Repository navigation
Expand file tree
/
Copy pathllama_cpp.py
More file actions
3348 lines (2891 loc) · 116 KB
/
Copy pathllama_cpp.py
File metadata and controls
3348 lines (2891 loc) · 116 KB
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
from __future__ import annotations
import sys
import os
import ctypes
import functools
import pathlib
from typing import (
Any,
Callable,
List,
Union,
NewType,
Optional,
TYPE_CHECKING,
TypeVar,
Generic,
)
from typing_extensions import TypeAlias
# Load the library
def _load_shared_library(lib_base_name: str):
# Construct the paths to the possible shared library names
_base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__)))
# Searching for the library in the current directory under the name "libllama" (default name
# for llamacpp) and "llama" (default name for this repo)
_lib_paths: List[pathlib.Path] = []
# Determine the file extension based on the platform
if sys.platform.startswith("linux"):
_lib_paths += [
_base_path / f"lib{lib_base_name}.so",
]
elif sys.platform == "darwin":
_lib_paths += [
_base_path / f"lib{lib_base_name}.so",
_base_path / f"lib{lib_base_name}.dylib",
]
elif sys.platform == "win32":
_lib_paths += [
_base_path / f"{lib_base_name}.dll",
_base_path / f"lib{lib_base_name}.dll",
]
else:
raise RuntimeError("Unsupported platform")
if "LLAMA_CPP_LIB" in os.environ:
lib_base_name = os.environ["LLAMA_CPP_LIB"]
_lib = pathlib.Path(lib_base_name)
_base_path = _lib.parent.resolve()
_lib_paths = [_lib.resolve()]
cdll_args = dict() # type: ignore
# Add the library directory to the DLL search path on Windows (if needed)
if sys.platform == "win32" and sys.version_info >= (3, 8):
os.add_dll_directory(str(_base_path))
if "CUDA_PATH" in os.environ:
os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin"))
os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib"))
if "HIP_PATH" in os.environ:
os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "bin"))
os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "lib"))
cdll_args["winmode"] = ctypes.RTLD_GLOBAL
# Try to load the shared library, handling potential errors
for _lib_path in _lib_paths:
if _lib_path.exists():
try:
return ctypes.CDLL(str(_lib_path), **cdll_args) # type: ignore
except Exception as e:
raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}")
raise FileNotFoundError(
f"Shared library with base name '{lib_base_name}' not found"
)
# Specify the base name of the shared library to load
_lib_base_name = "llama"
# Load the library
_lib = _load_shared_library(_lib_base_name)
# ctypes sane type hint helpers
#
# - Generic Pointer and Array types
# - PointerOrRef type with a type hinted byref function
#
# NOTE: Only use these for static type checking not for runtime checks
# no good will come of that
if TYPE_CHECKING:
CtypesCData = TypeVar("CtypesCData", bound=ctypes._CData) # type: ignore
CtypesArray: TypeAlias = ctypes.Array[CtypesCData] # type: ignore
CtypesPointer: TypeAlias = ctypes._Pointer[CtypesCData] # type: ignore
CtypesVoidPointer: TypeAlias = ctypes.c_void_p
class CtypesRef(Generic[CtypesCData]):
pass
CtypesPointerOrRef: TypeAlias = Union[
CtypesPointer[CtypesCData], CtypesRef[CtypesCData]
]
CtypesFuncPointer: TypeAlias = ctypes._FuncPointer # type: ignore
F = TypeVar("F", bound=Callable[..., Any])
def ctypes_function_for_shared_library(lib: ctypes.CDLL):
def ctypes_function(
name: str, argtypes: List[Any], restype: Any, enabled: bool = True
):
def decorator(f: F) -> F:
if enabled:
func = getattr(lib, name)
func.argtypes = argtypes
func.restype = restype
functools.wraps(f)(func)
return func
else:
return f
return decorator
return ctypes_function
ctypes_function = ctypes_function_for_shared_library(_lib)
def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCData]:
"""Type-annotated version of ctypes.byref"""
...
byref = ctypes.byref # type: ignore
# from ggml.h
# // NOTE: always add types at the end of the enum to keep backward compatibility
# enum ggml_type {
# GGML_TYPE_F32 = 0,
# GGML_TYPE_F16 = 1,
# GGML_TYPE_Q4_0 = 2,
# GGML_TYPE_Q4_1 = 3,
# // GGML_TYPE_Q4_2 = 4, support has been removed
# // GGML_TYPE_Q4_3 = 5, support has been removed
# GGML_TYPE_Q5_0 = 6,
# GGML_TYPE_Q5_1 = 7,
# GGML_TYPE_Q8_0 = 8,
# GGML_TYPE_Q8_1 = 9,
# GGML_TYPE_Q2_K = 10,
# GGML_TYPE_Q3_K = 11,
# GGML_TYPE_Q4_K = 12,
# GGML_TYPE_Q5_K = 13,
# GGML_TYPE_Q6_K = 14,
# GGML_TYPE_Q8_K = 15,
# GGML_TYPE_IQ2_XXS = 16,
# GGML_TYPE_IQ2_XS = 17,
# GGML_TYPE_IQ3_XXS = 18,
# GGML_TYPE_IQ1_S = 19,
# GGML_TYPE_IQ4_NL = 20,
# GGML_TYPE_IQ3_S = 21,
# GGML_TYPE_IQ2_S = 22,
# GGML_TYPE_IQ4_XS = 23,
# GGML_TYPE_I8 = 24,
# GGML_TYPE_I16 = 25,
# GGML_TYPE_I32 = 26,
# GGML_TYPE_I64 = 27,
# GGML_TYPE_F64 = 28,
# GGML_TYPE_IQ1_M = 29,
# GGML_TYPE_COUNT,
# };
GGML_TYPE_F32 = 0
GGML_TYPE_F16 = 1
GGML_TYPE_Q4_0 = 2
GGML_TYPE_Q4_1 = 3
GGML_TYPE_Q5_0 = 6
GGML_TYPE_Q5_1 = 7
GGML_TYPE_Q8_0 = 8
GGML_TYPE_Q8_1 = 9
GGML_TYPE_Q2_K = 10
GGML_TYPE_Q3_K = 11
GGML_TYPE_Q4_K = 12
GGML_TYPE_Q5_K = 13
GGML_TYPE_Q6_K = 14
GGML_TYPE_Q8_K = 15
GGML_TYPE_IQ2_XXS = 16
GGML_TYPE_IQ2_XS = 17
GGML_TYPE_IQ3_XXS = 18
GGML_TYPE_IQ1_S = 19
GGML_TYPE_IQ4_NL = 20
GGML_TYPE_IQ3_S = 21
GGML_TYPE_IQ2_S = 22
GGML_TYPE_IQ4_XS = 23
GGML_TYPE_I8 = 24
GGML_TYPE_I16 = 25
GGML_TYPE_I32 = 26
GGML_TYPE_I64 = 27
GGML_TYPE_F64 = 28
GGML_TYPE_IQ1_M = 29
GGML_TYPE_COUNT = 30
# from ggml-backend.h
# typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE(
ctypes.c_bool, ctypes.c_void_p, ctypes.c_bool, ctypes.c_void_p
)
# // Abort callback
# // If not NULL, called before ggml computation
# // If it returns true, the computation is aborted
# typedef bool (*ggml_abort_callback)(void * data);
ggml_abort_callback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_void_p)
# llama.h bindings
_lib.llama_max_devices.argtypes = []
_lib.llama_max_devices.restype = ctypes.c_size_t
LLAMA_MAX_DEVICES = _lib.llama_max_devices()
# define LLAMA_DEFAULT_SEED 0xFFFFFFFF
LLAMA_DEFAULT_SEED = 0xFFFFFFFF
# define LLAMA_MAX_RNG_STATE (64*1024)
LLAMA_MAX_RNG_STATE = 64 * 1024
# define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
LLAMA_FILE_MAGIC_GGLA = 0x67676C61
# define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
LLAMA_FILE_MAGIC_GGSN = 0x6767736E
# define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'
LLAMA_FILE_MAGIC_GGSQ = 0x67677371
# define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN
LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN
# define LLAMA_SESSION_VERSION 5
LLAMA_SESSION_VERSION = 5
# define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ
LLAMA_STATE_SEQ_MAGIC = LLAMA_FILE_MAGIC_GGSQ
# define LLAMA_STATE_SEQ_VERSION 1
LLAMA_STATE_SEQ_VERSION = 1
# struct llama_model;
llama_model_p = NewType("llama_model_p", int)
llama_model_p_ctypes = ctypes.c_void_p
# struct llama_context;
llama_context_p = NewType("llama_context_p", int)
llama_context_p_ctypes = ctypes.c_void_p
# typedef int32_t llama_pos;
llama_pos = ctypes.c_int32
# typedef int32_t llama_token;
llama_token = ctypes.c_int32
llama_token_p = ctypes.POINTER(llama_token)
# typedef int32_t llama_seq_id;
llama_seq_id = ctypes.c_int32
# enum llama_vocab_type {
# LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
# LLAMA_VOCAB_TYPE_SPM = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
# LLAMA_VOCAB_TYPE_BPE = 2, // GPT-2 tokenizer based on byte-level BPE
# LLAMA_VOCAB_TYPE_WPM = 3, // BERT tokenizer based on WordPiece
# };
LLAMA_VOCAB_TYPE_NONE = 0
"""For models without vocab"""
LLAMA_VOCAB_TYPE_SPM = 1
"""LLaMA tokenizer based on byte-level BPE with byte fallback"""
LLAMA_VOCAB_TYPE_BPE = 2
"""GPT-2 tokenizer based on byte-level BPE"""
LLAMA_VOCAB_TYPE_WPM = 3
"""BERT tokenizer based on WordPiece"""
# // note: these values should be synchronized with ggml_rope
# // TODO: maybe move this enum to ggml.h (ggml_rope_type)
# enum llama_rope_type {
# LLAMA_ROPE_TYPE_NONE = -1,
# LLAMA_ROPE_TYPE_NORM = 0,
# LLAMA_ROPE_TYPE_NEOX = 2,
# LLAMA_ROPE_TYPE_GLM = 4,
# };
LLAMA_ROPE_TYPE_NONE = -1
LLAMA_ROPE_TYPE_NORM = 0
LLAMA_ROPE_TYPE_NEOX = 2
LLAMA_ROPE_TYPE_GLM = 4
# enum llama_token_type {
# LLAMA_TOKEN_TYPE_UNDEFINED = 0,
# LLAMA_TOKEN_TYPE_NORMAL = 1,
# LLAMA_TOKEN_TYPE_UNKNOWN = 2,
# LLAMA_TOKEN_TYPE_CONTROL = 3,
# LLAMA_TOKEN_TYPE_USER_DEFINED = 4,
# LLAMA_TOKEN_TYPE_UNUSED = 5,
# LLAMA_TOKEN_TYPE_BYTE = 6,
# };
LLAMA_TOKEN_TYPE_UNDEFINED = 0
LLAMA_TOKEN_TYPE_NORMAL = 1
LLAMA_TOKEN_TYPE_UNKNOWN = 2
LLAMA_TOKEN_TYPE_CONTROL = 3
LLAMA_TOKEN_TYPE_USER_DEFINED = 4
LLAMA_TOKEN_TYPE_UNUSED = 5
LLAMA_TOKEN_TYPE_BYTE = 6
# // model file types
# enum llama_ftype {
# LLAMA_FTYPE_ALL_F32 = 0,
# LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
# // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed
# // LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed
# LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q3_K_S = 11, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q3_K_M = 12, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q3_K_L = 13, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_K_S = 14, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_K_M = 15, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_K_S = 16, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_K_M = 17, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q6_K = 18, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ3_XS = 22, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ1_S = 24, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ4_NL = 25, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ3_S = 26, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ3_M = 27, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ2_S = 28, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ2_M = 29, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors
# LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
# };
LLAMA_FTYPE_ALL_F32 = 0
LLAMA_FTYPE_MOSTLY_F16 = 1
LLAMA_FTYPE_MOSTLY_Q4_0 = 2
LLAMA_FTYPE_MOSTLY_Q4_1 = 3
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4
LLAMA_FTYPE_MOSTLY_Q8_0 = 7
LLAMA_FTYPE_MOSTLY_Q5_0 = 8
LLAMA_FTYPE_MOSTLY_Q5_1 = 9
LLAMA_FTYPE_MOSTLY_Q2_K = 10
LLAMA_FTYPE_MOSTLY_Q3_K_S = 11
LLAMA_FTYPE_MOSTLY_Q3_K_M = 12
LLAMA_FTYPE_MOSTLY_Q3_K_L = 13
LLAMA_FTYPE_MOSTLY_Q4_K_S = 14
LLAMA_FTYPE_MOSTLY_Q4_K_M = 15
LLAMA_FTYPE_MOSTLY_Q5_K_S = 16
LLAMA_FTYPE_MOSTLY_Q5_K_M = 17
LLAMA_FTYPE_MOSTLY_Q6_K = 18
LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19
LLAMA_FTYPE_MOSTLY_IQ2_XS = 20
LLAMA_FTYPE_MOSTLY_Q2_K_S = 21
LLAMA_FTYPE_MOSTLY_IQ3_XS = 22
LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23
LLAMA_FTYPE_MOSTLY_IQ1_S = 24
LLAMA_FTYPE_MOSTLY_IQ4_NL = 25
LLAMA_FTYPE_MOSTLY_IQ3_S = 26
LLAMA_FTYPE_MOSTLY_IQ3_M = 27
LLAMA_FTYPE_MOSTLY_IQ2_S = 28
LLAMA_FTYPE_MOSTLY_IQ2_M = 29
LLAMA_FTYPE_MOSTLY_IQ4_XS = 30
LLAMA_FTYPE_GUESSED = 1024
# enum llama_rope_scaling_type {
# LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1,
# LLAMA_ROPE_SCALING_TYPE_NONE = 0,
# LLAMA_ROPE_SCALING_TYPE_LINEAR = 1,
# LLAMA_ROPE_SCALING_TYPE_YARN = 2,
# LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN,
# };
LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1
LLAMA_ROPE_SCALING_TYPE_NONE = 0
LLAMA_ROPE_SCALING_TYPE_LINEAR = 1
LLAMA_ROPE_SCALING_TYPE_YARN = 2
LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN
# enum llama_pooling_type {
# LLAMA_POOLING_TYPE_UNSPECIFIED = -1,
# LLAMA_POOLING_TYPE_NONE = 0,
# LLAMA_POOLING_TYPE_MEAN = 1,
# LLAMA_POOLING_TYPE_CLS = 2,
# };
LLAMA_POOLING_TYPE_UNSPECIFIED = -1
LLAMA_POOLING_TYPE_NONE = 0
LLAMA_POOLING_TYPE_MEAN = 1
LLAMA_POOLING_TYPE_CLS = 2
# enum llama_split_mode {
# LLAMA_SPLIT_MODE_NONE = 0, // single GPU
# LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
# LLAMA_SPLIT_MODE_ROW = 2, // split rows across GPUs
# };
LLAMA_SPLIT_MODE_NONE = 0
LLAMA_SPLIT_MODE_LAYER = 1
LLAMA_SPLIT_MODE_ROW = 2
# typedef struct llama_token_data {
# llama_token id; // token id
# float logit; // log-odds of the token
# float p; // probability of the token
# } llama_token_data;
class llama_token_data(ctypes.Structure):
"""Used to store token data
Attributes:
id (llama_token): token id
logit (float): log-odds of the token
p (float): probability of the token"""
if TYPE_CHECKING:
id: llama_token
logit: float
p: float
_fields_ = [
("id", llama_token),
("logit", ctypes.c_float),
("p", ctypes.c_float),
]
llama_token_data_p = ctypes.POINTER(llama_token_data)
# typedef struct llama_token_data_array {
# llama_token_data * data;
# size_t size;
# bool sorted;
# } llama_token_data_array;
class llama_token_data_array(ctypes.Structure):
"""Used to sample tokens given logits
Attributes:
data (ctypes.Array[llama_token_data]): token data
size (int): size of the array
sorted (bool): whether the array is sorted"""
if TYPE_CHECKING:
data: CtypesArray[llama_token_data]
size: int
sorted: bool
_fields_ = [
("data", llama_token_data_p),
("size", ctypes.c_size_t),
("sorted", ctypes.c_bool),
]
llama_token_data_array_p = ctypes.POINTER(llama_token_data_array)
# typedef bool (*llama_progress_callback)(float progress, void *ctx);
llama_progress_callback = ctypes.CFUNCTYPE(
ctypes.c_bool, ctypes.c_float, ctypes.c_void_p
)
# // Input data for llama_decode
# // A llama_batch object can contain input about one or many sequences
# // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
# //
# // - token : the token ids of the input (used when embd is NULL)
# // - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
# // - pos : the positions of the respective token in the sequence
# // - seq_id : the sequence to which the respective token belongs
# // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
# //
# typedef struct llama_batch {
# int32_t n_tokens;
# llama_token * token;
# float * embd;
# llama_pos * pos;
# int32_t * n_seq_id;
# llama_seq_id ** seq_id;
# int8_t * logits; // TODO: rename this to "output"
# // NOTE: helpers for smooth API transition - can be deprecated in the future
# // for future-proof code, use the above fields instead and ignore everything below
# //
# // pos[i] = all_pos_0 + i*all_pos_1
# //
# llama_pos all_pos_0; // used if pos == NULL
# llama_pos all_pos_1; // used if pos == NULL
# llama_seq_id all_seq_id; // used if seq_id == NULL
# } llama_batch;
class llama_batch(ctypes.Structure):
"""Input data for llama_decode
A llama_batch object can contain input about one or many sequences
The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
Attributes:
n_tokens (int): number of tokens
token (ctypes.Array[llama_token]): the token ids of the input (used when embd is NULL)
embd (ctypes.Array[ctypes.ctypes.c_float]): token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
pos (ctypes.Array[ctypes.Array[llama_pos]]): the positions of the respective token in the sequence
seq_id (ctypes.Array[ctypes.Array[llama_seq_id]]): the sequence to which the respective token belongs
logits (ctypes.Array[ctypes.ctypes.c_int8]): if zero, the logits for the respective token will not be output
"""
if TYPE_CHECKING:
n_tokens: int
token: CtypesArray[llama_token]
embd: CtypesArray[ctypes.c_float]
pos: CtypesArray[CtypesArray[llama_pos]]
n_seq_id: CtypesArray[ctypes.c_int]
seq_id: CtypesArray[CtypesArray[llama_seq_id]]
logits: CtypesArray[ctypes.c_int8]
_fields_ = [
("n_tokens", ctypes.c_int32),
("token", ctypes.POINTER(llama_token)),
("embd", ctypes.POINTER(ctypes.c_float)),
("pos", ctypes.POINTER(llama_pos)),
("n_seq_id", ctypes.POINTER(ctypes.c_int32)),
("seq_id", ctypes.POINTER(ctypes.POINTER(llama_seq_id))),
("logits", ctypes.POINTER(ctypes.c_int8)),
("all_pos_0", llama_pos),
("all_pos_1", llama_pos),
("all_seq_id", llama_seq_id),
]
# enum llama_model_kv_override_type {
# LLAMA_KV_OVERRIDE_TYPE_INT,
# LLAMA_KV_OVERRIDE_TYPE_FLOAT,
# LLAMA_KV_OVERRIDE_TYPE_BOOL,
# };
LLAMA_KV_OVERRIDE_TYPE_INT = 0
LLAMA_KV_OVERRIDE_TYPE_FLOAT = 1
LLAMA_KV_OVERRIDE_TYPE_BOOL = 2
# struct llama_model_kv_override {
# char key[128];
# enum llama_model_kv_override_type tag;
# union {
# int64_t int_value;
# double float_value;
# bool bool_value;
# };
# };
class llama_model_kv_override_value(ctypes.Union):
_fields_ = [
("int_value", ctypes.c_int64),
("float_value", ctypes.c_double),
("bool_value", ctypes.c_bool),
]
class llama_model_kv_override(ctypes.Structure):
_fields_ = [
("key", ctypes.c_char * 128),
("tag", ctypes.c_int),
("value", llama_model_kv_override_value),
]
# struct llama_model_params {
# int32_t n_gpu_layers; // number of layers to store in VRAM
# enum llama_split_mode split_mode; // how to split the model across multiple GPUs
# // main_gpu interpretation depends on split_mode:
# // LLAMA_SPLIT_NONE: the GPU that is used for the entire model
# // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
# // LLAMA_SPLIT_LAYER: ignored
# int32_t main_gpu;
# // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
# const float * tensor_split;
# // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
# // If the provided progress_callback returns true, model loading continues.
# // If it returns false, model loading is immediately aborted.
# llama_progress_callback progress_callback;
# // context pointer passed to the progress callback
# void * progress_callback_user_data;
# // override key-value pairs of the model meta data
# const struct llama_model_kv_override * kv_overrides;
# // Keep the booleans together to avoid misalignment during copy-by-value.
# bool vocab_only; // only load the vocabulary, no weights
# bool use_mmap; // use mmap if possible
# bool use_mlock; // force system to keep model in RAM
# };
class llama_model_params(ctypes.Structure):
"""Parameters for llama_model
Attributes:
n_gpu_layers (int): number of layers to store in VRAM
split_mode (int): how to split the model across multiple GPUs
main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored
tensor_split (ctypes.Array[ctypes.ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted.
progress_callback_user_data (ctypes.ctypes.c_void_p): context pointer passed to the progress callback
kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data
vocab_only (bool): only load the vocabulary, no weights
use_mmap (bool): use mmap if possible
use_mlock (bool): force system to keep model in RAM"""
if TYPE_CHECKING:
n_gpu_layers: int
split_mode: int
main_gpu: int
tensor_split: CtypesArray[ctypes.c_float]
progress_callback: Callable[[float, ctypes.c_void_p], bool]
progress_callback_user_data: ctypes.c_void_p
kv_overrides: CtypesArray[llama_model_kv_override]
vocab_only: bool
use_mmap: bool
use_mlock: bool
_fields_ = [
("n_gpu_layers", ctypes.c_int32),
("split_mode", ctypes.c_int),
("main_gpu", ctypes.c_int32),
("tensor_split", ctypes.POINTER(ctypes.c_float)),
("progress_callback", llama_progress_callback),
("progress_callback_user_data", ctypes.c_void_p),
("kv_overrides", ctypes.POINTER(llama_model_kv_override)),
("vocab_only", ctypes.c_bool),
("use_mmap", ctypes.c_bool),
("use_mlock", ctypes.c_bool),
]
# struct llama_context_params {
# uint32_t seed; // RNG seed, -1 for random
# uint32_t n_ctx; // text context, 0 = from model
# uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode
# uint32_t n_ubatch; // physical maximum batch size
# uint32_t n_seq_max; // max number of sequences (i.e. distinct states for recurrent models)
# uint32_t n_threads; // number of threads to use for generation
# uint32_t n_threads_batch; // number of threads to use for batch processing
# enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
# enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id
# // (ignored if no pooling layer)
# // ref: https://github.com/ggerganov/llama.cpp/pull/2054
# float rope_freq_base; // RoPE base frequency, 0 = from model
# float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
# float yarn_ext_factor; // YaRN extrapolation mix factor, negative = from model
# float yarn_attn_factor; // YaRN magnitude scaling factor
# float yarn_beta_fast; // YaRN low correction dim
# float yarn_beta_slow; // YaRN high correction dim
# uint32_t yarn_orig_ctx; // YaRN original context size
# float defrag_thold; // defragment the KV cache if holes/size > thold, < 0 disabled (default)
# ggml_backend_sched_eval_callback cb_eval;
# void * cb_eval_user_data;
# enum ggml_type type_k; // data type for K cache
# enum ggml_type type_v; // data type for V cache
# // Keep the booleans together to avoid misalignment during copy-by-value.
# bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
# bool embeddings; // if true, extract embeddings (together with logits)
# bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
# // Abort callback
# // if it returns true, execution of llama_decode() will be aborted
# // currently works only with CPU execution
# ggml_abort_callback abort_callback;
# void * abort_callback_data;
# };
class llama_context_params(ctypes.Structure):
"""Parameters for llama_context
Attributes:
seed (int): RNG seed, -1 for random
n_ctx (int): text context, 0 = from model
n_batch (int): logical maximum batch size that can be submitted to llama_decode
n_ubatch (int): physical maximum batch size
n_seq_max (int): max number of sequences (i.e. distinct states for recurrent models)
n_threads (int): number of threads to use for generation
n_threads_batch (int): number of threads to use for batch processing
rope_scaling_type (int): RoPE scaling type, from `enum llama_rope_scaling_type`
pooling_type (int): whether to pool (sum) embedding results by sequence id (ignored if no pooling layer)
rope_freq_base (float): RoPE base frequency, 0 = from model
rope_freq_scale (float): RoPE frequency scaling factor, 0 = from model
yarn_ext_factor (float): YaRN extrapolation mix factor, negative = from model
yarn_attn_factor (float): YaRN magnitude scaling factor
yarn_beta_fast (float): YaRN low correction dim
yarn_beta_slow (float): YaRN high correction dim
yarn_orig_ctx (int): YaRN original context size
defrag_thold (float): defragment the KV cache if holes/size > thold, < 0 disabled (default)
cb_eval (ggml_backend_sched_eval_callback): callback for scheduling eval
cb_eval_user_data (ctypes.ctypes.c_void_p): user data for cb_eval
type_k (int): data type for K cache
type_v (int): data type for V cache
logits_all (bool): the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
embeddings (bool): if true, extract embeddings (together with logits)
offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU
abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted
abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback
"""
if TYPE_CHECKING:
seed: int
n_ctx: int
n_batch: int
n_ubatch: int
n_seq_max: int
n_threads: int
n_threads_batch: int
rope_scaling_type: int
pooling_type: int
rope_freq_base: float
rope_freq_scale: float
yarn_ext_factor: float
yarn_attn_factor: float
yarn_beta_fast: float
yarn_beta_slow: float
yarn_orig_ctx: int
defrag_thold: float
cb_eval: Callable[[ctypes.c_void_p, bool], bool]
cb_eval_user_data: ctypes.c_void_p
type_k: int
type_v: int
logits_all: bool
embeddings: bool
offload_kqv: bool
abort_callback: Callable[[ctypes.c_void_p], bool]
abort_callback_data: ctypes.c_void_p
_fields_ = [
("seed", ctypes.c_uint32),
("n_ctx", ctypes.c_uint32),
("n_batch", ctypes.c_uint32),
("n_ubatch", ctypes.c_uint32),
("n_seq_max", ctypes.c_uint32),
("n_threads", ctypes.c_uint32),
("n_threads_batch", ctypes.c_uint32),
("rope_scaling_type", ctypes.c_int),
("pooling_type", ctypes.c_int),
("rope_freq_base", ctypes.c_float),
("rope_freq_scale", ctypes.c_float),
("yarn_ext_factor", ctypes.c_float),
("yarn_attn_factor", ctypes.c_float),
("yarn_beta_fast", ctypes.c_float),
("yarn_beta_slow", ctypes.c_float),
("yarn_orig_ctx", ctypes.c_uint32),
("defrag_thold", ctypes.c_float),
("cb_eval", ggml_backend_sched_eval_callback),
("cb_eval_user_data", ctypes.c_void_p),
("type_k", ctypes.c_int),
("type_v", ctypes.c_int),
("logits_all", ctypes.c_bool),
("embeddings", ctypes.c_bool),
("offload_kqv", ctypes.c_bool),
("abort_callback", ggml_abort_callback),
("abort_callback_data", ctypes.c_void_p),
]
# // Signature for logging events
# // Note that text includes the new line character at the end for most events.
# // If your logging mechanism cannot handle that, check if the last character is '\n' and strip it
# // if it exists.
# // It might not exist for progress report where '.' is output repeatedly.
# typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data);
llama_log_callback = ctypes.CFUNCTYPE(
None, ctypes.c_int, ctypes.c_char_p, ctypes.c_void_p
)
"""Signature for logging events
Note that text includes the new line character at the end for most events.
If your logging mechanism cannot handle that, check if the last character is '\n' and strip it
if it exists.
It might not exist for progress report where '.' is output repeatedly."""
# // model quantization parameters
# typedef struct llama_model_quantize_params {
# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
# enum llama_ftype ftype; // quantize to this llama_ftype
# enum ggml_type output_tensor_type; // output tensor type
# enum ggml_type token_embedding_type; // itoken embeddings tensor type
# bool allow_requantize; // allow quantizing non-f32/f16 tensors
# bool quantize_output_tensor; // quantize output.weight
# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
# bool pure; // quantize all tensors to the default type
# void * imatrix; // pointer to importance matrix data
# void * kv_overrides; // pointer to vector containing overrides
# } llama_model_quantize_params;
class llama_model_quantize_params(ctypes.Structure):
"""Parameters for llama_model_quantize
Attributes:
nthread (int): number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
ftype (int): quantize to this llama_ftype
output_tensor_type (int): output tensor type
token_embedding_type (int): itoken embeddings tensor type
allow_requantize (bool): allow quantizing non-f32/f16 tensors
quantize_output_tensor (bool): quantize output.weight
only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
pure (bool): quantize all tensors to the default type
imatrix (ctypes.c_void_p): pointer to importance matrix data
kv_overrides (ctypes.c_void_p): pointer to vector containing overrides
"""
if TYPE_CHECKING:
nthread: int
ftype: int
output_tensor_type: int
token_embedding_type: int
allow_requantize: bool
quantize_output_tensor: bool
only_copy: bool
pure: bool
imatrix: ctypes.c_void_p
kv_overrides: ctypes.c_void_p
_fields_ = [
("nthread", ctypes.c_int32),
("ftype", ctypes.c_int),
("output_tensor_type", ctypes.c_int),
("token_embedding_type", ctypes.c_int),
("allow_requantize", ctypes.c_bool),
("quantize_output_tensor", ctypes.c_bool),
("only_copy", ctypes.c_bool),
("pure", ctypes.c_bool),
("imatrix", ctypes.c_void_p),
("kv_overrides", ctypes.c_void_p),
]
# // grammar types
# struct llama_grammar;
llama_grammar_p = ctypes.c_void_p
# // grammar element type
# enum llama_gretype {
# // end of rule definition
# LLAMA_GRETYPE_END = 0,
# // start of alternate definition for rule
# LLAMA_GRETYPE_ALT = 1,
# // non-terminal element: reference to rule
# LLAMA_GRETYPE_RULE_REF = 2,
# // terminal element: character (code point)
# LLAMA_GRETYPE_CHAR = 3,
# // inverse char(s) ([^a], [^a-b] [^abc])
# LLAMA_GRETYPE_CHAR_NOT = 4,
# // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
# // be an inclusive range ([a-z])
# LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
# // modifies a preceding LLAMA_GRETYPE_CHAR or
# // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
# LLAMA_GRETYPE_CHAR_ALT = 6,
# };
LLAMA_GRETYPE_END = 0
LLAMA_GRETYPE_ALT = 1
LLAMA_GRETYPE_RULE_REF = 2
LLAMA_GRETYPE_CHAR = 3
LLAMA_GRETYPE_CHAR_NOT = 4
LLAMA_GRETYPE_CHAR_RNG_UPPER = 5
LLAMA_GRETYPE_CHAR_ALT = 6
# typedef struct llama_grammar_element {
# enum llama_gretype type;
# uint32_t value; // Unicode code point or rule ID
# } llama_grammar_element;
class llama_grammar_element(ctypes.Structure):
if TYPE_CHECKING:
type: int
value: int
_fields_ = [
("type", ctypes.c_int),
("value", ctypes.c_uint32),
]
llama_grammar_element_p = ctypes.POINTER(llama_grammar_element)
# // performance timing information
# struct llama_timings {
# double t_start_ms;
# double t_end_ms;
# double t_load_ms;
# double t_sample_ms;
# double t_p_eval_ms;
# double t_eval_ms;
# int32_t n_sample;
# int32_t n_p_eval;
# int32_t n_eval;
# };
class llama_timings(ctypes.Structure):
if TYPE_CHECKING:
t_start_ms: float
t_end_ms: float
t_load_ms: float
t_sample_ms: float
t_p_eval_ms: float
t_eval_ms: float
n_sample: int
n_p_eval: int
n_eval: int
_fields_ = [
("t_start_ms", ctypes.c_double),
("t_end_ms", ctypes.c_double),
("t_load_ms", ctypes.c_double),
("t_sample_ms", ctypes.c_double),
("t_p_eval_ms", ctypes.c_double),
("t_eval_ms", ctypes.c_double),
("n_sample", ctypes.c_int32),
("n_p_eval", ctypes.c_int32),
("n_eval", ctypes.c_int32),
]
# // used in chat template
# typedef struct llama_chat_message {
# const char * role;
# const char * content;
# } llama_chat_message;
class llama_chat_message(ctypes.Structure):
_fields_ = [
("role", ctypes.c_char_p),
("content", ctypes.c_char_p),
]
# // Helpers for getting default parameters
# LLAMA_API struct llama_model_params llama_model_default_params(void);
@ctypes_function(
"llama_model_default_params",
[],
llama_model_params,
)
def llama_model_default_params() -> llama_model_params:
"""Get default parameters for llama_model"""
...
# LLAMA_API struct llama_context_params llama_context_default_params(void);
@ctypes_function(
"llama_context_default_params",
[],
llama_context_params,
)
def llama_context_default_params() -> llama_context_params:
"""Get default parameters for llama_context"""
...
# LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void);
@ctypes_function(
"llama_model_quantize_default_params",
[],
llama_model_quantize_params,
)
def llama_model_quantize_default_params() -> llama_model_quantize_params:
"""Get default parameters for llama_model_quantize"""
...
# // Initialize the llama + ggml backend
# // If numa is true, use NUMA optimizations