forked from abetlen/llama-cpp-python
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathllama.py
More file actions
2374 lines (2096 loc) · 90.8 KB
/
llama.py
File metadata and controls
2374 lines (2096 loc) · 90.8 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
import os
import sys
import uuid
import time
import multiprocessing
from abc import ABC, abstractmethod
from typing import (
List,
Optional,
Union,
Generator,
Sequence,
Iterator,
Deque,
Tuple,
Callable,
)
from collections import deque, OrderedDict
import diskcache
import ctypes
from .llama_types import *
from .llama_grammar import LlamaGrammar
import llama_cpp.llama_cpp as llama_cpp
import llama_cpp.llama_chat_format as llama_chat_format
import numpy as np
import numpy.typing as npt
from ._utils import suppress_stdout_stderr
class BaseLlamaCache(ABC):
"""Base cache class for a llama.cpp model."""
def __init__(self, capacity_bytes: int = (2 << 30)):
self.capacity_bytes = capacity_bytes
@property
@abstractmethod
def cache_size(self) -> int:
raise NotImplementedError
def _find_longest_prefix_key(
self,
key: Tuple[int, ...],
) -> Optional[Tuple[int, ...]]:
pass
@abstractmethod
def __getitem__(self, key: Sequence[int]) -> "LlamaState":
raise NotImplementedError
@abstractmethod
def __contains__(self, key: Sequence[int]) -> bool:
raise NotImplementedError
@abstractmethod
def __setitem__(self, key: Sequence[int], value: "LlamaState") -> None:
raise NotImplementedError
class LlamaRAMCache(BaseLlamaCache):
"""Cache for a llama.cpp model using RAM."""
def __init__(self, capacity_bytes: int = (2 << 30)):
super().__init__(capacity_bytes)
self.capacity_bytes = capacity_bytes
self.cache_state: OrderedDict[Tuple[int, ...], "LlamaState"] = OrderedDict()
@property
def cache_size(self):
return sum([state.llama_state_size for state in self.cache_state.values()])
def _find_longest_prefix_key(
self,
key: Tuple[int, ...],
) -> Optional[Tuple[int, ...]]:
min_len = 0
min_key = None
keys = (
(k, Llama.longest_token_prefix(k, key)) for k in self.cache_state.keys()
)
for k, prefix_len in keys:
if prefix_len > min_len:
min_len = prefix_len
min_key = k
return min_key
def __getitem__(self, key: Sequence[int]) -> "LlamaState":
key = tuple(key)
_key = self._find_longest_prefix_key(key)
if _key is None:
raise KeyError("Key not found")
value = self.cache_state[_key]
self.cache_state.move_to_end(_key)
return value
def __contains__(self, key: Sequence[int]) -> bool:
return self._find_longest_prefix_key(tuple(key)) is not None
def __setitem__(self, key: Sequence[int], value: "LlamaState"):
key = tuple(key)
if key in self.cache_state:
del self.cache_state[key]
self.cache_state[key] = value
while self.cache_size > self.capacity_bytes and len(self.cache_state) > 0:
self.cache_state.popitem(last=False)
# Alias for backwards compatibility
LlamaCache = LlamaRAMCache
class LlamaDiskCache(BaseLlamaCache):
"""Cache for a llama.cpp model using disk."""
def __init__(
self, cache_dir: str = ".cache/llama_cache", capacity_bytes: int = (2 << 30)
):
super().__init__(capacity_bytes)
self.cache = diskcache.Cache(cache_dir)
@property
def cache_size(self):
return int(self.cache.volume()) # type: ignore
def _find_longest_prefix_key(
self,
key: Tuple[int, ...],
) -> Optional[Tuple[int, ...]]:
min_len = 0
min_key: Optional[Tuple[int, ...]] = None
for k in self.cache.iterkeys(): # type: ignore
prefix_len = Llama.longest_token_prefix(k, key)
if prefix_len > min_len:
min_len = prefix_len
min_key = k # type: ignore
return min_key
def __getitem__(self, key: Sequence[int]) -> "LlamaState":
key = tuple(key)
_key = self._find_longest_prefix_key(key)
if _key is None:
raise KeyError("Key not found")
value: "LlamaState" = self.cache.pop(_key) # type: ignore
# NOTE: This puts an integer as key in cache, which breaks,
# Llama.longest_token_prefix(k, key) above since k is not a tuple of ints/tokens
# self.cache.push(_key, side="front") # type: ignore
return value
def __contains__(self, key: Sequence[int]) -> bool:
return self._find_longest_prefix_key(tuple(key)) is not None
def __setitem__(self, key: Sequence[int], value: "LlamaState"):
print("LlamaDiskCache.__setitem__: called", file=sys.stderr)
key = tuple(key)
if key in self.cache:
print("LlamaDiskCache.__setitem__: delete", file=sys.stderr)
del self.cache[key]
self.cache[key] = value
print("LlamaDiskCache.__setitem__: set", file=sys.stderr)
while self.cache_size > self.capacity_bytes and len(self.cache) > 0:
key_to_remove = next(iter(self.cache))
del self.cache[key_to_remove]
print("LlamaDiskCache.__setitem__: trim", file=sys.stderr)
class LlamaState:
def __init__(
self,
input_ids: npt.NDArray[np.intc],
scores: npt.NDArray[np.single],
n_tokens: int,
llama_state: bytes,
llama_state_size: int,
):
self.input_ids = input_ids
self.scores = scores
self.n_tokens = n_tokens
self.llama_state = llama_state
self.llama_state_size = llama_state_size
LogitsProcessor = Callable[
[npt.NDArray[np.intc], npt.NDArray[np.single]], npt.NDArray[np.single]
]
class LogitsProcessorList(List[LogitsProcessor]):
def __call__(
self, input_ids: npt.NDArray[np.intc], scores: npt.NDArray[np.single]
) -> npt.NDArray[np.single]:
for processor in self:
scores = processor(input_ids, scores)
return scores
StoppingCriteria = Callable[[npt.NDArray[np.intc], npt.NDArray[np.single]], bool]
class StoppingCriteriaList(List[StoppingCriteria]):
def __call__(
self, input_ids: npt.NDArray[np.intc], logits: npt.NDArray[np.single]
) -> bool:
return any([stopping_criteria(input_ids, logits) for stopping_criteria in self])
class _LlamaModel:
"""Intermediate Python wrapper for a llama.cpp llama_model.
NOTE: For stability it's recommended you use the Llama class instead."""
_llama_free_model = None
# NOTE: this must be "saved" here to avoid exceptions when calling __del__
suppress_stdout_stderr = suppress_stdout_stderr
def __init__(
self,
*,
path_model: str,
params: llama_cpp.llama_model_params,
verbose: bool = True,
):
self.path_model = path_model
self.params = params
self.verbose = verbose
self._llama_free_model = llama_cpp._lib.llama_free_model # type: ignore
if not os.path.exists(path_model):
raise ValueError(f"Model path does not exist: {path_model}")
with suppress_stdout_stderr(disable=self.verbose):
self.model = llama_cpp.llama_load_model_from_file(
self.path_model.encode("utf-8"), self.params
)
def __del__(self):
with self.suppress_stdout_stderr(disable=self.verbose):
if self.model is not None and self._llama_free_model is not None:
self._llama_free_model(self.model)
self.model = None
def vocab_type(self) -> int:
assert self.model is not None
return llama_cpp.llama_vocab_type(self.model)
def n_vocab(self) -> int:
assert self.model is not None
return llama_cpp.llama_n_vocab(self.model)
def n_ctx_train(self) -> int:
assert self.model is not None
return llama_cpp.llama_n_ctx_train(self.model)
def n_embd(self) -> int:
assert self.model is not None
return llama_cpp.llama_n_embd(self.model)
def rope_freq_scale_train(self) -> float:
assert self.model is not None
return llama_cpp.llama_rope_freq_scale_train(self.model)
def desc(self) -> str:
assert self.model is not None
buf = ctypes.create_string_buffer(1024)
llama_cpp.llama_model_desc(self.model, buf, 1024) # type: ignore
return buf.value.decode("utf-8")
def size(self) -> int:
assert self.model is not None
return llama_cpp.llama_model_size(self.model)
def n_params(self) -> int:
assert self.model is not None
return llama_cpp.llama_model_n_params(self.model)
def get_tensor(self, name: str) -> ctypes.c_void_p:
assert self.model is not None
return llama_cpp.llama_get_model_tensor(self.model, name.encode("utf-8"))
def apply_lora_from_file(
self,
lora_path: str,
scale: float,
path_base_model: Optional[str],
n_threads: int,
):
assert self.model is not None
return llama_cpp.llama_model_apply_lora_from_file(
self.model,
lora_path.encode("utf-8"),
scale,
path_base_model.encode("utf-8")
if path_base_model is not None
else llama_cpp.c_char_p(0),
n_threads,
)
# Vocab
def token_get_text(self, token: int) -> str:
# TODO: Fix
assert self.model is not None
return llama_cpp.llama_token_get_text(self.model, token).decode("utf-8")
def token_get_score(self, token: int) -> float:
assert self.model is not None
return llama_cpp.llama_token_get_score(self.model, token)
def token_get_type(self, token: int) -> int:
assert self.model is not None
return llama_cpp.llama_token_get_type(self.model, token)
# Special tokens
def token_bos(self) -> int:
assert self.model is not None
return llama_cpp.llama_token_bos(self.model)
def token_eos(self) -> int:
assert self.model is not None
return llama_cpp.llama_token_eos(self.model)
def token_nl(self) -> int:
assert self.model is not None
return llama_cpp.llama_token_nl(self.model)
def token_prefix(self) -> int:
assert self.model is not None
return llama_cpp.llama_token_prefix(self.model)
def token_middle(self) -> int:
assert self.model is not None
return llama_cpp.llama_token_middle(self.model)
def token_suffix(self) -> int:
assert self.model is not None
return llama_cpp.llama_token_suffix(self.model)
def token_eot(self) -> int:
assert self.model is not None
return llama_cpp.llama_token_eot(self.model)
# Tokenization
def tokenize(self, text: bytes, add_bos: bool, special: bool):
assert self.model is not None
n_ctx = self.n_ctx_train()
tokens = (llama_cpp.llama_token * n_ctx)()
n_tokens = llama_cpp.llama_tokenize(
self.model, text, len(text), tokens, n_ctx, add_bos, special
)
if n_tokens < 0:
n_tokens = abs(n_tokens)
tokens = (llama_cpp.llama_token * n_tokens)()
n_tokens = llama_cpp.llama_tokenize(
self.model, text, len(text), tokens, n_tokens, add_bos, special
)
if n_tokens < 0:
raise RuntimeError(
f'Failed to tokenize: text="{text}" n_tokens={n_tokens}'
)
return list(tokens[:n_tokens])
def token_to_piece(self, token: int) -> bytes:
assert self.model is not None
buf = ctypes.create_string_buffer(32)
llama_cpp.llama_token_to_piece(self.model, token, buf, 32) # type: ignore
return bytes(buf)
def detokenize(self, tokens: List[int]) -> bytes:
assert self.model is not None
output = b""
size = 32
buffer = (ctypes.c_char * size)()
for token in tokens:
n = llama_cpp.llama_token_to_piece(
self.model, llama_cpp.llama_token(token), buffer, size
)
assert n <= size
output += bytes(buffer[:n])
# NOTE: Llama1 models automatically added a space at the start of the prompt
# this line removes a leading space if the first token is a beginning of sentence token
return (
output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() else output
)
@staticmethod
def default_params():
"""Get the default llama_model_params."""
return llama_cpp.llama_model_default_params()
class _LlamaContext:
"""Intermediate Python wrapper for a llama.cpp llama_context.
NOTE: For stability it's recommended you use the Llama class instead."""
_llama_free = None
# NOTE: this must be "saved" here to avoid exceptions when calling __del__
suppress_stdout_stderr = suppress_stdout_stderr
def __init__(
self,
*,
model: _LlamaModel,
params: llama_cpp.llama_context_params,
verbose: bool = True,
):
self.model = model
self.params = params
self.verbose = verbose
self._llama_free = llama_cpp._lib.llama_free # type: ignore
with suppress_stdout_stderr(disable=self.verbose):
self.ctx = llama_cpp.llama_new_context_with_model(
self.model.model, self.params
)
def __del__(self):
with self.suppress_stdout_stderr(disable=self.verbose):
if self.ctx is not None and self._llama_free is not None:
self._llama_free(self.ctx)
self.ctx = None
def n_ctx(self) -> int:
assert self.ctx is not None
return llama_cpp.llama_n_ctx(self.ctx)
def kv_cache_clear(self):
assert self.ctx is not None
llama_cpp.llama_kv_cache_clear(self.ctx)
def kv_cache_seq_rm(self, seq_id: int, p0: int, p1: int):
assert self.ctx is not None
llama_cpp.llama_kv_cache_seq_rm(self.ctx, seq_id, p0, p1)
def kv_cache_seq_cp(self, seq_id_src: int, seq_id_dst: int, p0: int, p1: int):
assert self.ctx is not None
llama_cpp.llama_kv_cache_seq_cp(self.ctx, seq_id_src, seq_id_dst, p0, p1)
def kv_cache_seq_keep(self, seq_id: int):
assert self.ctx is not None
llama_cpp.llama_kv_cache_seq_keep(self.ctx, seq_id)
def kv_cache_seq_shift(self, seq_id: int, p0: int, p1: int, shift: int):
assert self.ctx is not None
llama_cpp.llama_kv_cache_seq_shift(self.ctx, seq_id, p0, p1, shift)
def get_state_size(self) -> int:
assert self.ctx is not None
return llama_cpp.llama_get_state_size(self.ctx)
# TODO: copy_state_data
# TODO: set_state_data
# TODO: llama_load_session_file
# TODO: llama_save_session_file
def decode(self, batch: "_LlamaBatch"):
assert self.ctx is not None
assert batch.batch is not None
return_code = llama_cpp.llama_decode(
ctx=self.ctx,
batch=batch.batch,
)
if return_code != 0:
raise RuntimeError(f"llama_decode returned {return_code}")
def set_n_threads(self, n_threads: int, n_threads_batch: int):
assert self.ctx is not None
llama_cpp.llama_set_n_threads(self.ctx, n_threads, n_threads_batch)
def get_logits(self):
assert self.ctx is not None
return llama_cpp.llama_get_logits(self.ctx)
def get_logits_ith(self, i: int):
assert self.ctx is not None
return llama_cpp.llama_get_logits_ith(self.ctx, i)
def get_embeddings(self):
assert self.ctx is not None
return llama_cpp.llama_get_embeddings(self.ctx)
# Sampling functions
def set_rng_seed(self, seed: int):
assert self.ctx is not None
llama_cpp.llama_set_rng_seed(self.ctx, seed)
def sample_repetition_penalties(
self,
candidates: "_LlamaTokenDataArray",
last_tokens_data: "llama_cpp.Array[llama_cpp.llama_token]",
penalty_last_n: int,
penalty_repeat: float,
penalty_freq: float,
penalty_present: float,
):
assert self.ctx is not None
llama_cpp.llama_sample_repetition_penalties(
self.ctx,
ctypes.byref(candidates.candidates), # type: ignore
last_tokens_data,
penalty_last_n,
penalty_repeat,
penalty_freq,
penalty_present,
)
def sample_classifier_free_guidance(
self,
candidates: "_LlamaTokenDataArray",
guidance_ctx: "_LlamaContext",
scale: float,
):
assert self.ctx is not None
assert guidance_ctx.ctx is not None
llama_cpp.llama_sample_classifier_free_guidance(
self.ctx,
ctypes.byref(candidates.candidates), # type: ignore
guidance_ctx.ctx,
scale,
)
def sample_softmax(self, candidates: "_LlamaTokenDataArray"):
assert self.ctx is not None
llama_cpp.llama_sample_softmax(
self.ctx,
ctypes.byref(candidates.candidates), # type: ignore
)
def sample_top_k(self, candidates: "_LlamaTokenDataArray", k: int, min_keep: int):
assert self.ctx is not None
llama_cpp.llama_sample_top_k(
self.ctx, ctypes.byref(candidates.candidates), k, min_keep # type: ignore
)
def sample_top_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int):
assert self.ctx is not None
llama_cpp.llama_sample_top_p(
self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore
)
def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int):
assert self.ctx is not None
llama_cpp.llama_sample_min_p(
self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore
)
def sample_tail_free(
self, candidates: "_LlamaTokenDataArray", z: float, min_keep: int
):
assert self.ctx is not None
llama_cpp.llama_sample_tail_free(
self.ctx, ctypes.byref(candidates.candidates), z, min_keep # type: ignore
)
def sample_typical(
self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int
):
assert self.ctx is not None
llama_cpp.llama_sample_typical(
self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore
)
def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float):
assert self.ctx is not None
llama_cpp.llama_sample_temp(
self.ctx, ctypes.byref(candidates.candidates), temp # type: ignore
)
def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGrammar):
assert self.ctx is not None
assert grammar.grammar is not None
llama_cpp.llama_sample_grammar(
self.ctx,
ctypes.byref(candidates.candidates), # type: ignore
grammar.grammar,
)
def sample_token_mirostat(
self,
candidates: "_LlamaTokenDataArray",
tau: float,
eta: float,
m: int,
mu: float,
) -> int:
assert self.ctx is not None
return llama_cpp.llama_sample_token_mirostat(
self.ctx,
ctypes.byref(candidates.candidates), # type: ignore
tau,
eta,
m,
ctypes.pointer(ctypes.c_float(mu)),
)
def sample_token_mirostat_v2(
self, candidates: "_LlamaTokenDataArray", tau: float, eta: float, mu: float
) -> int:
assert self.ctx is not None
return llama_cpp.llama_sample_token_mirostat_v2(
self.ctx,
ctypes.byref(candidates.candidates), # type: ignore
tau,
eta,
ctypes.pointer(ctypes.c_float(mu)),
)
def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int:
assert self.ctx is not None
return llama_cpp.llama_sample_token_greedy(
self.ctx,
ctypes.byref(candidates.candidates), # type: ignore
)
def sample_token(self, candidates: "_LlamaTokenDataArray") -> int:
assert self.ctx is not None
return llama_cpp.llama_sample_token(
self.ctx,
ctypes.byref(candidates.candidates), # type: ignore
)
# Grammar
def grammar_accept_token(self, grammar: LlamaGrammar, token: int):
assert self.ctx is not None
assert grammar.grammar is not None
llama_cpp.llama_grammar_accept_token(self.ctx, grammar.grammar, token)
def reset_timings(self):
assert self.ctx is not None
llama_cpp.llama_reset_timings(self.ctx)
def print_timings(self):
assert self.ctx is not None
llama_cpp.llama_print_timings(self.ctx)
# Utility functions
@staticmethod
def default_params():
"""Get the default llama_context_params."""
return llama_cpp.llama_context_default_params()
class _LlamaBatch:
_llama_batch_free = None
# NOTE: this must be "saved" here to avoid exceptions when calling __del__
suppress_stdout_stderr = suppress_stdout_stderr
def __init__(
self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True
):
self.n_tokens = n_tokens
self.embd = embd
self.n_seq_max = n_seq_max
self.verbose = verbose
self._llama_batch_free = llama_cpp._lib.llama_batch_free # type: ignore
with suppress_stdout_stderr(disable=self.verbose):
self.batch = llama_cpp.llama_batch_init(
self.n_tokens, self.embd, self.n_seq_max
)
def __del__(self):
with self.suppress_stdout_stderr(disable=self.verbose):
if self.batch is not None and self._llama_batch_free is not None:
self._llama_batch_free(self.batch)
self.batch = None
def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool):
assert self.batch is not None
n_tokens = len(batch)
self.batch.n_tokens = n_tokens
for i in range(n_tokens):
self.batch.token[i] = batch[i]
self.batch.pos[i] = n_past + i
self.batch.seq_id[i][0] = 0
self.batch.n_seq_id[i] = 1
self.batch.logits[i] = logits_all
self.batch.logits[n_tokens - 1] = True
class _LlamaTokenDataArray:
def __init__(self, *, n_vocab: int):
self.n_vocab = n_vocab
self.candidates_data = np.array(
[],
dtype=np.dtype(
[("id", np.intc), ("logit", np.single), ("p", np.single)], align=True
),
)
self.candidates_data.resize(3, self.n_vocab, refcheck=False)
self.candidates = llama_cpp.llama_token_data_array(
data=self.candidates_data.ctypes.data_as(llama_cpp.llama_token_data_p),
size=self.n_vocab,
sorted=False,
)
self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc)
self.default_candidates_data_p = np.zeros(self.n_vocab, dtype=np.single)
def copy_logits(self, logits: npt.NDArray[np.single]):
self.candidates_data["id"][:] = self.default_candidates_data_id
self.candidates_data["logit"][:] = logits
self.candidates_data["p"][:] = self.default_candidates_data_p
self.candidates.data = self.candidates_data.ctypes.data_as(
llama_cpp.llama_token_data_p
)
self.candidates.sorted = llama_cpp.c_bool(False)
self.candidates.size = llama_cpp.c_size_t(self.n_vocab)
class Llama:
"""High-level Python wrapper for a llama.cpp model."""
__backend_initialized = False
def __init__(
self,
model_path: str,
*,
# Model Params
n_gpu_layers: int = 0,
split_mode: int = llama_cpp.LLAMA_SPLIT_LAYER,
main_gpu: int = 0,
tensor_split: Optional[List[float]] = None,
vocab_only: bool = False,
use_mmap: bool = True,
use_mlock: bool = False,
kv_overrides: Optional[Dict[str, Union[bool, int, float]]] = None,
# Context Params
seed: int = llama_cpp.LLAMA_DEFAULT_SEED,
n_ctx: int = 512,
n_batch: int = 512,
n_threads: Optional[int] = None,
n_threads_batch: Optional[int] = None,
rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED,
rope_freq_base: float = 0.0,
rope_freq_scale: float = 0.0,
yarn_ext_factor: float = -1.0,
yarn_attn_factor: float = 1.0,
yarn_beta_fast: float = 32.0,
yarn_beta_slow: float = 1.0,
yarn_orig_ctx: int = 0,
mul_mat_q: bool = True,
logits_all: bool = False,
embedding: bool = False,
offload_kqv: bool = False,
# Sampling Params
last_n_tokens_size: int = 64,
# LoRA Params
lora_base: Optional[str] = None,
lora_scale: float = 1.0,
lora_path: Optional[str] = None,
# Backend Params
numa: bool = False,
# Chat Format Params
chat_format: str = "llama-2",
chat_handler: Optional[llama_chat_format.LlamaChatCompletionHandler] = None,
# Misc
verbose: bool = True,
# Extra Params
**kwargs, # type: ignore
):
"""Load a llama.cpp model from `model_path`.
Examples:
Basic usage
>>> import llama_cpp
>>> model = llama_cpp.Llama(
... model_path="path/to/model",
... )
>>> print(model("The quick brown fox jumps ", stop=["."])["choices"][0]["text"])
the lazy dog
Loading a chat model
>>> import llama_cpp
>>> model = llama_cpp.Llama(
... model_path="path/to/model",
... chat_format="llama-2",
... )
>>> print(model.create_chat_completion(
... messages=[{
... "role": "user",
... "content": "what is the meaning of life?"
... }]
... ))
Args:
model_path: Path to the model.
n_gpu_layers: Number of layers to offload to GPU (-ngl). If -1, all layers are offloaded.
split_mode: How to split the model across GPUs. See llama_cpp.LLAMA_SPLIT_* for options.
main_gpu: 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: How split tensors should be distributed across GPUs. If None, the model is not split.
vocab_only: Only load the vocabulary no weights.
use_mmap: Use mmap if possible.
use_mlock: Force the system to keep the model in RAM.
kv_overrides: Key-value overrides for the model.
seed: RNG seed, -1 for random
n_ctx: Text context, 0 = from model
n_batch: Prompt processing maximum batch size
n_threads: Number of threads to use for generation
n_threads_batch: Number of threads to use for batch processing
rope_scaling_type: RoPE scaling type, from `enum llama_rope_scaling_type`. ref: https://github.com/ggerganov/llama.cpp/pull/2054
rope_freq_base: RoPE base frequency, 0 = from model
rope_freq_scale: RoPE frequency scaling factor, 0 = from model
yarn_ext_factor: YaRN extrapolation mix factor, negative = from model
yarn_attn_factor: YaRN magnitude scaling factor
yarn_beta_fast: YaRN low correction dim
yarn_beta_slow: YaRN high correction dim
yarn_orig_ctx: YaRN original context size
logits_all: Return logits for all tokens, not just the last token. Must be True for completion to return logprobs.
embedding: Embedding mode only.
offload_kqv: Offload K, Q, V to GPU.
last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque.
lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model.
lora_path: Path to a LoRA file to apply to the model.
numa: Enable NUMA support. (NOTE: The initial value of this parameter is used for the remainder of the program as this value is set in llama_backend_init)
chat_format: String specifying the chat format to use when calling create_chat_completion.
chat_handler: Optional chat handler to use when calling create_chat_completion.
verbose: Print verbose output to stderr.
Raises:
ValueError: If the model path does not exist.
Returns:
A Llama instance.
"""
self.verbose = verbose
self.numa = numa
if not Llama.__backend_initialized:
with suppress_stdout_stderr(disable=self.verbose):
llama_cpp.llama_backend_init(self.numa)
Llama.__backend_initialized = True
self.model_path = model_path
# Model Params
self.model_params = llama_cpp.llama_model_default_params()
self.model_params.n_gpu_layers = (
0x7FFFFFFF if n_gpu_layers == -1 else n_gpu_layers
) # 0x7FFFFFFF is INT32 max, will be auto set to all layers
self.model_params.split_mode = split_mode
self.model_params.main_gpu = main_gpu
self.tensor_split = tensor_split
self._c_tensor_split = None
if self.tensor_split is not None:
if len(self.tensor_split) > llama_cpp.LLAMA_MAX_DEVICES:
raise ValueError(
f"Attempt to split tensors that exceed maximum supported devices. Current LLAMA_MAX_DEVICES={llama_cpp.LLAMA_MAX_DEVICES}"
)
# Type conversion and expand the list to the length of LLAMA_MAX_DEVICES
FloatArray = ctypes.c_float * llama_cpp.LLAMA_MAX_DEVICES
self._c_tensor_split = FloatArray(
*tensor_split # type: ignore
) # keep a reference to the array so it is not gc'd
self.model_params.tensor_split = self._c_tensor_split
self.model_params.vocab_only = vocab_only
self.model_params.use_mmap = use_mmap if lora_path is None else False
self.model_params.use_mlock = use_mlock
self.kv_overrides = kv_overrides
if kv_overrides is not None:
n_overrides = len(kv_overrides)
self._kv_overrides_array = llama_cpp.llama_model_kv_override * (n_overrides + 1)
self._kv_overrides_array_keys = []
for k, v in kv_overrides.items():
key_buf = ctypes.create_string_buffer(k.encode("utf-8"))
self._kv_overrides_array_keys.append(key_buf)
self._kv_overrides_array[i].key = key_buf
if isinstance(v, int):
self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_INT
self._kv_overrides_array[i].value.int_value = v
elif isinstance(v, float):
self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_FLOAT
self._kv_overrides_array[i].value.float_value = v
elif isinstance(v, bool):
self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_BOOL
self._kv_overrides_array[i].value.bool_value = v
else:
raise ValueError(f"Unknown value type for {k}: {v}")
self._kv_overrides_array_sentinel_key = b'\0'
# null array sentinel
self._kv_overrides_array[n_overrides].key = self._kv_overrides_array_sentinel_key
self.model_params.kv_overrides = self._kv_overrides_array
self.n_batch = min(n_ctx, n_batch) # ???
self.n_threads = n_threads or max(multiprocessing.cpu_count() // 2, 1)
self.n_threads_batch = n_threads_batch or max(
multiprocessing.cpu_count() // 2, 1
)
# Context Params
self.context_params = llama_cpp.llama_context_default_params()
self.context_params.seed = seed
self.context_params.n_ctx = n_ctx
self.context_params.n_batch = self.n_batch
self.context_params.n_threads = self.n_threads
self.context_params.n_threads_batch = self.n_threads_batch
self.context_params.rope_scaling_type = (
rope_scaling_type
if rope_scaling_type is not None
else llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED
)
self.context_params.rope_freq_base = (
rope_freq_base if rope_freq_base != 0.0 else 0
)
self.context_params.rope_freq_scale = (
rope_freq_scale if rope_freq_scale != 0.0 else 0
)
self.context_params.yarn_ext_factor = (
yarn_ext_factor if yarn_ext_factor != 0.0 else 0
)
self.context_params.yarn_attn_factor = (
yarn_attn_factor if yarn_attn_factor != 0.0 else 0
)
self.context_params.yarn_beta_fast = (
yarn_beta_fast if yarn_beta_fast != 0.0 else 0
)
self.context_params.yarn_beta_slow = (
yarn_beta_slow if yarn_beta_slow != 0.0 else 0
)
self.context_params.yarn_orig_ctx = yarn_orig_ctx if yarn_orig_ctx != 0 else 0
self.context_params.mul_mat_q = mul_mat_q
self.context_params.logits_all = logits_all
self.context_params.embedding = embedding
self.context_params.offload_kqv = offload_kqv
# Sampling Params
self.last_n_tokens_size = last_n_tokens_size
self.cache: Optional[BaseLlamaCache] = None
self.lora_base = lora_base
self.lora_scale = lora_scale
self.lora_path = lora_path
if not os.path.exists(model_path):
raise ValueError(f"Model path does not exist: {model_path}")
self._model = _LlamaModel(
path_model=self.model_path, params=self.model_params, verbose=self.verbose
)
# Set the default value for the context and correct the batch
if n_ctx == 0:
n_ctx = self._model.n_ctx_train()
self.n_batch = min(n_ctx, n_batch)
self.context_params.n_ctx = self._model.n_ctx_train()
self.context_params.n_batch = self.n_batch
self._ctx = _LlamaContext(
model=self._model,
params=self.context_params,
verbose=self.verbose,
)
self._batch = _LlamaBatch(
n_tokens=self.n_batch,
embd=0,
n_seq_max=self.context_params.n_ctx,
verbose=self.verbose,
)
if self.lora_path:
if self._model.apply_lora_from_file(
self.lora_path,
self.lora_scale,
self.lora_base,
self.n_threads,
):
raise RuntimeError(
f"Failed to apply LoRA from lora path: {self.lora_path} to base path: {self.lora_base}"
)
if self.verbose:
print(llama_cpp.llama_print_system_info().decode("utf-8"), file=sys.stderr)
self.chat_format = chat_format
self.chat_handler = chat_handler
self._n_vocab = self.n_vocab()
self._n_ctx = self.n_ctx()
self._token_nl = self.token_nl()
self._token_eos = self.token_eos()