forked from LostRuins/koboldcpp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
koboldcpp.py
executable file
·1941 lines (1707 loc) · 89.6 KB
/
koboldcpp.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/env python3
#-*- coding: utf-8 -*-
# KoboldCpp is an easy-to-use AI text-generation software for GGML models.
# It's a single self contained distributable from Concedo, that builds off llama.cpp,
# and adds a versatile Kobold API endpoint, additional format support,
# backward compatibility, as well as a fancy UI with persistent stories,
# editing tools, save formats, memory, world info, author's note, characters,
# scenarios and everything Kobold and Kobold Lite have to offer.
import ctypes
import os
import argparse
import json, sys, http.server, time, asyncio, socket, threading
from concurrent.futures import ThreadPoolExecutor
sampler_order_max = 7
stop_token_max = 16
ban_token_max = 16
tensor_split_max = 16
class load_model_inputs(ctypes.Structure):
_fields_ = [("threads", ctypes.c_int),
("blasthreads", ctypes.c_int),
("max_context_length", ctypes.c_int),
("batch_size", ctypes.c_int),
("f16_kv", ctypes.c_bool),
("low_vram", ctypes.c_bool),
("use_mmq", ctypes.c_bool),
("executable_path", ctypes.c_char_p),
("model_filename", ctypes.c_char_p),
("lora_filename", ctypes.c_char_p),
("lora_base", ctypes.c_char_p),
("use_mmap", ctypes.c_bool),
("use_mlock", ctypes.c_bool),
("use_smartcontext", ctypes.c_bool),
("clblast_info", ctypes.c_int),
("cublas_info", ctypes.c_int),
("blasbatchsize", ctypes.c_int),
("debugmode", ctypes.c_int),
("forceversion", ctypes.c_int),
("gpulayers", ctypes.c_int),
("rope_freq_scale", ctypes.c_float),
("rope_freq_base", ctypes.c_float),
("banned_tokens", ctypes.c_char_p * ban_token_max),
("tensor_split", ctypes.c_float * tensor_split_max)]
class generation_inputs(ctypes.Structure):
_fields_ = [("seed", ctypes.c_int),
("prompt", ctypes.c_char_p),
("max_context_length", ctypes.c_int),
("max_length", ctypes.c_int),
("temperature", ctypes.c_float),
("top_k", ctypes.c_int),
("top_a", ctypes.c_float),
("top_p", ctypes.c_float),
("typical_p", ctypes.c_float),
("tfs", ctypes.c_float),
("rep_pen", ctypes.c_float),
("rep_pen_range", ctypes.c_int),
("mirostat", ctypes.c_int),
("mirostat_tau", ctypes.c_float),
("mirostat_eta", ctypes.c_float),
("sampler_order", ctypes.c_int * sampler_order_max),
("sampler_len", ctypes.c_int),
("unban_tokens_rt", ctypes.c_bool),
("stop_sequence", ctypes.c_char_p * stop_token_max),
("stream_sse", ctypes.c_bool),
("grammar", ctypes.c_char_p),
("grammar_retain_state", ctypes.c_bool)]
class generation_outputs(ctypes.Structure):
_fields_ = [("status", ctypes.c_int),
("text", ctypes.c_char * 24576)]
handle = None
def getdirpath():
return os.path.dirname(os.path.realpath(__file__))
def getabspath():
return os.path.dirname(os.path.abspath(__file__))
def file_exists(filename):
return os.path.exists(os.path.join(getdirpath(), filename))
def pick_existant_file(ntoption,nonntoption):
precompiled_prefix = "precompiled_"
ntexist = file_exists(ntoption)
nonntexist = file_exists(nonntoption)
precompiled_ntexist = file_exists(precompiled_prefix+ntoption)
precompiled_nonntexist = file_exists(precompiled_prefix+nonntoption)
if os.name == 'nt':
if not ntexist and precompiled_ntexist:
return (precompiled_prefix+ntoption)
if nonntexist and not ntexist:
return nonntoption
return ntoption
else:
if not nonntexist and precompiled_nonntexist:
return (precompiled_prefix+nonntoption)
if ntexist and not nonntexist:
return ntoption
return nonntoption
lib_default = pick_existant_file("koboldcpp_default.dll","koboldcpp_default.so")
lib_failsafe = pick_existant_file("koboldcpp_failsafe.dll","koboldcpp_failsafe.so")
lib_openblas = pick_existant_file("koboldcpp_openblas.dll","koboldcpp_openblas.so")
lib_noavx2 = pick_existant_file("koboldcpp_noavx2.dll","koboldcpp_noavx2.so")
lib_clblast = pick_existant_file("koboldcpp_clblast.dll","koboldcpp_clblast.so")
lib_cublas = pick_existant_file("koboldcpp_cublas.dll","koboldcpp_cublas.so")
lib_hipblas = pick_existant_file("koboldcpp_hipblas.dll","koboldcpp_hipblas.so")
def init_library():
global handle, args
global lib_default,lib_failsafe,lib_openblas,lib_noavx2,lib_clblast,lib_cublas
libname = ""
use_openblas = False # if true, uses OpenBLAS for acceleration. libopenblas.dll must exist in the same dir.
use_clblast = False #uses CLBlast instead
use_cublas = False #uses cublas instead
use_hipblas = False #uses hipblas instead
use_noavx2 = False #uses no avx2 instructions
use_failsafe = False #uses no intrinsics, failsafe mode
if args.noavx2:
use_noavx2 = True
if not file_exists(lib_noavx2):
print("Warning: NoAVX2 library file not found. Failsafe library will be used.")
elif (args.noblas and args.nommap):
use_failsafe = True
print("!!! Attempting to use FAILSAFE MODE !!!")
else:
print("Attempting to use non-avx2 compatibility library.")
elif args.useclblast:
if not file_exists(lib_clblast) or (os.name=='nt' and not file_exists("clblast.dll")):
print("Warning: CLBlast library file not found. Non-BLAS library will be used.")
else:
print("Attempting to use CLBlast library for faster prompt ingestion. A compatible clblast will be required.")
use_clblast = True
elif (args.usecublas is not None):
if not file_exists(lib_cublas) and not file_exists(lib_hipblas):
print("Warning: CuBLAS library file not found. Non-BLAS library will be used.")
else:
if file_exists(lib_cublas):
print("Attempting to use CuBLAS library for faster prompt ingestion. A compatible CuBLAS will be required.")
use_cublas = True
elif file_exists(lib_hipblas):
print("Attempting to use hipBLAS library for faster prompt ingestion. A compatible AMD GPU will be required.")
use_hipblas = True
else:
if not file_exists(lib_openblas) or (os.name=='nt' and not file_exists("libopenblas.dll")):
print("Warning: OpenBLAS library file not found. Non-BLAS library will be used.")
elif args.noblas:
print("Attempting to library without OpenBLAS.")
else:
use_openblas = True
print("Attempting to use OpenBLAS library for faster prompt ingestion. A compatible libopenblas will be required.")
if sys.platform=="darwin":
print("Mac OSX note: Some people have found Accelerate actually faster than OpenBLAS. To compare, run Koboldcpp with --noblas instead.")
if use_noavx2:
if use_failsafe:
libname = lib_failsafe
else:
libname = lib_noavx2
else:
if use_clblast:
libname = lib_clblast
elif use_cublas:
libname = lib_cublas
elif use_hipblas:
libname = lib_hipblas
elif use_openblas:
libname = lib_openblas
else:
libname = lib_default
print("Initializing dynamic library: " + libname)
dir_path = getdirpath()
abs_path = getabspath()
#add all potential paths
if os.name=='nt':
os.add_dll_directory(dir_path)
os.add_dll_directory(abs_path)
os.add_dll_directory(os.getcwd())
if libname == lib_hipblas and "HIP_PATH" in os.environ:
os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "bin"))
if args.debugmode == 1:
print(f"HIP/ROCm SDK at {os.environ['HIP_PATH']} included in .DLL load path")
handle = ctypes.CDLL(os.path.join(dir_path, libname))
handle.load_model.argtypes = [load_model_inputs]
handle.load_model.restype = ctypes.c_bool
handle.generate.argtypes = [generation_inputs, ctypes.c_wchar_p] #apparently needed for osx to work. i duno why they need to interpret it that way but whatever
handle.generate.restype = generation_outputs
handle.new_token.restype = ctypes.c_char_p
handle.new_token.argtypes = [ctypes.c_int]
handle.get_stream_count.restype = ctypes.c_int
handle.has_finished.restype = ctypes.c_bool
handle.get_last_eval_time.restype = ctypes.c_float
handle.get_last_process_time.restype = ctypes.c_float
handle.get_last_token_count.restype = ctypes.c_int
handle.get_last_stop_reason.restype = ctypes.c_int
handle.abort_generate.restype = ctypes.c_bool
handle.token_count.restype = ctypes.c_int
handle.get_pending_output.restype = ctypes.c_char_p
def load_model(model_filename):
global args
inputs = load_model_inputs()
inputs.model_filename = model_filename.encode("UTF-8")
inputs.batch_size = 8
inputs.max_context_length = maxctx #initial value to use for ctx, can be overwritten
inputs.threads = args.threads
inputs.low_vram = (True if (args.usecublas and "lowvram" in args.usecublas) else False)
inputs.use_mmq = (True if (args.usecublas and "mmq" in args.usecublas) else False)
inputs.blasthreads = args.blasthreads
inputs.f16_kv = True
inputs.use_mmap = (not args.nommap)
inputs.use_mlock = args.usemlock
inputs.lora_filename = "".encode("UTF-8")
inputs.lora_base = "".encode("UTF-8")
if args.lora:
inputs.lora_filename = args.lora[0].encode("UTF-8")
inputs.use_mmap = False
if len(args.lora) > 1:
inputs.lora_base = args.lora[1].encode("UTF-8")
inputs.use_smartcontext = args.smartcontext
inputs.blasbatchsize = args.blasbatchsize
inputs.forceversion = args.forceversion
inputs.gpulayers = args.gpulayers
inputs.rope_freq_scale = args.ropeconfig[0]
if len(args.ropeconfig)>1:
inputs.rope_freq_base = args.ropeconfig[1]
else:
inputs.rope_freq_base = 10000
clblastids = 0
if args.useclblast:
clblastids = 100 + int(args.useclblast[0])*10 + int(args.useclblast[1])
inputs.clblast_info = clblastids
for n in range(tensor_split_max):
if args.tensor_split and n < len(args.tensor_split):
inputs.tensor_split[n] = float(args.tensor_split[n])
else:
inputs.tensor_split[n] = 0
# we must force an explicit tensor split
# otherwise the default will divide equally and multigpu crap will slow it down badly
inputs.cublas_info = 0
if not args.tensor_split:
if (args.usecublas and "0" in args.usecublas):
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
os.environ["HIP_VISIBLE_DEVICES"] = "0"
elif (args.usecublas and "1" in args.usecublas):
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
os.environ["HIP_VISIBLE_DEVICES"] = "1"
elif (args.usecublas and "2" in args.usecublas):
os.environ["CUDA_VISIBLE_DEVICES"] = "2"
os.environ["HIP_VISIBLE_DEVICES"] = "2"
elif (args.usecublas and "3" in args.usecublas):
os.environ["CUDA_VISIBLE_DEVICES"] = "3"
os.environ["HIP_VISIBLE_DEVICES"] = "3"
else:
if (args.usecublas and "0" in args.usecublas):
inputs.cublas_info = 0
elif (args.usecublas and "1" in args.usecublas):
inputs.cublas_info = 1
elif (args.usecublas and "2" in args.usecublas):
inputs.cublas_info = 2
elif (args.usecublas and "3" in args.usecublas):
inputs.cublas_info = 3
inputs.executable_path = (getdirpath()+"/").encode("UTF-8")
inputs.debugmode = args.debugmode
banned_tokens = args.bantokens
for n in range(ban_token_max):
if not banned_tokens or n >= len(banned_tokens):
inputs.banned_tokens[n] = "".encode("UTF-8")
else:
inputs.banned_tokens[n] = banned_tokens[n].encode("UTF-8")
ret = handle.load_model(inputs)
return ret
def generate(prompt,max_length=20, max_context_length=512, temperature=0.8, top_k=120, top_a=0.0, top_p=0.85, typical_p=1.0, tfs=1.0, rep_pen=1.1, rep_pen_range=128, mirostat=0, mirostat_tau=5.0, mirostat_eta=0.1, sampler_order=[6,0,1,3,4,2,5], seed=-1, stop_sequence=[], use_default_badwordsids=False, stream_sse=False, grammar='', grammar_retain_state=False, genkey=''):
global maxctx, args, currentusergenkey, totalgens
inputs = generation_inputs()
outputs = ctypes.create_unicode_buffer(ctypes.sizeof(generation_outputs))
inputs.prompt = prompt.encode("UTF-8")
if max_length >= max_context_length:
max_length = max_context_length-1
inputs.max_context_length = max_context_length # this will resize the context buffer if changed
global showmaxctxwarning
if showmaxctxwarning and max_context_length > maxctx:
print(f"\n(Warning! Request max_context_length={max_context_length} exceeds allocated context size of {maxctx}. Consider launching with increased --contextsize to avoid errors. This message will only show once per session.)")
showmaxctxwarning = False
inputs.max_length = max_length
inputs.temperature = temperature
inputs.top_k = top_k
inputs.top_a = top_a
inputs.top_p = top_p
inputs.typical_p = typical_p
inputs.tfs = tfs
inputs.rep_pen = rep_pen
inputs.rep_pen_range = rep_pen_range
inputs.stream_sse = stream_sse
inputs.grammar = grammar.encode("UTF-8")
inputs.grammar_retain_state = grammar_retain_state
inputs.unban_tokens_rt = not use_default_badwordsids
if mirostat in (1, 2):
inputs.mirostat = mirostat
inputs.mirostat_tau = mirostat_tau
inputs.mirostat_eta = mirostat_eta
else:
inputs.mirostat = inputs.mirostat_tau = inputs.mirostat_eta = 0
if sampler_order and 0 < len(sampler_order) <= sampler_order_max:
try:
for i, sampler in enumerate(sampler_order):
inputs.sampler_order[i] = sampler
inputs.sampler_len = len(sampler_order)
global showsamplerwarning
if showsamplerwarning and inputs.mirostat==0 and inputs.sampler_len>0 and (inputs.sampler_order[0]!=6 or inputs.sampler_order[inputs.sampler_len-1]!=5):
print("\n(Note: Sub-optimal sampler_order detected. You may have reduced quality. Recommended sampler values are [6,0,1,3,4,2,5]. This message will only show once per session.)")
showsamplerwarning = False
except TypeError as e:
print("ERROR: sampler_order must be a list of integers: " + str(e))
inputs.seed = seed
for n in range(stop_token_max):
if not stop_sequence or n >= len(stop_sequence):
inputs.stop_sequence[n] = "".encode("UTF-8")
else:
inputs.stop_sequence[n] = stop_sequence[n].encode("UTF-8")
currentusergenkey = genkey
totalgens += 1
ret = handle.generate(inputs,outputs)
if(ret.status==1):
return ret.text.decode("UTF-8","ignore")
return ""
def utfprint(str):
try:
print(str)
except UnicodeEncodeError:
# Replace or omit the problematic character
utf_string = str.encode('ascii', 'ignore').decode('ascii')
utf_string = utf_string.replace('\a', '') #remove bell characters
print(utf_string)
def bring_terminal_to_foreground():
if os.name=='nt':
ctypes.windll.user32.ShowWindow(ctypes.windll.kernel32.GetConsoleWindow(), 9)
ctypes.windll.user32.SetForegroundWindow(ctypes.windll.kernel32.GetConsoleWindow())
#################################################################
### A hacky simple HTTP server simulating a kobold api by Concedo
### we are intentionally NOT using flask, because we want MINIMAL dependencies
#################################################################
friendlymodelname = "concedo/koboldcpp" # local kobold api apparently needs a hardcoded known HF model name
maxctx = 2048
maxhordectx = 2048
maxhordelen = 256
modelbusy = threading.Lock()
requestsinqueue = 0
defaultport = 5001
KcppVersion = "1.47.2"
showdebug = True
showsamplerwarning = True
showmaxctxwarning = True
session_kudos_earned = 0
session_jobs = 0
session_starttime = None
exitcounter = 0
punishcounter = 0 #causes a timeout if too many errors
rewardcounter = 0 #reduces error counts for successful jobs
totalgens = 0
currentusergenkey = "" #store a special key so polled streaming works even in multiuser
args = None #global args
class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
sys_version = ""
server_version = "ConcedoLlamaForKoboldServer"
def __init__(self, addr, port, embedded_kailite, embedded_kcpp_docs):
self.addr = addr
self.port = port
self.embedded_kailite = embedded_kailite
self.embedded_kcpp_docs = embedded_kcpp_docs
def __call__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
def log_message(self, format, *args):
global showdebug
if showdebug:
super().log_message(format, *args)
pass
async def generate_text(self, genparams, api_format, stream_flag):
global friendlymodelname
def run_blocking():
if api_format==1:
genparams["prompt"] = genparams.get('text', "")
genparams["top_k"] = int(genparams.get('top_k', 120))
genparams["max_length"] = genparams.get('max', 80)
elif api_format==3:
frqp = genparams.get('frequency_penalty', 0.1)
scaled_rep_pen = genparams.get('presence_penalty', frqp) + 1
genparams["max_length"] = genparams.get('max_tokens', 80)
genparams["rep_pen"] = scaled_rep_pen
# openai allows either a string or a list as a stop sequence
if isinstance(genparams.get('stop',[]), list):
genparams["stop_sequence"] = genparams.get('stop', [])
else:
genparams["stop_sequence"] = [genparams.get('stop')]
elif api_format==4:
# translate openai chat completion messages format into one big string.
messages_array = genparams.get('messages', [])
adapter_obj = genparams.get('adapter', {})
messages_string = ""
system_message_start = adapter_obj.get("system_start", "\n### Instruction:\n")
system_message_end = adapter_obj.get("system_end", "")
user_message_start = adapter_obj.get("user_start", "\n### Instruction:\n")
user_message_end = adapter_obj.get("user_end", "")
assistant_message_start = adapter_obj.get("assistant_start", "\n### Response:\n")
assistant_message_end = adapter_obj.get("assistant_end", "")
for message in messages_array:
if message['role'] == "system":
messages_string += system_message_start
elif message['role'] == "user":
messages_string += user_message_start
elif message['role'] == "assistant":
messages_string += assistant_message_start
messages_string += message['content']
if message['role'] == "system":
messages_string += system_message_end
elif message['role'] == "user":
messages_string += user_message_end
elif message['role'] == "assistant":
messages_string += assistant_message_end
messages_string += assistant_message_start
genparams["prompt"] = messages_string
frqp = genparams.get('frequency_penalty', 0.1)
scaled_rep_pen = genparams.get('presence_penalty', frqp) + 1
genparams["max_length"] = genparams.get('max_tokens', 80)
genparams["rep_pen"] = scaled_rep_pen
# openai allows either a string or a list as a stop sequence
if isinstance(genparams.get('stop',[]), list):
genparams["stop_sequence"] = genparams.get('stop', [])
else:
genparams["stop_sequence"] = [genparams.get('stop')]
return generate(
prompt=genparams.get('prompt', ""),
max_context_length=genparams.get('max_context_length', maxctx),
max_length=genparams.get('max_length', 80),
temperature=genparams.get('temperature', 0.8),
top_k=genparams.get('top_k', 120),
top_a=genparams.get('top_a', 0.0),
top_p=genparams.get('top_p', 0.85),
typical_p=genparams.get('typical', 1.0),
tfs=genparams.get('tfs', 1.0),
rep_pen=genparams.get('rep_pen', 1.1),
rep_pen_range=genparams.get('rep_pen_range', 256),
mirostat=genparams.get('mirostat', 0),
mirostat_tau=genparams.get('mirostat_tau', 5.0),
mirostat_eta=genparams.get('mirostat_eta', 0.1),
sampler_order=genparams.get('sampler_order', [6,0,1,3,4,2,5]),
seed=genparams.get('sampler_seed', -1),
stop_sequence=genparams.get('stop_sequence', []),
use_default_badwordsids=genparams.get('use_default_badwordsids', False),
stream_sse=stream_flag,
grammar=genparams.get('grammar', ''),
grammar_retain_state = genparams.get('grammar_retain_state', False),
genkey=genparams.get('genkey', ''))
recvtxt = ""
if stream_flag:
loop = asyncio.get_event_loop()
executor = ThreadPoolExecutor()
recvtxt = await loop.run_in_executor(executor, run_blocking)
else:
recvtxt = run_blocking()
if args.debugmode!=-1:
utfprint("\nOutput: " + recvtxt)
if api_format==1:
res = {"data": {"seqs":[recvtxt]}}
elif api_format==3:
res = {"id": "cmpl-1", "object": "text_completion", "created": 1, "model": friendlymodelname,
"choices": [{"text": recvtxt, "index": 0, "finish_reason": "length"}]}
elif api_format==4:
res = {"id": "chatcmpl-1", "object": "chat.completion", "created": 1, "model": friendlymodelname,
"choices": [{"index": 0, "message":{"role": "assistant", "content": recvtxt,}, "finish_reason": "length"}]}
else:
res = {"results": [{"text": recvtxt}]}
try:
return res
except Exception as e:
print(f"Generate: Error while generating: {e}")
async def send_oai_sse_event(self, data):
self.wfile.write(f'data: {data}\r\n\r\n'.encode())
self.wfile.flush()
async def send_kai_sse_event(self, data):
self.wfile.write(f'event: message\n'.encode())
self.wfile.write(f'data: {data}\n\n'.encode())
self.wfile.flush()
async def handle_sse_stream(self, api_format):
global friendlymodelname
self.send_response(200)
self.send_header("cache-control", "no-cache")
self.send_header("connection", "keep-alive")
self.end_headers(content_type='text/event-stream')
current_token = 0
incomplete_token_buffer = bytearray()
await asyncio.sleep(0.1) #anti race condition, prevent check from overtaking generate
while True:
streamDone = handle.has_finished() #exit next loop on done
tokenStr = ""
streamcount = handle.get_stream_count()
while current_token < streamcount:
token = handle.new_token(current_token)
if token is None: # Token isnt ready yet, received nullpointer
break
current_token += 1
newbyte = ctypes.string_at(token)
incomplete_token_buffer += bytearray(newbyte)
tokenSeg = incomplete_token_buffer.decode("UTF-8","ignore")
if tokenSeg!="":
incomplete_token_buffer.clear()
tokenStr += tokenSeg
if tokenStr!="":
if api_format == 4: # if oai chat, set format to expected openai streaming response
event_str = json.dumps({"id":"koboldcpp","object":"chat.completion.chunk","created":1,"model":friendlymodelname,"choices":[{"index":0,"finish_reason":"length","delta":{'role':'assistant','content':tokenStr}}]})
await self.send_oai_sse_event(event_str)
else:
event_str = json.dumps({"token": tokenStr})
await self.send_kai_sse_event(event_str)
tokenStr = ""
else:
await asyncio.sleep(0.02) #this should keep things responsive
if streamDone:
if api_format == 4: # if oai chat, send last [DONE] message consistent with openai format
await self.send_oai_sse_event('[DONE]')
break
# flush buffers, sleep a bit to make sure all data sent, and then force close the connection
self.wfile.flush()
await asyncio.sleep(0.2)
self.close_connection = True
await asyncio.sleep(0.1)
async def handle_request(self, genparams, api_format, stream_flag):
tasks = []
if stream_flag:
tasks.append(self.handle_sse_stream(api_format))
generate_task = asyncio.create_task(self.generate_text(genparams, api_format, stream_flag))
tasks.append(generate_task)
try:
await asyncio.gather(*tasks)
generate_result = generate_task.result()
return generate_result
except Exception as e:
print(e)
def do_GET(self):
global maxctx, maxhordelen, friendlymodelname, KcppVersion, totalgens
self.path = self.path.rstrip('/')
response_body = None
content_type = 'application/json'
if self.path in ["", "/?"] or self.path.startswith(('/?','?')): #it's possible for the root url to have ?params without /
content_type = 'text/html'
if self.embedded_kailite is None:
response_body = (f"Embedded Kobold Lite is not found.<br>You will have to connect via the main KoboldAI client, or <a href='https://lite.koboldai.net?local=1&port={self.port}'>use this URL</a> to connect.").encode()
else:
response_body = self.embedded_kailite
elif self.path.endswith(('/api/v1/model', '/api/latest/model')):
response_body = (json.dumps({'result': friendlymodelname }).encode())
elif self.path.endswith(('/api/v1/config/max_length', '/api/latest/config/max_length')):
response_body = (json.dumps({"value": maxhordelen}).encode())
elif self.path.endswith(('/api/v1/config/max_context_length', '/api/latest/config/max_context_length')):
response_body = (json.dumps({"value": min(maxctx,maxhordectx)}).encode())
elif self.path.endswith(('/api/v1/config/soft_prompt', '/api/latest/config/soft_prompt')):
response_body = (json.dumps({"value":""}).encode())
elif self.path.endswith(('/api/v1/config/soft_prompts_list', '/api/latest/config/soft_prompts_list')):
response_body = (json.dumps({"values": []}).encode())
elif self.path.endswith(('/api/v1/info/version', '/api/latest/info/version')):
response_body = (json.dumps({"result":"1.2.5"}).encode())
elif self.path.endswith(('/api/extra/true_max_context_length')): #do not advertise this to horde
response_body = (json.dumps({"value": maxctx}).encode())
elif self.path.endswith(('/api/extra/version')):
response_body = (json.dumps({"result":"KoboldCpp","version":KcppVersion}).encode())
elif self.path.endswith(('/api/extra/perf')):
lastp = handle.get_last_process_time()
laste = handle.get_last_eval_time()
lastc = handle.get_last_token_count()
stopreason = handle.get_last_stop_reason()
response_body = (json.dumps({"last_process":lastp,"last_eval":laste,"last_token_count":lastc, "stop_reason":stopreason, "queue":requestsinqueue, "idle":(0 if modelbusy.locked() else 1)}).encode())
elif self.path.endswith('/api/extra/generate/check'):
pendtxtStr = ""
if requestsinqueue==0 and totalgens>0:
pendtxt = handle.get_pending_output()
pendtxtStr = ctypes.string_at(pendtxt).decode("UTF-8","ignore")
response_body = (json.dumps({"results": [{"text": pendtxtStr}]}).encode())
elif self.path.endswith('/v1/models'):
response_body = (json.dumps({"object":"list","data":[{"id":friendlymodelname,"object":"model","created":1,"owned_by":"koboldcpp","permission":[],"root":"koboldcpp"}]}).encode())
elif self.path=="/api":
content_type = 'text/html'
if self.embedded_kcpp_docs is None:
response_body = (f"KoboldCpp partial API reference can be found at the wiki: https://github.com/LostRuins/koboldcpp/wiki").encode()
else:
response_body = self.embedded_kcpp_docs
elif self.path.endswith(('/api')) or self.path.endswith(('/api/v1')):
self.path = "/api"
self.send_response(302)
self.send_header("location", self.path)
self.end_headers(content_type='text/html')
return None
if response_body is None:
self.send_response(404)
self.end_headers(content_type='text/html')
rp = 'Error: HTTP Server is running, but this endpoint does not exist. Please check the URL.'
self.wfile.write(rp.encode())
else:
self.send_response(200)
self.send_header('content-length', str(len(response_body)))
self.end_headers(content_type=content_type)
self.wfile.write(response_body)
return
def do_POST(self):
global modelbusy, requestsinqueue, currentusergenkey, totalgens
content_length = int(self.headers['content-length'])
body = self.rfile.read(content_length)
self.path = self.path.rstrip('/')
if self.path.endswith(('/api/extra/tokencount')):
try:
genparams = json.loads(body)
countprompt = genparams.get('prompt', "")
count = handle.token_count(countprompt.encode("UTF-8"))
self.send_response(200)
self.end_headers(content_type='application/json')
self.wfile.write(json.dumps({"value": count}).encode())
except Exception as e:
utfprint("Count Tokens - Body Error: " + str(e))
self.send_response(400)
self.end_headers(content_type='application/json')
self.wfile.write(json.dumps({"value": -1}).encode())
return
if self.path.endswith('/api/extra/abort'):
multiuserkey = ""
try:
tempbody = json.loads(body)
if isinstance(tempbody, dict):
multiuserkey = tempbody.get('genkey', "")
except Exception as e:
multiuserkey = ""
pass
if (multiuserkey=="" and requestsinqueue==0) or (multiuserkey!="" and multiuserkey==currentusergenkey):
ag = handle.abort_generate()
time.sleep(0.1) #short delay before replying
self.send_response(200)
self.end_headers(content_type='application/json')
self.wfile.write(json.dumps({"success": ("true" if ag else "false")}).encode())
print("\nGeneration Aborted")
else:
self.send_response(200)
self.end_headers(content_type='application/json')
self.wfile.write(json.dumps({"success": "false"}).encode())
return
if self.path.endswith('/api/extra/generate/check'):
pendtxtStr = ""
multiuserkey = ""
try:
tempbody = json.loads(body)
if isinstance(tempbody, dict):
multiuserkey = tempbody.get('genkey', "")
except Exception as e:
multiuserkey = ""
pass
if totalgens>0:
if (multiuserkey=="" and requestsinqueue==0) or (multiuserkey!="" and multiuserkey==currentusergenkey):
pendtxt = handle.get_pending_output()
pendtxtStr = ctypes.string_at(pendtxt).decode("UTF-8","ignore")
self.send_response(200)
self.end_headers(content_type='application/json')
self.wfile.write(json.dumps({"results": [{"text": pendtxtStr}]}).encode())
return
reqblocking = False
if args.multiuser and requestsinqueue < 4: #up to 5 concurrent requests
reqblocking = True
requestsinqueue += 1
if not modelbusy.acquire(blocking=reqblocking):
self.send_response(503)
self.end_headers(content_type='application/json')
self.wfile.write(json.dumps({"detail": {
"msg": "Server is busy; please try again later.",
"type": "service_unavailable",
}}).encode())
return
if reqblocking:
requestsinqueue = (requestsinqueue - 1) if requestsinqueue > 0 else 0
try:
sse_stream_flag = False
api_format = 0 #1=basic,2=kai,3=oai,4=oai-chat
if self.path.endswith('/request'):
api_format = 1
if self.path.endswith(('/api/v1/generate', '/api/latest/generate')):
api_format = 2
if self.path.endswith('/api/extra/generate/stream'):
api_format = 2
sse_stream_flag = True
if self.path.endswith('/v1/completions'):
api_format = 3
if self.path.endswith('/v1/chat/completions'):
api_format = 4
if api_format > 0:
genparams = None
try:
genparams = json.loads(body)
except Exception as e:
utfprint("Body Err: " + str(body))
return self.send_response(503)
if args.debugmode!=-1:
utfprint("\nInput: " + json.dumps(genparams))
if args.foreground:
bring_terminal_to_foreground()
# Check if streaming chat completions, if so, set stream mode to true
if api_format == 4 and "stream" in genparams and genparams["stream"]:
sse_stream_flag = True
gen = asyncio.run(self.handle_request(genparams, api_format, sse_stream_flag))
try:
# Headers are already sent when streaming
if not sse_stream_flag:
self.send_response(200)
self.end_headers(content_type='application/json')
self.wfile.write(json.dumps(gen).encode())
except:
print("Generate: The response could not be sent, maybe connection was terminated?")
return
finally:
modelbusy.release()
self.send_response(404)
self.end_headers(content_type='text/html')
def do_OPTIONS(self):
self.send_response(200)
self.end_headers(content_type='text/html')
def do_HEAD(self):
self.send_response(200)
self.end_headers(content_type='text/html')
def end_headers(self, content_type=None):
self.send_header('access-control-allow-origin', '*')
self.send_header('access-control-allow-methods', '*')
self.send_header('access-control-allow-headers', '*, Accept, Content-Type, Content-Length, Accept-Encoding, X-CSRF-Token, Client-Agent, X-Fields, Content-Type, Authorization, X-Requested-With, X-HTTP-Method-Override, apikey, genkey')
if content_type is not None:
self.send_header('content-type', content_type)
return super(ServerRequestHandler, self).end_headers()
def RunServerMultiThreaded(addr, port, embedded_kailite = None, embedded_kcpp_docs = None):
global exitcounter
sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
sock.bind((addr, port))
sock.listen(5)
class Thread(threading.Thread):
def __init__(self, i):
threading.Thread.__init__(self)
self.i = i
self.daemon = True
self.start()
def run(self):
global exitcounter
handler = ServerRequestHandler(addr, port, embedded_kailite, embedded_kcpp_docs)
with http.server.HTTPServer((addr, port), handler, False) as self.httpd:
try:
self.httpd.socket = sock
self.httpd.server_bind = self.server_close = lambda self: None
self.httpd.serve_forever()
except (KeyboardInterrupt,SystemExit):
exitcounter = 999
self.httpd.server_close()
sys.exit(0)
finally:
exitcounter = 999
self.httpd.server_close()
sys.exit(0)
def stop(self):
global exitcounter
exitcounter = 999
self.httpd.server_close()
numThreads = 12
threadArr = []
for i in range(numThreads):
threadArr.append(Thread(i))
while 1:
try:
time.sleep(10)
except KeyboardInterrupt:
exitcounter = 999
for i in range(numThreads):
threadArr[i].stop()
sys.exit(0)
# note: customtkinter-5.2.0
def show_new_gui():
from tkinter.filedialog import askopenfilename
from tkinter.filedialog import asksaveasfile
# if args received, launch
if len(sys.argv) != 1:
import tkinter as tk
root = tk.Tk() #we dont want the useless window to be visible, but we want it in taskbar
root.attributes("-alpha", 0)
args.model_param = askopenfilename(title="Select ggml model .bin or .gguf file or .kcpps config")
root.destroy()
if args.model_param and args.model_param!="" and args.model_param.lower().endswith('.kcpps'):
loadconfigfile(args.model_param)
if not args.model_param:
print("\nNo ggml model or kcpps file was selected. Exiting.")
time.sleep(3)
sys.exit(2)
return
import customtkinter as ctk
nextstate = 0 #0=exit, 1=launch
windowwidth = 530
windowheight = 500
ctk.set_appearance_mode("dark")
root = ctk.CTk()
root.geometry(str(windowwidth) + "x" + str(windowheight))
root.title("KoboldCpp v"+KcppVersion)
root.resizable(False,False)
tabs = ctk.CTkFrame(root, corner_radius = 0, width=windowwidth, height=windowheight-50)
tabs.grid(row=0, stick="nsew")
tabnames= ["Quick Launch", "Hardware", "Tokens", "Model", "Network"]
navbuttons = {}
navbuttonframe = ctk.CTkFrame(tabs, width=100, height=int(tabs.cget("height")))
navbuttonframe.grid(row=0, column=0, padx=2,pady=2)
navbuttonframe.grid_propagate(False)
tabcontentframe = ctk.CTkFrame(tabs, width=windowwidth - int(navbuttonframe.cget("width")), height=int(tabs.cget("height")))
tabcontentframe.grid(row=0, column=1, sticky="nsew", padx=2, pady=2)
tabcontentframe.grid_propagate(False)
tabcontent = {}
lib_option_pairs = [
(lib_openblas, "Use OpenBLAS"),
(lib_clblast, "Use CLBlast"),
(lib_cublas, "Use CuBLAS"),
(lib_hipblas, "Use hipBLAS (ROCm)"),
(lib_default, "Use No BLAS"),
(lib_noavx2, "NoAVX2 Mode (Old CPU)"),
(lib_failsafe, "Failsafe Mode (Old CPU)")]
openblas_option, clblast_option, cublas_option, hipblas_option, default_option, noavx2_option, failsafe_option = (opt if file_exists(lib) or (os.name == 'nt' and file_exists(opt + ".dll")) else None for lib, opt in lib_option_pairs)
# slider data
blasbatchsize_values = ["-1", "32", "64", "128", "256", "512", "1024", "2048"]
blasbatchsize_text = ["Don't Batch BLAS","32","64","128","256","512","1024","2048"]
contextsize_text = ["512", "1024", "2048", "3072", "4096", "6144", "8192", "12288", "16384", "24576", "32768", "65536"]
runopts = [opt for lib, opt in lib_option_pairs if file_exists(lib)]
antirunopts = [opt.replace("Use ", "") for lib, opt in lib_option_pairs if not (opt in runopts)]
if not any(runopts):
show_gui_msgbox("No Backends Available!","KoboldCPP couldn't locate any backends to use (i.e Default, OpenBLAS, CLBlast, CuBLAS).\n\nTo use the program, please run the 'make' command from the directory.")
time.sleep(3)
sys.exit(2)
def tabbuttonaction(name):
for t in tabcontent:
if name == t:
tabcontent[t].grid(row=0, column=0)
navbuttons[t].configure(fg_color="#6f727b")
else:
tabcontent[t].grid_forget()
navbuttons[t].configure(fg_color="transparent")
# Dynamically create tabs + buttons based on values of [tabnames]
for idx, name in enumerate(tabnames):
tabcontent[name] = ctk.CTkFrame(tabcontentframe, width=int(tabcontentframe.cget("width")), height=int(tabcontentframe.cget("height")), fg_color="transparent")
tabcontent[name].grid_propagate(False)
if idx == 0:
tabcontent[name].grid(row=idx, sticky="nsew")
ctk.CTkLabel(tabcontent[name], text= name, font=ctk.CTkFont(None, 14, 'bold')).grid(row=0, padx=12, pady = 5, stick='nw')
navbuttons[name] = ctk.CTkButton(navbuttonframe, text=name, width = 100, corner_radius=0 , command = lambda d=name:tabbuttonaction(d), hover_color="#868a94" )
navbuttons[name].grid(row=idx)
tabbuttonaction(tabnames[0])
# helper functions
def makecheckbox(parent, text, variable=None, row=0, column=0, command=None, onvalue=1, offvalue=0):
temp = ctk.CTkCheckBox(parent, text=text,variable=variable, onvalue=onvalue, offvalue=offvalue)
if command is not None and variable is not None:
variable.trace("w", command)
temp.grid(row=row,column=column, padx=8, pady=1, stick="nw")
return temp
def makelabel(parent, text, row, column=0):
temp = ctk.CTkLabel(parent, text=text)
temp.grid(row=row, column=column, padx=8, pady=1, stick="nw")
return temp
def makeslider(parent, label, options, var, from_ , to, row=0, width=160, height=10, set=0):
sliderLabel = makelabel(parent, options[set], row + 1, 1)
makelabel(parent, label, row)
def sliderUpdate(a,b,c):
sliderLabel.configure(text = options[int(var.get())])
var.trace("w", sliderUpdate)
slider = ctk.CTkSlider(parent, from_=from_, to=to, variable = var, width = width, height=height, border_width=5,number_of_steps=len(options) - 1)
slider.grid(row=row+1, column=0, padx = 8, stick="w")
slider.set(set)
return slider
def makelabelentry(parent, text, var, row=0, width= 50):
label = makelabel(parent, text, row)
entry = ctk.CTkEntry(parent, width=width, textvariable=var) #you cannot set placeholder text for SHARED variables
entry.grid(row=row, column=1, padx= 8, stick="nw")
return entry, label
def makefileentry(parent, text, searchtext, var, row=0, width=250, filetypes=[]):
makelabel(parent, text, row)
def getfilename(var, text):
var.set(askopenfilename(title=text,filetypes=filetypes))
entry = ctk.CTkEntry(parent, width, textvariable=var)
entry.grid(row=row+1, column=0, padx=8, stick="nw")
button = ctk.CTkButton(parent, 50, text="Browse", command= lambda a=var,b=searchtext:getfilename(a,b))
button.grid(row=row+1, column=1, stick="nw")
return
def show_tooltip(event, tooltip_text=None):
if hasattr(show_tooltip, "_tooltip"):
tooltip = show_tooltip._tooltip
else: