Repository navigation
Expand file tree
/
Copy path_internals.py
More file actions
868 lines (722 loc) · 27.5 KB
/
Copy path_internals.py
File metadata and controls
868 lines (722 loc) · 27.5 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
from __future__ import annotations
import os
import ctypes
from typing import (
Dict,
List,
Optional,
Sequence,
)
from dataclasses import dataclass, field
from contextlib import ExitStack
import numpy as np
import numpy.typing as npt
from .llama_types import *
from .llama_grammar import LlamaGrammar
from ._utils import suppress_stdout_stderr
import llama_cpp.llama_cpp as llama_cpp
# Python wrappers over llama.h structs
class _LlamaModel:
"""Intermediate Python wrapper for a llama.cpp llama_model.
NOTE: For stability it's recommended you use the Llama class instead."""
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._exit_stack = ExitStack()
self.model = None
if not os.path.exists(path_model):
raise ValueError(f"Model path does not exist: {path_model}")
with suppress_stdout_stderr(disable=verbose):
self.model = llama_cpp.llama_load_model_from_file(
self.path_model.encode("utf-8"), self.params
)
if self.model is None:
raise ValueError(f"Failed to load model from file: {path_model}")
def free_model():
if self.model is None:
return
llama_cpp.llama_free_model(self.model)
self.model = None
self._exit_stack.callback(free_model)
def close(self):
self._exit_stack.close()
def __del__(self):
self.close()
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)
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 ctypes.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_attr(self, token: int) -> int:
assert self.model is not None
return llama_cpp.llama_token_get_attr(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_cls(self) -> int:
assert self.model is not None
return llama_cpp.llama_token_cls(self.model)
def token_sep(self) -> int:
assert self.model is not None
return llama_cpp.llama_token_sep(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)
def add_bos_token(self) -> int:
assert self.model is not None
return llama_cpp.llama_add_bos_token(self.model)
def add_eos_token(self) -> int:
assert self.model is not None
return llama_cpp.llama_add_eos_token(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, special: bool = False) -> bytes:
assert self.model is not None
buf = ctypes.create_string_buffer(32)
llama_cpp.llama_token_to_piece(self.model, token, buf, 32, 0, special)
return bytes(buf)
def detokenize(self, tokens: List[int], special: bool = False) -> 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, 0, special
)
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() and output[0:1] == b" "
else output
)
# Extra
def metadata(self) -> Dict[str, str]:
assert self.model is not None
metadata: Dict[str, str] = {}
buffer_size = 1024
buffer = ctypes.create_string_buffer(buffer_size)
# zero the buffer
buffer.value = b"\0" * buffer_size
# iterate over model keys
for i in range(llama_cpp.llama_model_meta_count(self.model)):
nbytes = llama_cpp.llama_model_meta_key_by_index(
self.model, i, buffer, buffer_size
)
if nbytes > buffer_size:
buffer_size = nbytes + 1
buffer = ctypes.create_string_buffer(buffer_size)
nbytes = llama_cpp.llama_model_meta_key_by_index(
self.model, i, buffer, buffer_size
)
key = buffer.value.decode("utf-8")
nbytes = llama_cpp.llama_model_meta_val_str_by_index(
self.model, i, buffer, buffer_size
)
if nbytes > buffer_size:
buffer_size = nbytes + 1
buffer = ctypes.create_string_buffer(buffer_size)
nbytes = llama_cpp.llama_model_meta_val_str_by_index(
self.model, i, buffer, buffer_size
)
value = buffer.value.decode("utf-8")
metadata[key] = value
return metadata
@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."""
def __init__(
self,
*,
model: _LlamaModel,
params: llama_cpp.llama_context_params,
verbose: bool = True,
):
self.model = model
self.params = params
self.verbose = verbose
self._exit_stack = ExitStack()
self.ctx = None
assert self.model.model is not None
self.ctx = llama_cpp.llama_new_context_with_model(self.model.model, self.params)
if self.ctx is None:
raise ValueError("Failed to create llama_context")
def free_ctx():
if self.ctx is None:
return
llama_cpp.llama_free(self.ctx)
self.ctx = None
self._exit_stack.callback(free_ctx)
def close(self):
self._exit_stack.close()
def __del__(self):
self.close()
def n_ctx(self) -> int:
assert self.ctx is not None
return llama_cpp.llama_n_ctx(self.ctx)
def pooling_type(self) -> int:
assert self.ctx is not None
return llama_cpp.llama_pooling_type(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_add(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(
self.ctx,
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,
llama_cpp.byref(candidates.candidates),
last_tokens_data,
penalty_last_n,
penalty_repeat,
penalty_freq,
penalty_present,
)
def sample_softmax(self, candidates: "_LlamaTokenDataArray"):
assert self.ctx is not None
llama_cpp.llama_sample_softmax(
self.ctx,
llama_cpp.byref(candidates.candidates),
)
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, llama_cpp.byref(candidates.candidates), k, min_keep
)
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, llama_cpp.byref(candidates.candidates), p, min_keep
)
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, llama_cpp.byref(candidates.candidates), p, min_keep
)
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, llama_cpp.byref(candidates.candidates), z, min_keep
)
def sample_typical(
self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int
):
assert self.ctx is not None
llama_cpp.llama_sample_typical(
self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep
)
def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float):
assert self.ctx is not None
llama_cpp.llama_sample_temp(
self.ctx, llama_cpp.byref(candidates.candidates), temp
)
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,
llama_cpp.byref(candidates.candidates),
grammar.grammar,
)
def sample_token_mirostat(
self,
candidates: "_LlamaTokenDataArray",
tau: float,
eta: float,
m: int,
mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float],
) -> int:
assert self.ctx is not None
return llama_cpp.llama_sample_token_mirostat(
self.ctx,
llama_cpp.byref(candidates.candidates),
tau,
eta,
m,
mu,
)
def sample_token_mirostat_v2(
self,
candidates: "_LlamaTokenDataArray",
tau: float,
eta: float,
mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float],
) -> int:
assert self.ctx is not None
return llama_cpp.llama_sample_token_mirostat_v2(
self.ctx,
llama_cpp.byref(candidates.candidates),
tau,
eta,
mu,
)
def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int:
assert self.ctx is not None
return llama_cpp.llama_sample_token_greedy(
self.ctx,
llama_cpp.byref(candidates.candidates),
)
def sample_token(self, candidates: "_LlamaTokenDataArray") -> int:
assert self.ctx is not None
return llama_cpp.llama_sample_token(
self.ctx,
llama_cpp.byref(candidates.candidates),
)
# 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(grammar.grammar, self.ctx, 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:
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._exit_stack = ExitStack()
self.batch = None
self.batch = llama_cpp.llama_batch_init(
self._n_tokens, self.embd, self.n_seq_max
)
def free_batch():
if self.batch is None:
return
llama_cpp.llama_batch_free(self.batch)
self.batch = None
self._exit_stack.callback(free_batch)
def close(self):
self._exit_stack.close()
def __del__(self):
self.close()
def n_tokens(self) -> int:
assert self.batch is not None
return self.batch.n_tokens
def reset(self):
assert self.batch is not None
self.batch.n_tokens = 0
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
def add_sequence(self, batch: Sequence[int], seq_id: int, logits_all: bool):
assert self.batch is not None
n_tokens = len(batch)
n_tokens0 = self.batch.n_tokens
self.batch.n_tokens += n_tokens
for i in range(n_tokens):
j = n_tokens0 + i
self.batch.token[j] = batch[i]
self.batch.pos[j] = i
self.batch.seq_id[j][0] = seq_id
self.batch.n_seq_id[j] = 1
self.batch.logits[j] = 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.recarray(
(self.n_vocab,),
dtype=np.dtype(
[("id", np.intc), ("logit", np.single), ("p", np.single)], align=True
),
)
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) # type: ignore
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.sorted = False
self.candidates.size = self.n_vocab
# Python wrappers over common/common
def _tokenize(model: _LlamaModel, text: str, add_bos: bool, special: bool) -> list[int]:
assert model.model is not None
n_tokens = len(text) + 1 if add_bos else len(text)
result = (llama_cpp.llama_token * n_tokens)()
n_tokens = llama_cpp.llama_tokenize(
model.model,
text.encode("utf-8"),
len(text),
result,
n_tokens,
add_bos,
special,
)
if n_tokens < 0:
result = (llama_cpp.llama_token * -n_tokens)()
check = llama_cpp.llama_tokenize(
model.model,
text.encode("utf-8"),
len(text),
result,
len(result),
add_bos,
special,
)
if check != -n_tokens:
raise RuntimeError(f'Failed to tokenize: text="{text}" n_tokens={n_tokens}')
else:
result = result[:n_tokens]
return list(result)
def _token_to_piece(model: _LlamaModel, token: int, special: bool = False) -> str:
assert model.model is not None
result = (ctypes.c_char * 8)(0)
n_tokens = llama_cpp.llama_token_to_piece(
model.model, token, result, 0, len(result), special
)
if n_tokens < 0:
result = (ctypes.c_char * -n_tokens)(0)
check = llama_cpp.llama_token_to_piece(
model.model, token, result, 0, len(result), special
)
if check != -n_tokens:
raise RuntimeError(f"Failed to get piece: token={token}")
else:
result = result[:n_tokens]
return bytes(result).decode("utf-8")
def _detokenize_spm(model: _LlamaModel, tokens: List[int]) -> str:
bos_id = model.token_bos()
result = ""
for i, token in enumerate(tokens):
piece = _token_to_piece(model, token)
if (
(tokens[0] == bos_id and i == 1) or (tokens[0] != bos_id and i == 0)
) and piece[0] == " ":
piece = piece[1:]
result += piece
return result
def _detokenize_bpe(model: _LlamaModel, tokens: List[int]) -> str:
result = ""
for token in tokens:
piece = _token_to_piece(model, token)
result += piece
return result
def _should_add_bos(model: _LlamaModel) -> bool:
assert model.model is not None
add_bos = llama_cpp.llama_add_bos_token(model.model)
if add_bos != -1:
return add_bos != 0
else:
return llama_cpp.llama_vocab_type(model.model) == llama_cpp.LLAMA_VOCAB_TYPE_SPM
# Embedding functions
def _normalize_embedding(embedding):
norm = float(np.linalg.norm(embedding))
if norm == 0.0:
return embedding
return [v / norm for v in embedding]
# Python wrappers over common/sampling structs
@dataclass
class _LlamaSamplingParams:
n_prev: int = 64
n_probs: int = 0
top_k: int = 40
top_p: float = 0.95
min_p: float = 0.05
tfs_z: float = 1.00
typical_p: float = 1.00
temp: float = 0.80
penalty_last_n: int = 64
penalty_repeat: float = 1.0
penalty_freq: float = 0.00
penalty_present: float = 0.00
mirostat: int = 0
mirostat_tau: float = 5.00
mirostat_eta: float = 0.10
penalize_nl: bool = True
grammar: str = ""
cfg_negative_prompt: str = ""
cfg_scale: float = 1.00
logit_bias: dict[int, float] = field(default_factory=dict)
@dataclass
class _LlamaSamplingContext:
params: _LlamaSamplingParams = field(default_factory=_LlamaSamplingParams)
mirostat_mu: ctypes.c_float = field(default_factory=ctypes.c_float)
grammar: Optional[LlamaGrammar] = None
# NOTE: Missing parsed_grammar
prev: list[int] = field(default_factory=list)
cur: list[llama_cpp.llama_token_data] = field(default_factory=list)
def reset(self):
self.prev = []
self.cur = []
if self.grammar is not None:
self.grammar.reset()
def cp(self):
return _LlamaSamplingContext(
params=self.params,
mirostat_mu=self.mirostat_mu,
grammar=self.grammar,
prev=self.prev.copy(),
cur=self.cur.copy(),
)
def last(self) -> Optional[int]:
if len(self.prev) > 0:
return self.prev[-1]
else:
return None
def prev_str(self, ctx_main: _LlamaContext, n: int) -> str:
return ctx_main.model.detokenize(self.prev[-n:]).decode("utf-8")
def sample(
self,
ctx_main: _LlamaContext,
idx: int = 0,
logits_array: Optional[npt.NDArray[np.single]] = None,
):
n_vocab = ctx_main.model.n_vocab()
id: int = 0
if logits_array is None:
logits = ctx_main.get_logits_ith(idx)
logits_array = np.array(
ctypes.cast(logits, ctypes.POINTER(ctypes.c_float * n_vocab)).contents,
dtype=np.single,
)
# apply logit_bias
for token, logit_bias in self.params.logit_bias.items():
logits_array[token] += logit_bias
token_data_array = _LlamaTokenDataArray(
n_vocab=n_vocab
) # TODO: Only create this once
token_data_array.copy_logits(logits_array)
# apply penalties
if len(self.prev) > 0:
nl_token = ctx_main.model.token_nl()
nl_logit = logits_array[nl_token]
last_tokens = self.prev[-self.params.penalty_last_n :]
last_tokens_size = min(len(last_tokens), self.params.penalty_last_n)
if last_tokens_size > 0:
last_tokens_p = (llama_cpp.llama_token * len(last_tokens))(*last_tokens)
ctx_main.sample_repetition_penalties(
token_data_array,
last_tokens_p,
last_tokens_size,
self.params.penalty_repeat,
self.params.penalty_freq,
self.params.penalty_present,
)
if not self.params.penalize_nl:
token_data_array.candidates_data.logit[nl_token] = nl_logit
if self.grammar is not None:
ctx_main.sample_grammar(token_data_array, self.grammar)
if self.params.temp < 0:
ctx_main.sample_softmax(token_data_array)
id = token_data_array.candidates_data.id[0]
elif self.params.temp == 0:
id = ctx_main.sample_token_greedy(token_data_array)
else:
if self.params.mirostat == 1:
mirostat_m = 100
ctx_main.sample_temp(token_data_array, self.params.temp)
id = ctx_main.sample_token_mirostat(
token_data_array,
self.params.mirostat_tau,
self.params.mirostat_eta,
mirostat_m,
ctypes.pointer(self.mirostat_mu),
)
elif self.params.mirostat == 2:
ctx_main.sample_temp(token_data_array, self.params.temp)
id = ctx_main.sample_token_mirostat_v2(
token_data_array,
self.params.mirostat_tau,
self.params.mirostat_eta,
ctypes.pointer(self.mirostat_mu),
)
else:
min_keep = max(1, self.params.n_probs)
ctx_main.sample_top_k(
token_data_array, self.params.top_k, min_keep=min_keep
)
ctx_main.sample_tail_free(
token_data_array, self.params.tfs_z, min_keep=min_keep
)
ctx_main.sample_typical(
token_data_array, self.params.typical_p, min_keep=min_keep
)
ctx_main.sample_top_p(
token_data_array, self.params.top_p, min_keep=min_keep
)
ctx_main.sample_min_p(
token_data_array, self.params.min_p, min_keep=min_keep
)
ctx_main.sample_temp(token_data_array, self.params.temp)
id = ctx_main.sample_token(token_data_array)
return id
def accept(self, ctx_main: _LlamaContext, id: int, apply_grammar: bool):
if apply_grammar and self.grammar is not None:
ctx_main.grammar_accept_token(self.grammar, id)
self.prev.append(id)