = [] # List of World Info key/value objects sans uninitialized entries
worldinfo_u = {} # Dictionary of World Info UID - key/value pairs
wifolders_d = {} # Dictionary of World Info folder UID-info pairs
wifolders_l = [] # List of World Info folder UIDs
wifolders_u = {} # Dictionary of pairs of folder UID - list of WI UID
modelconfig = {} # Raw contents of the model's config.json, or empty dictionary if none found
lua_state = None # Lua state of the Lua scripting system
lua_koboldbridge = None # `koboldbridge` from bridge.lua
lua_kobold = None # `kobold` from` bridge.lua
lua_koboldcore = None # `koboldcore` from bridge.lua
lua_logname = ... # Name of previous userscript that logged to terminal
lua_running = False # Whether or not Lua is running (i.e. wasn't stopped due to an error)
lua_edited = set() # Set of chunk numbers that were edited from a Lua generation modifier
lua_deleted = set() # Set of chunk numbers that were deleted from a Lua generation modifier
generated_tkns = 0 # If using a backend that supports Lua generation modifiers, how many tokens have already been generated, otherwise 0
abort = False # Whether or not generation was aborted by clicking on the submit button during generation
compiling = False # If using a TPU Colab, this will be set to True when the TPU backend starts compiling and then set to False again
checking = False # Whether or not we are actively checking to see if TPU backend is compiling or not
sp_changed = False # This gets set to True whenever a userscript changes the soft prompt so that check_for_sp_change() can alert the browser that the soft prompt has changed
spfilename = "" # Filename of soft prompt to load, or an empty string if not using a soft prompt
userscripts = [] # List of userscripts to load
last_userscripts = [] # List of previous userscript filenames from the previous time userscripts were send via usstatitems
corescript = "default.lua" # Filename of corescript to load
# badwords = [] # Array of str/chr values that should be removed from output
badwordsids = []
badwordsids_default = [[13460], [6880], [50256], [42496], [4613], [17414], [22039], [16410], [27], [29], [38430], [37922], [15913], [24618], [28725], [58], [47175], [36937], [26700], [12878], [16471], [37981], [5218], [29795], [13412], [45160], [3693], [49778], [4211], [20598], [36475], [33409], [44167], [32406], [29847], [29342], [42669], [685], [25787], [7359], [3784], [5320], [33994], [33490], [34516], [43734], [17635], [24293], [9959], [23785], [21737], [28401], [18161], [26358], [32509], [1279], [38155], [18189], [26894], [6927], [14610], [23834], [11037], [14631], [26933], [46904], [22330], [25915], [47934], [38214], [1875], [14692], [41832], [13163], [25970], [29565], [44926], [19841], [37250], [49029], [9609], [44438], [16791], [17816], [30109], [41888], [47527], [42924], [23984], [49074], [33717], [31161], [49082], [30138], [31175], [12240], [14804], [7131], [26076], [33250], [3556], [38381], [36338], [32756], [46581], [17912], [49146]] # Tokenized array of badwords used to prevent AI artifacting
badwordsids_neox = [[0], [1], [44162], [9502], [12520], [31841], [36320], [49824], [34417], [6038], [34494], [24815], [26635], [24345], [3455], [28905], [44270], [17278], [32666], [46880], [7086], [43189], [37322], [17778], [20879], [49821], [3138], [14490], [4681], [21391], [26786], [43134], [9336], [683], [48074], [41256], [19181], [29650], [28532], [36487], [45114], [46275], [16445], [15104], [11337], [1168], [5647], [29], [27482], [44965], [43782], [31011], [42944], [47389], [6334], [17548], [38329], [32044], [35487], [2239], [34761], [7444], [1084], [12399], [18990], [17636], [39083], [1184], [35830], [28365], [16731], [43467], [47744], [1138], [16079], [40116], [45564], [18297], [42368], [5456], [18022], [42696], [34476], [23505], [23741], [39334], [37944], [45382], [38709], [33440], [26077], [43600], [34418], [36033], [6660], [48167], [48471], [15775], [19884], [41533], [1008], [31053], [36692], [46576], [20095], [20629], [31759], [46410], [41000], [13488], [30952], [39258], [16160], [27655], [22367], [42767], [43736], [49694], [13811], [12004], [46768], [6257], [37471], [5264], [44153], [33805], [20977], [21083], [25416], [14277], [31096], [42041], [18331], [33376], [22372], [46294], [28379], [38475], [1656], [5204], [27075], [50001], [16616], [11396], [7748], [48744], [35402], [28120], [41512], [4207], [43144], [14767], [15640], [16595], [41305], [44479], [38958], [18474], [22734], [30522], [46267], [60], [13976], [31830], [48701], [39822], [9014], [21966], [31422], [28052], [34607], [2479], [3851], [32214], [44082], [45507], [3001], [34368], [34758], [13380], [38363], [4299], [46802], [30996], [12630], [49236], [7082], [8795], [5218], [44740], [9686], [9983], [45301], [27114], [40125], [1570], [26997], [544], [5290], [49193], [23781], [14193], [40000], [2947], [43781], [9102], [48064], [42274], [18772], [49384], [9884], [45635], [43521], [31258], [32056], [47686], [21760], [13143], [10148], [26119], [44308], [31379], [36399], [23983], [46694], [36134], [8562], [12977], [35117], [28591], [49021], [47093], [28653], [29013], [46468], [8605], [7254], [25896], [5032], [8168], [36893], [38270], [20499], [27501], [34419], [29547], [28571], [36586], [20871], [30537], [26842], [21375], [31148], [27618], [33094], [3291], [31789], [28391], [870], [9793], [41361], [47916], [27468], [43856], [8850], [35237], [15707], [47552], [2730], [41449], [45488], [3073], [49806], [21938], [24430], [22747], [20924], [46145], [20481], [20197], [8239], [28231], [17987], [42804], [47269], [29972], [49884], [21382], [46295], [36676], [34616], [3921], [26991], [27720], [46265], [654], [9855], [40354], [5291], [34904], [44342], [2470], [14598], [880], [19282], [2498], [24237], [21431], [16369], [8994], [44524], [45662], [13663], [37077], [1447], [37786], [30863], [42854], [1019], [20322], [4398], [12159], [44072], [48664], [31547], [18736], [9259], [31], [16354], [21810], [4357], [37982], [5064], [2033], [32871], [47446], [62], [22158], [37387], [8743], [47007], [17981], [11049], [4622], [37916], [36786], [35138], [29925], [14157], [18095], [27829], [1181], [22226], [5709], [4725], [30189], [37014], [1254], [11380], [42989], [696], [24576], [39487], [30119], [1092], [8088], [2194], [9899], [14412], [21828], [3725], [13544], [5180], [44679], [34398], [3891], [28739], [14219], [37594], [49550], [11326], [6904], [17266], [5749], [10174], [23405], [9955], [38271], [41018], [13011], [48392], [36784], [24254], [21687], [23734], [5413], [41447], [45472], [10122], [17555], [15830], [47384], [12084], [31350], [47940], [11661], [27988], [45443], [905], [49651], [16614], [34993], [6781], [30803], [35869], [8001], [41604], [28118], [46462], [46762], [16262], [17281], [5774], [10943], [5013], [18257], [6750], [4713], [3951], [11899], [38791], [16943], [37596], [9318], [18413], [40473], [13208], [16375]]
badwordsids_opt = [[44717], [46613], [48513], [49923], [50185], [48755], [8488], [43303], [49659], [48601], [49817], [45405], [48742], [49925], [47720], [11227], [48937], [48784], [50017], [42248], [49310], [48082], [49895], [50025], [49092], [49007], [8061], [44226], [0], [742], [28578], [15698], [49784], [46679], [39365], [49281], [49609], [48081], [48906], [46161], [48554], [49670], [48677], [49721], [49632], [48610], [48462], [47457], [10975], [46077], [28696], [48709], [43839], [49798], [49154], [48203], [49625], [48395], [50155], [47161], [49095], [48833], [49420], [49666], [48443], [22176], [49242], [48651], [49138], [49750], [40389], [48021], [21838], [49070], [45333], [40862], [1], [49915], [33525], [49858], [50254], [44403], [48992], [48872], [46117], [49853], [47567], [50206], [41552], [50068], [48999], [49703], [49940], [49329], [47620], [49868], [49962], [2], [44082], [50236], [31274], [50260], [47052], [42645], [49177], [17523], [48691], [49900], [49069], [49358], [48794], [47529], [46479], [48457], [646], [49910], [48077], [48935], [46386], [48902], [49151], [48759], [49803], [45587], [48392], [47789], [48654], [49836], [49230], [48188], [50264], [46844], [44690], [48505], [50161], [27779], [49995], [41833], [50154], [49097], [48520], [50018], [8174], [50084], [49366], [49526], [50193], [7479], [49982], [3]]
fp32_model = False # Whether or not the most recently loaded HF model was in fp32 format
deletewi = None # Temporary storage for UID to delete
wirmvwhtsp = True # Whether to remove leading whitespace from WI entries
widepth = 3 # How many historical actions to scan for WI hits
mode = "play" # Whether the interface is in play, memory, or edit mode
editln = 0 # Which line was last selected in Edit Mode
gpu_device = 0 # Which PyTorch device to use when using pure GPU generation
url = "https://api.inferkit.com/v1/models/standard/generate" # InferKit API URL
oaiurl = "" # OpenAI API URL
oaiengines = "https://api.openai.com/v1/engines"
colaburl = "" # Ngrok url for Google Colab mode
apikey = "" # API key to use for InferKit API calls
oaiapikey = "" # API key to use for OpenAI API calls
cluster_requested_models = [] # The models which we allow to generate during cluster mode
savedir = getcwd()+"\\stories"
hascuda = False # Whether torch has detected CUDA on the system
usegpu = False # Whether to launch pipeline with GPU support
custmodpth = "" # Filesystem location of custom model to run
formatoptns = {'frmttriminc': True, 'frmtrmblln': False, 'frmtrmspch': False, 'frmtadsnsp': True, 'singleline': False} # Container for state of formatting options
importnum = -1 # Selection on import popup list
importjs = {} # Temporary storage for import data
loadselect = "" # Temporary storage for story filename to load
spselect = "" # Temporary storage for soft prompt filename to load
spmeta = None # Metadata of current soft prompt, or None if not using a soft prompt
sp = None # Current soft prompt tensor (as a NumPy array)
sp_length = 0 # Length of current soft prompt in tokens, or 0 if not using a soft prompt
has_genmod = False # Whether or not at least one loaded Lua userscript has a generation modifier
svowname = "" # Filename that was flagged for overwrite confirm
saveow = False # Whether or not overwrite confirm has been displayed
autosave = False # Whether or not to automatically save after each action
genseqs = [] # Temporary storage for generated sequences
recentback = False # Whether Back button was recently used without Submitting or Retrying after
recentrng = None # If a new random game was recently generated without Submitting after, this is the topic used (as a string), otherwise this is None
recentrngm = None # If a new random game was recently generated without Submitting after, this is the memory used (as a string), otherwise this is None
useprompt = False # Whether to send the full prompt with every submit action
breakmodel = False # For GPU users, whether to use both system RAM and VRAM to conserve VRAM while offering speedup compared to CPU-only
bmsupported = False # Whether the breakmodel option is supported (GPT-Neo/GPT-J/XGLM/OPT only, currently)
nobreakmodel = False # Something specifically requested Breakmodel to be disabled (For example a models config)
smandelete = False # Whether stories can be deleted from inside the browser
smanrename = False # Whether stories can be renamed from inside the browser
allowsp = False # Whether we are allowed to use soft prompts (by default enabled if we're using GPT-2, GPT-Neo or GPT-J)
modeldim = -1 # Embedding dimension of your model (e.g. it's 4096 for GPT-J-6B and 2560 for GPT-Neo-2.7B)
laststory = None # Filename (without extension) of most recent story JSON file we loaded
regex_sl = re.compile(r'\n*(?<=.) *\n(.|\n)*') # Pattern for limiting the output to a single line
acregex_ai = re.compile(r'\n* *>(.|\n)*') # Pattern for matching adventure actions from the AI so we can remove them
acregex_ui = re.compile(r'^ *(>.*)$', re.MULTILINE) # Pattern for matching actions in the HTML-escaped story so we can apply colouring, etc (make sure to encase part to format in parentheses)
comregex_ai = re.compile(r'(?:\n<\|(?:.|\n)*?\|>(?=\n|$))|(?:<\|(?:.|\n)*?\|>\n?)') # Pattern for matching comments to remove them before sending them to the AI
comregex_ui = re.compile(r'(<\|(?:.|\n)*?\|>)') # Pattern for matching comments in the editor
sampler_order = utils.default_sampler_order.copy()
rng_states = {} # Used by the POST /generate endpoint to store sampler RNG states
chatmode = False
chatname = "You"
adventure = False
actionmode = 1
dynamicscan = False
host = False
nopromptgen = False
rngpersist = False
nogenmod = False
welcome = False # Custom Welcome Text (False is default)
newlinemode = "ns"
quiet = False # If set will suppress any story text from being printed to the console (will only be seen on the client web page)
debug = False # If set to true, will send debug information to the client for display
lazy_load = True # Whether or not to use torch_lazy_loader.py for transformers models in order to reduce CPU memory usage
use_colab_tpu = os.environ.get("COLAB_TPU_ADDR", "") != "" or os.environ.get("TPU_NAME", "") != "" # Whether or not we're in a Colab TPU instance or Kaggle TPU instance and are going to use the TPU rather than the CPU
revision = None
standalone = False
api_tokenizer_id = None
disable_set_aibusy = False
disable_input_formatting = False
disable_output_formatting = False
output_streaming = True
token_stream_queue = TokenStreamQueue() # Queue for the token streaming
show_probs = False # Whether or not to show token probabilities
show_budget = False # Whether or not to show token probabilities
configname = None
utils.vars = vars
class Send_to_socketio(object):
def write(self, bar):
print(bar, end="")
time.sleep(0.01)
try:
gui_msg = bar.replace(f"{colors.PURPLE}INIT{colors.END} | ","").replace(" ", " ")
emit('from_server', {'cmd': 'model_load_status', 'data': gui_msg}, broadcast=True)
except:
pass
# Set logging level to reduce chatter from Flask
import logging
log = logging.getLogger('werkzeug')
log.setLevel(logging.ERROR)
from flask import Flask, render_template, Response, request, copy_current_request_context, send_from_directory, session, jsonify, abort, redirect
from flask_socketio import SocketIO
from flask_socketio import emit as _emit
from flask_session import Session
from werkzeug.exceptions import HTTPException, NotFound, InternalServerError
import secrets
app = Flask(__name__, root_path=os.getcwd())
app.secret_key = secrets.token_hex()
app.config['SESSION_TYPE'] = 'filesystem'
app.config['TEMPLATES_AUTO_RELOAD'] = True
socketio = SocketIO(app, async_method="eventlet")
old_socketio_on = socketio.on
def new_socketio_on(*a, **k):
decorator = old_socketio_on(*a, **k)
def new_decorator(f):
@functools.wraps(f)
def g(*a, **k):
if args.no_ui:
return
return f(*a, **k)
return decorator(g)
return new_decorator
socketio.on = new_socketio_on
def emit(*args, **kwargs):
try:
return _emit(*args, **kwargs)
except AttributeError:
return socketio.emit(*args, **kwargs)
utils.emit = emit
# marshmallow/apispec setup
from apispec import APISpec
from apispec.ext.marshmallow import MarshmallowPlugin
from apispec.ext.marshmallow.field_converter import make_min_max_attributes
from apispec_webframeworks.flask import FlaskPlugin
from marshmallow import Schema, fields, validate, EXCLUDE
from marshmallow.exceptions import ValidationError
class KoboldSchema(Schema):
pass
def new_make_min_max_attributes(validators, min_attr, max_attr) -> dict:
# Patched apispec function that creates "exclusiveMinimum"/"exclusiveMaximum" OpenAPI attributes insteaed of "minimum"/"maximum" when using validators.Range or validators.Length with min_inclusive=False or max_inclusive=False
attributes = {}
min_list = [validator.min for validator in validators if validator.min is not None]
max_list = [validator.max for validator in validators if validator.max is not None]
min_inclusive_list = [getattr(validator, "min_inclusive", True) for validator in validators if validator.min is not None]
max_inclusive_list = [getattr(validator, "max_inclusive", True) for validator in validators if validator.max is not None]
if min_list:
if min_attr == "minimum" and not min_inclusive_list[max(range(len(min_list)), key=min_list.__getitem__)]:
min_attr = "exclusiveMinimum"
attributes[min_attr] = max(min_list)
if max_list:
if min_attr == "maximum" and not max_inclusive_list[min(range(len(max_list)), key=max_list.__getitem__)]:
min_attr = "exclusiveMaximum"
attributes[max_attr] = min(max_list)
return attributes
make_min_max_attributes.__code__ = new_make_min_max_attributes.__code__
def api_format_docstring(f):
f.__doc__ = eval('f"""{}"""'.format(f.__doc__.replace("\\", "\\\\")))
return f
def api_catch_out_of_memory_errors(f):
@functools.wraps(f)
def decorated(*args, **kwargs):
try:
return f(*args, **kwargs)
except Exception as e:
if any (s in traceback.format_exc().lower() for s in ("out of memory", "not enough memory")):
for line in reversed(traceback.format_exc().split("\n")):
if any(s in line.lower() for s in ("out of memory", "not enough memory")) and line.count(":"):
line = line.split(":", 1)[1]
line = re.sub(r"\[.+?\] +data\.", "", line).strip()
raise KoboldOutOfMemoryError("KoboldAI ran out of memory: " + line, type="out_of_memory.gpu.cuda" if "cuda out of memory" in line.lower() else "out_of_memory.gpu.hip" if "hip out of memory" in line.lower() else "out_of_memory.tpu.hbm" if "memory space hbm" in line.lower() else "out_of_memory.cpu.default_memory_allocator" if "defaultmemoryallocator" in line.lower() else "out_of_memory.unknown.unknown")
raise KoboldOutOfMemoryError(type="out_of_memory.unknown.unknown")
raise e
return decorated
def api_schema_wrap(f):
try:
input_schema: Type[Schema] = next(iter(inspect.signature(f).parameters.values())).annotation
except:
HAS_SCHEMA = False
else:
HAS_SCHEMA = inspect.isclass(input_schema) and issubclass(input_schema, Schema)
f = api_format_docstring(f)
f = api_catch_out_of_memory_errors(f)
@functools.wraps(f)
def decorated(*args, **kwargs):
if HAS_SCHEMA:
body = request.get_json()
schema = input_schema.from_dict(input_schema().load(body))
response = f(schema, *args, **kwargs)
else:
response = f(*args, **kwargs)
if not isinstance(response, Response):
response = jsonify(response)
return response
return decorated
@app.errorhandler(HTTPException)
def handler(e):
if request.path != "/api" and not request.path.startswith("/api/"):
return e
resp = jsonify(detail={"msg": str(e), "type": "generic.error_" + str(e.code)})
if e.code == 405 and e.valid_methods is not None:
resp.headers["Allow"] = ", ".join(e.valid_methods)
return resp, e.code
class KoboldOutOfMemoryError(HTTPException):
code = 507
description = "KoboldAI ran out of memory."
type = "out_of_memory.unknown.unknown"
def __init__(self, *args, type=None, **kwargs):
super().__init__(*args, **kwargs)
if type is not None:
self.type = type
@app.errorhandler(KoboldOutOfMemoryError)
def handler(e):
if request.path != "/api" and not request.path.startswith("/api/"):
return InternalServerError()
return jsonify(detail={"type": e.type, "msg": e.description}), e.code
@app.errorhandler(ValidationError)
def handler(e):
if request.path != "/api" and not request.path.startswith("/api/"):
return InternalServerError()
return jsonify(detail=e.messages), 422
@app.errorhandler(NotImplementedError)
def handler(e):
if request.path != "/api" and not request.path.startswith("/api/"):
return InternalServerError()
return jsonify(detail={"type": "not_implemented", "msg": str(e).strip()}), 501
api_versions: List[str] = []
class KoboldAPISpec(APISpec):
class KoboldFlaskPlugin(FlaskPlugin):
def __init__(self, api: "KoboldAPISpec", *args, **kwargs):
self._kobold_api_spec = api
super().__init__(*args, **kwargs)
def path_helper(self, *args, **kwargs):
return super().path_helper(*args, **kwargs)[len(self._kobold_api_spec._prefixes[0]):]
def __init__(self, *args, title: str = "KoboldAI API", openapi_version: str = "3.0.3", version: str = "1.0.0", prefixes: List[str] = None, **kwargs):
plugins = [KoboldAPISpec.KoboldFlaskPlugin(self), MarshmallowPlugin()]
self._prefixes = prefixes if prefixes is not None else [""]
self._kobold_api_spec_version = version
api_versions.append(version)
api_versions.sort(key=lambda x: [int(e) for e in x.split(".")])
super().__init__(*args, title=title, openapi_version=openapi_version, version=version, plugins=plugins, servers=[{"url": self._prefixes[0]}], **kwargs)
for prefix in self._prefixes:
app.route(prefix, endpoint="~KoboldAPISpec~" + prefix)(lambda: redirect(request.path + "/docs/"))
app.route(prefix + "/", endpoint="~KoboldAPISpec~" + prefix + "/")(lambda: redirect("docs/"))
app.route(prefix + "/docs", endpoint="~KoboldAPISpec~" + prefix + "/docs")(lambda: redirect("docs/"))
app.route(prefix + "/docs/", endpoint="~KoboldAPISpec~" + prefix + "/docs/")(lambda: render_template("swagger-ui.html", url=self._prefixes[0] + "/openapi.json"))
app.route(prefix + "/openapi.json", endpoint="~KoboldAPISpec~" + prefix + "/openapi.json")(lambda: jsonify(self.to_dict()))
def route(self, rule: str, methods=["GET"], **kwargs):
__F = TypeVar("__F", bound=Callable[..., Any])
if "strict_slashes" not in kwargs:
kwargs["strict_slashes"] = False
def new_decorator(f: __F) -> __F:
@functools.wraps(f)
def g(*args, **kwargs):
global api_version
api_version = self._kobold_api_spec_version
try:
return f(*args, **kwargs)
finally:
api_version = None
for prefix in self._prefixes:
g = app.route(prefix + rule, methods=methods, **kwargs)(g)
with app.test_request_context():
self.path(view=g, **kwargs)
return g
return new_decorator
def get(self, rule: str, **kwargs):
return self.route(rule, methods=["GET"], **kwargs)
def post(self, rule: str, **kwargs):
return self.route(rule, methods=["POST"], **kwargs)
def put(self, rule: str, **kwargs):
return self.route(rule, methods=["PUT"], **kwargs)
def patch(self, rule: str, **kwargs):
return self.route(rule, methods=["PATCH"], **kwargs)
def delete(self, rule: str, **kwargs):
return self.route(rule, methods=["DELETE"], **kwargs)
tags = [
{"name": "info", "description": "Metadata about this API"},
{"name": "generate", "description": "Text generation endpoints"},
{"name": "model", "description": "Information about the current text generation model"},
{"name": "story", "description": "Endpoints for managing the story in the KoboldAI GUI"},
{"name": "world_info", "description": "Endpoints for managing the world info in the KoboldAI GUI"},
{"name": "config", "description": "Allows you to get/set various setting values"},
]
api_version = None # This gets set automatically so don't change this value
api_v1 = KoboldAPISpec(
version="1.2.1",
prefixes=["/api/v1", "/api/latest"],
tags=tags,
)
# Returns the expected config filename for the current setup.
# If the model_name is specified, it returns what the settings file would be for that model
def get_config_filename(model_name = None):
if model_name:
return(f"settings/{model_name.replace('/', '_')}.settings")
elif args.configname:
return(f"settings/{args.configname.replace('/', '_')}.settings")
elif vars.configname != '':
return(f"settings/{vars.configname.replace('/', '_')}.settings")
else:
logger.warning(f"Empty configfile name sent back. Defaulting to ReadOnly")
return(f"settings/ReadOnly.settings")
#==================================================================#
# Function to get model selection at startup
#==================================================================#
def sendModelSelection(menu="mainmenu", folder="./models"):
#If we send one of the manual load options, send back the list of model directories, otherwise send the menu
if menu in ('NeoCustom', 'GPT2Custom'):
(paths, breadcrumbs) = get_folder_path_info(folder)
if vars.host:
breadcrumbs = []
menu_list = [[folder, menu, "", False] for folder in paths]
menu_list.append(["Return to Main Menu", "mainmenu", "", True])
if os.path.abspath("{}/models".format(os.getcwd())) == os.path.abspath(folder):
showdelete=True
else:
showdelete=False
emit('from_server', {'cmd': 'show_model_menu', 'data': menu_list, 'menu': menu, 'breadcrumbs': breadcrumbs, "showdelete": showdelete}, broadcast=True)
else:
emit('from_server', {'cmd': 'show_model_menu', 'data': model_menu[menu], 'menu': menu, 'breadcrumbs': [], "showdelete": False}, broadcast=True)
def get_folder_path_info(base):
if base == 'This PC':
breadcrumbs = [['This PC', 'This PC']]
paths = [["{}:\\".format(chr(i)), "{}:\\".format(chr(i))] for i in range(65, 91) if os.path.exists("{}:".format(chr(i)))]
else:
path = os.path.abspath(base)
if path[-1] == "\\":
path = path[:-1]
breadcrumbs = []
for i in range(len(path.replace("/", "\\").split("\\"))):
breadcrumbs.append(["\\".join(path.replace("/", "\\").split("\\")[:i+1]),
path.replace("/", "\\").split("\\")[i]])
if len(breadcrumbs) == 1:
breadcrumbs = [["{}:\\".format(chr(i)), "{}:\\".format(chr(i))] for i in range(65, 91) if os.path.exists("{}:".format(chr(i)))]
else:
if len([["{}:\\".format(chr(i)), "{}:\\".format(chr(i))] for i in range(65, 91) if os.path.exists("{}:".format(chr(i)))]) > 0:
breadcrumbs.insert(0, ['This PC', 'This PC'])
paths = []
base_path = os.path.abspath(base)
for item in os.listdir(base_path):
if os.path.isdir(os.path.join(base_path, item)):
paths.append([os.path.join(base_path, item), item])
# Paths/breadcrumbs is a list of lists, where the first element in the sublist is the full path and the second is the folder name
return (paths, breadcrumbs)
def getModelSelection(modellist):
print(" # Model\t\t\t\t\t\tVRAM\n ========================================================")
i = 1
for m in modellist:
print(" {0} - {1}\t\t\t{2}".format("{:<2}".format(i), m[0].ljust(25), m[2]))
i += 1
print(" ");
modelsel = 0
vars.model = ''
while(vars.model == ''):
modelsel = input("Model #> ")
if(modelsel.isnumeric() and int(modelsel) > 0 and int(modelsel) <= len(modellist)):
vars.model = modellist[int(modelsel)-1][1]
else:
print("{0}Please enter a valid selection.{1}".format(colors.RED, colors.END))
# Model Lists
try:
getModelSelection(eval(vars.model))
except Exception as e:
if(vars.model == "Return"):
getModelSelection(mainmenu)
# If custom model was selected, get the filesystem location and store it
if(vars.model == "NeoCustom" or vars.model == "GPT2Custom"):
print("{0}Please choose the folder where pytorch_model.bin is located:{1}\n".format(colors.CYAN, colors.END))
modpath = fileops.getdirpath(getcwd() + "/models", "Select Model Folder")
if(modpath):
# Save directory to vars
vars.custmodpth = modpath
else:
# Print error and retry model selection
print("{0}Model select cancelled!{1}".format(colors.RED, colors.END))
print("{0}Select an AI model to continue:{1}\n".format(colors.CYAN, colors.END))
getModelSelection(mainmenu)
def check_if_dir_is_model(path):
return os.path.exists(os.path.join(path, 'config.json'))
#==================================================================#
# Return all keys in tokenizer dictionary containing char
#==================================================================#
#def gettokenids(char):
# keys = []
# for key in vocab_keys:
# if(key.find(char) != -1):
# keys.append(key)
# return keys
#==================================================================#
# Return Model Name
#==================================================================#
def getmodelname():
if(vars.online_model != ''):
return(f"{vars.model}/{vars.online_model}")
if(vars.model in ("NeoCustom", "GPT2Custom", "TPUMeshTransformerGPTJ", "TPUMeshTransformerGPTNeoX")):
modelname = os.path.basename(os.path.normpath(vars.custmodpth))
return modelname
else:
modelname = vars.model
return modelname
#==================================================================#
# Get hidden size from model
#==================================================================#
def get_hidden_size_from_model(model):
return model.get_input_embeddings().embedding_dim
#==================================================================#
# Breakmodel configuration functions
#==================================================================#
def device_list(n_layers, primary=None, selected=None):
device_count = torch.cuda.device_count()
if(device_count < 2):
primary = None
gpu_blocks = breakmodel.gpu_blocks + (device_count - len(breakmodel.gpu_blocks))*[0]
print(f"{colors.YELLOW} DEVICE ID | LAYERS | DEVICE NAME{colors.END}")
for i in range(device_count):
name = torch.cuda.get_device_name(i)
if(len(name) > 47):
name = "..." + name[-44:]
row_color = colors.END
sep_color = colors.YELLOW
print(f"{row_color}{colors.YELLOW + '->' + row_color if i == selected else ' '} {'(primary)' if i == primary else ' '*9} {i:3} {sep_color}|{row_color} {gpu_blocks[i]:3} {sep_color}|{row_color} {name}{colors.END}")
row_color = colors.END
sep_color = colors.YELLOW
if(utils.HAS_ACCELERATE):
print(f"{row_color}{colors.YELLOW + '->' + row_color if -1 == selected else ' '} {' '*9} N/A {sep_color}|{row_color} {breakmodel.disk_blocks:3} {sep_color}|{row_color} (Disk cache){colors.END}")
print(f"{row_color} {' '*9} N/A {sep_color}|{row_color} {n_layers:3} {sep_color}|{row_color} (CPU){colors.END}")
def device_config(config):
global breakmodel, generator
import breakmodel
n_layers = utils.num_layers(config)
if args.cpu:
breakmodel.gpu_blocks = [0]*n_layers
return
elif(args.breakmodel_gpulayers is not None or (utils.HAS_ACCELERATE and args.breakmodel_disklayers is not None)):
try:
if(not args.breakmodel_gpulayers):
breakmodel.gpu_blocks = []
else:
breakmodel.gpu_blocks = list(map(int, args.breakmodel_gpulayers.split(',')))
assert len(breakmodel.gpu_blocks) <= torch.cuda.device_count()
s = n_layers
for i in range(len(breakmodel.gpu_blocks)):
if(breakmodel.gpu_blocks[i] <= -1):
breakmodel.gpu_blocks[i] = s
break
else:
s -= breakmodel.gpu_blocks[i]
assert sum(breakmodel.gpu_blocks) <= n_layers
n_layers -= sum(breakmodel.gpu_blocks)
if(args.breakmodel_disklayers is not None):
assert args.breakmodel_disklayers <= n_layers
breakmodel.disk_blocks = args.breakmodel_disklayers
n_layers -= args.breakmodel_disklayers
except:
logger.warning("--breakmodel_gpulayers is malformatted. Please use the --help option to see correct usage of --breakmodel_gpulayers. Defaulting to all layers on device 0.")
breakmodel.gpu_blocks = [n_layers]
n_layers = 0
elif(args.breakmodel_layers is not None):
breakmodel.gpu_blocks = [n_layers - max(0, min(n_layers, args.breakmodel_layers))]
n_layers -= sum(breakmodel.gpu_blocks)
elif(args.model is not None):
logger.info("Breakmodel not specified, assuming GPU 0")
breakmodel.gpu_blocks = [n_layers]
n_layers = 0
else:
device_count = torch.cuda.device_count()
if(device_count > 1):
print(colors.CYAN + "\nPlease select one of your GPUs to be your primary GPU.")
print("VRAM usage in your primary GPU will be higher than for your other ones.")
print("It is recommended you make your fastest GPU your primary GPU.")
device_list(n_layers)
while(True):
primaryselect = input("device ID> ")
if(primaryselect.isnumeric() and 0 <= int(primaryselect) < device_count):
breakmodel.primary_device = int(primaryselect)
break
else:
print(f"{colors.RED}Please enter an integer between 0 and {device_count-1}.{colors.END}")
else:
breakmodel.primary_device = 0
print(colors.PURPLE + "\nIf you don't have enough VRAM to run the model on a single GPU")
print("you can split the model between your CPU and your GPU(s), or between")
print("multiple GPUs if you have more than one.")
print("By putting more 'layers' on a GPU or CPU, more computations will be")
print("done on that device and more VRAM or RAM will be required on that device")
print("(roughly proportional to number of layers).")
print("It should be noted that GPUs are orders of magnitude faster than the CPU.")
print(f"This model has{colors.YELLOW} {n_layers} {colors.PURPLE}layers.{colors.END}\n")
for i in range(device_count):
device_list(n_layers, primary=breakmodel.primary_device, selected=i)
print(f"{colors.CYAN}\nHow many of the remaining{colors.YELLOW} {n_layers} {colors.CYAN}layers would you like to put into device {i}?\nYou can also enter -1 to allocate all remaining layers to this device.{colors.END}\n")
while(True):
layerselect = input("# of layers> ")
if((layerselect.isnumeric() or layerselect.strip() == '-1') and -1 <= int(layerselect) <= n_layers):
layerselect = int(layerselect)
layerselect = n_layers if layerselect == -1 else layerselect
breakmodel.gpu_blocks.append(layerselect)
n_layers -= layerselect
break
else:
print(f"{colors.RED}Please enter an integer between -1 and {n_layers}.{colors.END}")
if(n_layers == 0):
break
if(utils.HAS_ACCELERATE and n_layers > 0):
device_list(n_layers, primary=breakmodel.primary_device, selected=-1)
print(f"{colors.CYAN}\nHow many of the remaining{colors.YELLOW} {n_layers} {colors.CYAN}layers would you like to put into the disk cache?\nYou can also enter -1 to allocate all remaining layers to this device.{colors.END}\n")
while(True):
layerselect = input("# of layers> ")
if((layerselect.isnumeric() or layerselect.strip() == '-1') and -1 <= int(layerselect) <= n_layers):
layerselect = int(layerselect)
layerselect = n_layers if layerselect == -1 else layerselect
breakmodel.disk_blocks = layerselect
n_layers -= layerselect
break
else:
print(f"{colors.RED}Please enter an integer between -1 and {n_layers}.{colors.END}")
logger.init_ok("Final device configuration:", status="Info")
device_list(n_layers, primary=breakmodel.primary_device)
# If all layers are on the same device, use the old GPU generation mode
while(len(breakmodel.gpu_blocks) and breakmodel.gpu_blocks[-1] == 0):
breakmodel.gpu_blocks.pop()
if(len(breakmodel.gpu_blocks) and breakmodel.gpu_blocks[-1] in (-1, utils.num_layers(config))):
vars.breakmodel = False
vars.usegpu = True
vars.gpu_device = len(breakmodel.gpu_blocks)-1
return
if(not breakmodel.gpu_blocks):
logger.warning("Nothing assigned to a GPU, reverting to CPU only mode")
import breakmodel
breakmodel.primary_device = "cpu"
vars.breakmodel = False
vars.usegpu = False
return
def move_model_to_devices(model):
global generator
if(not utils.HAS_ACCELERATE and not vars.breakmodel):
if(vars.usegpu):
model = model.half().to(vars.gpu_device)
else:
model = model.to('cpu').float()
generator = model.generate
return
import breakmodel
if(utils.HAS_ACCELERATE):
import accelerate.utils
for key, value in model.state_dict().items():
target_dtype = torch.float32 if breakmodel.primary_device == "cpu" else torch.float16
if(value.dtype is not target_dtype):
accelerate.utils.set_module_tensor_to_device(model, key, target_dtype)
disk_blocks = breakmodel.disk_blocks
gpu_blocks = breakmodel.gpu_blocks
ram_blocks = len(utils.layers_module_names) - sum(gpu_blocks)
cumulative_gpu_blocks = tuple(itertools.accumulate(gpu_blocks))
device_map = {}
for name in utils.layers_module_names:
layer = int(name.rsplit(".", 1)[1])
device = ("disk" if layer < disk_blocks else "cpu") if layer < ram_blocks else bisect.bisect_right(cumulative_gpu_blocks, layer - ram_blocks)
device_map[name] = device
for name in utils.get_missing_module_names(model, list(device_map.keys())):
device_map[name] = breakmodel.primary_device
breakmodel.dispatch_model_ex(model, device_map, main_device=breakmodel.primary_device, offload_buffers=True, offload_dir="accelerate-disk-cache")
gc.collect()
generator = model.generate
return
model.half()
gc.collect()
if(hasattr(model, "transformer")):
model.transformer.wte.to(breakmodel.primary_device)
model.transformer.ln_f.to(breakmodel.primary_device)
if(hasattr(model, 'lm_head')):
model.lm_head.to(breakmodel.primary_device)
if(hasattr(model.transformer, 'wpe')):
model.transformer.wpe.to(breakmodel.primary_device)
elif(not hasattr(model.model, "decoder")):
model.model.embed_tokens.to(breakmodel.primary_device)
model.model.layer_norm.to(breakmodel.primary_device)
model.lm_head.to(breakmodel.primary_device)
model.model.embed_positions.to(breakmodel.primary_device)
else:
model.model.decoder.embed_tokens.to(breakmodel.primary_device)
if(model.model.decoder.project_in is not None):
model.model.decoder.project_in.to(breakmodel.primary_device)
if(model.model.decoder.project_out is not None):
model.model.decoder.project_out.to(breakmodel.primary_device)
model.model.decoder.embed_positions.to(breakmodel.primary_device)
gc.collect()
GPTNeoModel.forward = breakmodel.new_forward_neo
if("GPTJModel" in globals()):
GPTJModel.forward = breakmodel.new_forward_neo # type: ignore
if("XGLMModel" in globals()):
XGLMModel.forward = breakmodel.new_forward_xglm # type: ignore
if("OPTDecoder" in globals()):
OPTDecoder.forward = breakmodel.new_forward_opt # type: ignore
generator = model.generate
if(hasattr(model, "transformer")):
breakmodel.move_hidden_layers(model.transformer)
elif(not hasattr(model.model, "decoder")):
breakmodel.move_hidden_layers(model.model, model.model.layers)
else:
breakmodel.move_hidden_layers(model.model.decoder, model.model.decoder.layers)
#==================================================================#
# Allow the models to override some settings
#==================================================================#
def loadmodelsettings():
try: