forked from abetlen/llama-cpp-python
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathllama_chat_format.py
More file actions
4001 lines (3639 loc) · 156 KB
/
llama_chat_format.py
File metadata and controls
4001 lines (3639 loc) · 156 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 os
import sys
import json
import ctypes
import dataclasses
import random
import string
import warnings
from datetime import datetime
from contextlib import ExitStack
from typing import (
Any,
Dict,
Iterator,
List,
Literal,
Optional,
Tuple,
Union,
Protocol,
cast,
)
import jinja2
from jinja2.sandbox import ImmutableSandboxedEnvironment
import numpy as np
import numpy.typing as npt
import llama_cpp.llama_cpp as llama_cpp
from llama_cpp import llama, llama_grammar, llama_types
from ._logger import logger
from ._utils import suppress_stdout_stderr, Singleton
### Common Chat Templates and Special Tokens ###
# Source: https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/blob/main/tokenizer_config.json
CHATML_CHAT_TEMPLATE = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
CHATML_BOS_TOKEN = "<s>"
CHATML_EOS_TOKEN = "<|im_end|>"
# Source: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1/blob/main/tokenizer_config.json
MISTRAL_INSTRUCT_CHAT_TEMPLATE = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token + ' ' }}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}"
MISTRAL_INSTRUCT_BOS_TOKEN = "<s>"
MISTRAL_INSTRUCT_EOS_TOKEN = "</s>"
# Source: https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1/blob/main/tokenizer_config.json
MIXTRAL_INSTRUCT_CHAT_TEMPLATE = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}"
# Source: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/blob/main/tokenizer_config.json
LLAMA3_INSTRUCT_CHAT_TEMPLATE = "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}"
### Chat Completion Handler ###
class LlamaChatCompletionHandler(Protocol):
"""Base Protocol for a llama chat completion handler.
Very generic protocol that can be used to implement any chat format.
The only hard requirement is that it must return a ChatCompletion when
stream=False and an iterator of ChatCompletionChunks when stream=True."""
def __call__(
self,
*,
# llama.cpp instance
llama: llama.Llama,
# openai api parameters
messages: List[llama_types.ChatCompletionRequestMessage],
functions: Optional[List[llama_types.ChatCompletionFunction]] = None,
function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None,
tools: Optional[List[llama_types.ChatCompletionTool]] = None,
tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None,
temperature: float = 0.2,
top_p: float = 0.95,
top_k: int = 40,
stream: bool = False,
stop: Optional[Union[str, List[str]]] = [],
seed: Optional[int] = None,
response_format: Optional[
llama_types.ChatCompletionRequestResponseFormat
] = None,
max_tokens: Optional[int] = None,
presence_penalty: float = 0.0,
frequency_penalty: float = 0.0,
repeat_penalty: float = 1.1,
model: Optional[str] = None,
logit_bias: Optional[Dict[str, float]] = None,
# llama.cpp parameters
min_p: float = 0.05,
typical_p: float = 1.0,
tfs_z: float = 1.0,
mirostat_mode: int = 0,
mirostat_tau: float = 5.0,
mirostat_eta: float = 0.1,
logits_processor: Optional[llama.LogitsProcessorList] = None,
grammar: Optional[llama.LlamaGrammar] = None,
logprobs: Optional[bool] = None,
top_logprobs: Optional[int] = None,
**kwargs, # type: ignore
) -> Union[
llama_types.CreateChatCompletionResponse,
Iterator[llama_types.CreateChatCompletionStreamResponse],
]: ...
class LlamaChatCompletionHandlerNotFoundException(Exception):
pass
class LlamaChatCompletionHandlerRegistry(Singleton):
_chat_handlers: Dict[str, LlamaChatCompletionHandler] = {}
def register_chat_completion_handler(
self,
name: str,
chat_handler: LlamaChatCompletionHandler,
overwrite: bool = False,
):
if not overwrite and name in self._chat_handlers:
raise ValueError(
f"Formatter with name '{name}' is already registered. Use `overwrite=True` to overwrite it."
)
self._chat_handlers[name] = chat_handler
def unregister_chat_handler(self, name: str):
if name in self._chat_handlers:
del self._chat_handlers[name]
else:
raise ValueError(f"No formatter registered under the name '{name}'.")
def get_chat_completion_handler_by_name(
self, name: str
) -> LlamaChatCompletionHandler:
try:
chat_handler = self._chat_handlers[name]
return chat_handler
except KeyError:
raise LlamaChatCompletionHandlerNotFoundException(
f"Invalid chat handler: {name} (valid formats: {list(self._chat_handlers.keys())})"
)
def get_chat_completion_handler(name: str) -> LlamaChatCompletionHandler:
return LlamaChatCompletionHandlerRegistry().get_chat_completion_handler_by_name(
name
)
def register_chat_completion_handler(name: str):
def decorator(f: LlamaChatCompletionHandler):
LlamaChatCompletionHandlerRegistry().register_chat_completion_handler(name, f)
return f
return decorator
### Chat Formatter ###
@dataclasses.dataclass
class ChatFormatterResponse:
"""Dataclass that stores completion parameters for a given chat format and
create_chat_completion request.
prompt contains the formatted prompt generated from the chat format and messages.
stop contains the stop token or list of stop tokens to use for the chat format."""
prompt: str
stop: Optional[Union[str, List[str]]] = None
stopping_criteria: Optional[llama.StoppingCriteriaList] = None
added_special: bool = False
class ChatFormatter(Protocol):
"""Base Protocol for a chat formatter. A chat formatter is a function that
takes a list of messages and returns a chat format response which can be used
to generate a completion. The response can also include a stop token or list
of stop tokens to use for the completion."""
def __call__(
self,
*,
messages: List[llama_types.ChatCompletionRequestMessage],
**kwargs: Any,
) -> ChatFormatterResponse: ...
class Jinja2ChatFormatter(ChatFormatter):
def __init__(
self,
template: str,
eos_token: str,
bos_token: str,
add_generation_prompt: bool = True,
stop_token_ids: Optional[List[int]] = None,
):
"""A chat formatter that uses jinja2 templates to format the prompt."""
self.template = template
self.eos_token = eos_token
self.bos_token = bos_token
self.add_generation_prompt = add_generation_prompt
self.stop_token_ids = (
set(stop_token_ids) if stop_token_ids is not None else None
)
self._environment = ImmutableSandboxedEnvironment(
loader=jinja2.BaseLoader(),
trim_blocks=True,
lstrip_blocks=True,
).from_string(self.template)
@staticmethod
def strftime_now(f: str) -> str:
return datetime.now().strftime(f)
def __call__(
self,
*,
messages: List[llama_types.ChatCompletionRequestMessage],
functions: Optional[List[llama_types.ChatCompletionFunction]] = None,
function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None,
tools: Optional[List[llama_types.ChatCompletionTool]] = None,
tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None,
**kwargs: Any,
) -> ChatFormatterResponse:
def raise_exception(message: str):
raise ValueError(message)
prompt = self._environment.render(
messages=messages,
eos_token=self.eos_token,
bos_token=self.bos_token,
raise_exception=raise_exception,
add_generation_prompt=self.add_generation_prompt,
functions=functions,
function_call=function_call,
tools=tools,
tool_choice=tool_choice,
strftime_now=self.strftime_now,
)
stopping_criteria = None
if self.stop_token_ids is not None:
def stop_on_last_token(
tokens: npt.NDArray[np.intc], logits: npt.NDArray[np.single]
) -> bool:
return tokens[-1] in self.stop_token_ids
stopping_criteria = llama.StoppingCriteriaList([stop_on_last_token])
return ChatFormatterResponse(
prompt=prompt,
stop=[self.eos_token],
stopping_criteria=stopping_criteria,
added_special=True,
)
def to_chat_handler(self) -> LlamaChatCompletionHandler:
return chat_formatter_to_chat_completion_handler(self)
def _convert_text_completion_logprobs_to_chat(
logprobs: Optional[llama_types.CompletionLogprobs],
) -> llama_types.ChatCompletionLogprobs:
if logprobs is None:
return None
return {
"content": [
{
"token": token,
"bytes": None,
"logprob": logprob,
"top_logprobs": [
{
"token": top_token,
"logprob": top_logprob,
"bytes": None,
}
for top_token, top_logprob in top_logprobs.items()
],
} for (token, logprob, top_logprobs) in zip(logprobs["tokens"], logprobs["token_logprobs"], logprobs["top_logprobs"])
],
"refusal": None,
}
def _convert_text_completion_to_chat(
completion: llama_types.Completion,
) -> llama_types.ChatCompletion:
assert "usage" in completion
return {
"id": "chat" + completion["id"],
"object": "chat.completion",
"created": completion["created"],
"model": completion["model"],
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": completion["choices"][0]["text"],
},
"logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]),
"finish_reason": completion["choices"][0]["finish_reason"],
}
],
"usage": completion["usage"],
}
def _convert_text_completion_chunks_to_chat(
chunks: Iterator[llama_types.CreateCompletionStreamResponse],
) -> Iterator[llama_types.ChatCompletionChunk]:
for i, chunk in enumerate(chunks):
if i == 0:
yield {
"id": "chat" + chunk["id"],
"model": chunk["model"],
"created": chunk["created"],
"object": "chat.completion.chunk",
"choices": [
{
"index": 0,
"delta": {
"role": "assistant",
},
"logprobs": None,
"finish_reason": None,
}
],
}
yield {
"id": "chat" + chunk["id"],
"model": chunk["model"],
"created": chunk["created"],
"object": "chat.completion.chunk",
"choices": [
{
"index": 0,
"delta": (
{
"content": chunk["choices"][0]["text"],
}
if chunk["choices"][0]["finish_reason"] is None
else {}
),
"logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]),
"finish_reason": chunk["choices"][0]["finish_reason"],
}
],
}
def _convert_completion_to_chat(
completion_or_chunks: Union[
llama_types.CreateCompletionResponse,
Iterator[llama_types.CreateCompletionStreamResponse],
],
stream: bool = False,
) -> Union[
llama_types.CreateChatCompletionResponse, Iterator[llama_types.ChatCompletionChunk]
]:
if stream:
chunks: Iterator[llama_types.CreateCompletionStreamResponse] = completion_or_chunks # type: ignore
return _convert_text_completion_chunks_to_chat(chunks)
else:
completion: llama_types.Completion = completion_or_chunks # type: ignore
return _convert_text_completion_to_chat(completion)
def _convert_completion_to_chat_function(
tool_name: str,
completion_or_chunks: Union[
llama_types.CreateCompletionResponse,
Iterator[llama_types.CreateCompletionStreamResponse],
],
stream: bool,
):
if not stream:
completion: llama_types.CreateCompletionResponse = completion_or_chunks # type: ignore
assert "usage" in completion
tool_id = "call_" + "_0_" + tool_name + "_" + completion["id"]
# TODO: Fix for legacy function calls
chat_completion: llama_types.CreateChatCompletionResponse = {
"id": "chat" + completion["id"],
"object": "chat.completion",
"created": completion["created"],
"model": completion["model"],
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": None,
"function_call": {
"name": tool_name,
"arguments": completion["choices"][0]["text"],
},
"tool_calls": [
{
"id": tool_id,
"type": "function",
"function": {
"name": tool_name,
"arguments": completion["choices"][0]["text"],
},
}
],
},
"logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]),
"finish_reason": "tool_calls",
}
],
"usage": completion["usage"],
}
return chat_completion
else:
chunks: Iterator[llama_types.CreateCompletionStreamResponse] = completion_or_chunks # type: ignore
def _stream_response_to_function_stream(
chunks: Iterator[llama_types.CreateCompletionStreamResponse],
) -> Iterator[llama_types.CreateChatCompletionStreamResponse]:
# blank first message
first = True
id_ = None
created = None
model = None
tool_id = None
for chunk in chunks:
if first:
id_ = "chat" + chunk["id"]
created = chunk["created"]
model = chunk["model"]
tool_id = "call_" + "_0_" + tool_name + "_" + chunk["id"]
yield {
"id": id_,
"object": "chat.completion.chunk",
"created": created,
"model": model,
"choices": [
{
"index": 0,
"finish_reason": None,
"logprobs": None,
"delta": {
"role": "assistant",
"content": None,
"function_call": None,
"tool_calls": None,
},
}
],
}
yield {
"id": "chat" + chunk["id"],
"object": "chat.completion.chunk",
"created": chunk["created"],
"model": chunk["model"],
"choices": [
{
"index": 0,
"finish_reason": None,
"logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]),
"delta": {
"role": None,
"content": None,
"function_call": {
"name": tool_name,
"arguments": chunk["choices"][0]["text"],
},
"tool_calls": [
{
"index": 0,
"id": tool_id,
"type": "function",
"function": {
"name": tool_name,
"arguments": chunk["choices"][0][
"text"
],
},
}
],
},
}
],
}
first = False
continue
assert tool_id is not None
yield {
"id": "chat" + chunk["id"],
"object": "chat.completion.chunk",
"created": chunk["created"],
"model": chunk["model"],
"choices": [
{
"index": 0,
"finish_reason": None,
"logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]),
"delta": {
"role": None,
"content": None,
"function_call": {
"name": tool_name,
"arguments": chunk["choices"][0]["text"],
},
"tool_calls": [
{
"index": 0,
"id": tool_id,
"type": "function",
"function": {
"name": tool_name,
"arguments": chunk["choices"][0]["text"],
},
}
],
},
}
],
}
if id_ is not None and created is not None and model is not None:
yield {
"id": id_,
"object": "chat.completion.chunk",
"created": created,
"model": model,
"choices": [
{
"index": 0,
"finish_reason": "tool_calls",
"logprobs": None,
"delta": {
"role": None,
"content": None,
"function_call": None,
"tool_calls": None,
},
}
],
}
return _stream_response_to_function_stream(chunks)
def chat_formatter_to_chat_completion_handler(
chat_formatter: ChatFormatter,
) -> LlamaChatCompletionHandler:
def chat_completion_handler(
*,
llama: llama.Llama,
messages: List[llama_types.ChatCompletionRequestMessage],
functions: Optional[List[llama_types.ChatCompletionFunction]] = None,
function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None,
tools: Optional[List[llama_types.ChatCompletionTool]] = None,
tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None,
temperature: float = 0.2,
top_p: float = 0.95,
top_k: int = 40,
min_p: float = 0.05,
typical_p: float = 1.0,
stream: bool = False,
stop: Optional[Union[str, List[str]]] = [],
seed: Optional[int] = None,
response_format: Optional[
llama_types.ChatCompletionRequestResponseFormat
] = None,
max_tokens: Optional[int] = None,
presence_penalty: float = 0.0,
frequency_penalty: float = 0.0,
repeat_penalty: float = 1.1,
tfs_z: float = 1.0,
mirostat_mode: int = 0,
mirostat_tau: float = 5.0,
mirostat_eta: float = 0.1,
model: Optional[str] = None,
logits_processor: Optional[llama.LogitsProcessorList] = None,
grammar: Optional[llama.LlamaGrammar] = None,
logit_bias: Optional[Dict[str, float]] = None,
logprobs: Optional[bool] = None,
top_logprobs: Optional[int] = None,
**kwargs, # type: ignore
) -> Union[
llama_types.CreateChatCompletionResponse,
Iterator[llama_types.CreateChatCompletionStreamResponse],
]:
result = chat_formatter(
messages=messages,
functions=functions,
function_call=function_call,
tools=tools,
tool_choice=tool_choice,
)
prompt = llama.tokenize(
result.prompt.encode("utf-8"),
add_bos=not result.added_special,
special=True,
)
if result.stop is not None:
stop = [] if stop is None else [stop] if isinstance(stop, str) else stop
rstop = result.stop if isinstance(result.stop, list) else [result.stop]
stop = stop + rstop
stopping_criteria = None
if result.stopping_criteria is not None:
stopping_criteria = result.stopping_criteria
if response_format is not None and response_format["type"] == "json_object":
grammar = _grammar_for_response_format(
response_format, verbose=llama.verbose
)
# Convert legacy functions to tools
if functions is not None:
tools = [
{
"type": "function",
"function": function,
}
for function in functions
]
# Convert legacy function_call to tool_choice
if function_call is not None:
if isinstance(function_call, str) and (
function_call == "none" or function_call == "auto"
):
tool_choice = function_call
if isinstance(function_call, dict) and "name" in function_call:
tool_choice = {
"type": "function",
"function": {
"name": function_call["name"],
},
}
tool = None
if (
tool_choice is not None
and isinstance(tool_choice, dict)
and tools is not None
):
name = tool_choice["function"]["name"]
tool = next((t for t in tools if t["function"]["name"] == name), None)
if tool is None:
raise ValueError(f"Tool choice '{name}' not found in tools.")
schema = tool["function"]["parameters"]
try:
# create grammar from json schema
grammar = llama_grammar.LlamaGrammar.from_json_schema(
json.dumps(schema), verbose=llama.verbose
)
except Exception as e:
if llama.verbose:
print(str(e), file=sys.stderr)
grammar = llama_grammar.LlamaGrammar.from_string(
llama_grammar.JSON_GBNF, verbose=llama.verbose
)
completion_or_chunks = llama.create_completion(
prompt=prompt,
temperature=temperature,
top_p=top_p,
top_k=top_k,
min_p=min_p,
typical_p=typical_p,
logprobs=top_logprobs if logprobs else None,
stream=stream,
stop=stop,
seed=seed,
max_tokens=max_tokens,
presence_penalty=presence_penalty,
frequency_penalty=frequency_penalty,
repeat_penalty=repeat_penalty,
tfs_z=tfs_z,
mirostat_mode=mirostat_mode,
mirostat_tau=mirostat_tau,
mirostat_eta=mirostat_eta,
model=model,
logits_processor=logits_processor,
stopping_criteria=stopping_criteria,
grammar=grammar,
logit_bias=logit_bias,
)
if tool is not None:
tool_name = tool["function"]["name"]
return _convert_completion_to_chat_function(
tool_name, completion_or_chunks, stream
)
return _convert_completion_to_chat(completion_or_chunks, stream=stream)
return chat_completion_handler
def hf_autotokenizer_to_chat_formatter(
pretrained_model_name_or_path: Union[str, os.PathLike[str]]
) -> ChatFormatter:
# https://huggingface.co/docs/transformers/main/chat_templating
# https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1#instruction-format
# https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1/blob/main/tokenizer_config.json
from transformers import AutoTokenizer # type: ignore
tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path) # type: ignore
def format_autotokenizer(
messages: List[llama_types.ChatCompletionRequestMessage],
**kwargs: Any,
) -> ChatFormatterResponse:
tokenizer.use_default_system_prompt = False # type: ignore
prompt: str = tokenizer.apply_chat_template(messages, tokenize=False) # type: ignore
assert isinstance(prompt, str)
# Return formatted prompt and eos token by default
return ChatFormatterResponse(
prompt=prompt, stop=tokenizer.eos_token, added_special=True
)
return format_autotokenizer
def hf_autotokenizer_to_chat_completion_handler(
pretrained_model_name_or_path: Union[str, os.PathLike[str]]
) -> LlamaChatCompletionHandler:
chat_formatter = hf_autotokenizer_to_chat_formatter(pretrained_model_name_or_path)
return chat_formatter_to_chat_completion_handler(chat_formatter)
def hf_tokenizer_config_to_chat_formatter(
tokenizer_config: Dict[str, Any],
add_generation_prompt: bool = True,
) -> ChatFormatter:
assert isinstance(tokenizer_config, dict)
assert "chat_template" in tokenizer_config
assert isinstance(tokenizer_config["chat_template"], str)
chat_template = tokenizer_config["chat_template"]
assert "bos_token" in tokenizer_config
assert isinstance(tokenizer_config["bos_token"], str)
bos_token = tokenizer_config["bos_token"]
assert "eos_token" in tokenizer_config
assert isinstance(tokenizer_config["eos_token"], str)
eos_token = tokenizer_config["eos_token"]
env = ImmutableSandboxedEnvironment(
trim_blocks=True,
lstrip_blocks=True,
).from_string(chat_template)
def format_tokenizer_config(
messages: List[llama_types.ChatCompletionRequestMessage],
**kwargs: Any,
) -> ChatFormatterResponse:
# TODO: veryify this is correct
# Add a blank assistant message to the end of the messages to prompt the model to generate a response
if add_generation_prompt:
messages = [
*messages,
llama_types.ChatCompletionRequestAssistantMessage(
role="assistant", content=""
),
]
prompt = env.render(
messages=messages,
bos_token=bos_token,
eos_token=eos_token,
)
return ChatFormatterResponse(
prompt=prompt, stop=[eos_token, bos_token], added_special=True
)
return format_tokenizer_config
def hf_tokenizer_config_to_chat_completion_handler(
tokenizer_config: Dict[str, Any],
add_generation_prompt: bool = True,
) -> LlamaChatCompletionHandler:
chat_formatter = hf_tokenizer_config_to_chat_formatter(
tokenizer_config, add_generation_prompt=add_generation_prompt
)
return chat_formatter_to_chat_completion_handler(chat_formatter)
def guess_chat_format_from_gguf_metadata(metadata: Dict[str, str]) -> Optional[str]:
if "tokenizer.chat_template" not in metadata:
return None
if metadata["tokenizer.chat_template"] == CHATML_CHAT_TEMPLATE:
return "chatml"
if (
metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE
or metadata["tokenizer.chat_template"] == MIXTRAL_INSTRUCT_CHAT_TEMPLATE
):
return "mistral-instruct"
if metadata["tokenizer.chat_template"] == LLAMA3_INSTRUCT_CHAT_TEMPLATE:
return "llama-3"
return None
### Utility functions for formatting chat prompts ###
# TODO: Replace these with jinja2 templates
def _get_system_message(
messages: List[llama_types.ChatCompletionRequestMessage],
) -> str:
"""Get the first system message."""
for message in messages:
if message["role"] == "system":
return message["content"] or ""
return ""
def _map_roles(
messages: List[llama_types.ChatCompletionRequestMessage],
role_map: Dict[str, str],
) -> List[Tuple[str, Optional[str]]]:
"""Map the message roles."""
output: List[Tuple[str, Optional[str]]] = []
for message in messages:
role = message["role"]
if role in role_map:
content: str | None = (
message["content"] if isinstance(message["content"], str) else None
)
output.append((role_map[role], content))
return output
def _format_llama2(
system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str
) -> str:
"""Format the prompt with the llama2 style."""
seps = [sep, sep2]
ret = system_message + sep
for i, (role, message) in enumerate(messages):
if system_message and i == 0:
m = message or ""
ret += m + seps[i % 2]
elif message:
ret += role + message + " " + seps[i % 2]
else:
ret += role + " "
return ret
def _format_add_colon_single(
system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str
) -> str:
"""Format the prompt with the add-colon-single style."""
ret = system_message + sep
for role, message in messages:
if message:
ret += role + ": " + message + sep
else:
ret += role + ":"
return ret
def _format_add_colon_two(
system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str
) -> str:
"""Format the prompt with the add-colon-two style."""
seps = [sep, sep2]
ret = system_message + seps[0]
for i, (role, message) in enumerate(messages):
if message:
ret += role + ": " + message + seps[i % 2]
else:
ret += role + ":"
return ret
def _format_no_colon_single(
system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str
) -> str:
"""Format the prompt with the no-colon-single style."""
ret = system_message
for role, message in messages:
if message:
ret += role + message + sep
else:
ret += role
return ret
def _format_add_colon_space_single(
system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str
) -> str:
"""Format the prompt with the add-colon-space-single style."""
ret = system_message + sep
for role, message in messages:
if message:
ret += role + ": " + message + sep
else:
ret += role + ": " # must be end with a space
return ret
def _format_chatml(
system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str
) -> str:
"""Format the prompt with the chatml style."""
ret = "" if system_message == "" else system_message + sep + "\n"
for role, message in messages:
if message:
ret += role + "\n" + message + sep + "\n"
else:
ret += role + "\n"
return ret
def _format_chatglm3(
system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str
) -> str:
"""Format the prompt with the chatglm3 style."""
ret = ""
if system_message:
ret += system_message
for role, message in messages:
if message:
ret += role + "\n" + " " + message
else:
ret += role
return ret
def _grammar_for_json(verbose: bool = False):
return llama_grammar.LlamaGrammar.from_string(
llama_grammar.JSON_GBNF, verbose=verbose
)
def _grammar_for_json_schema(
schema: str, verbose: bool = False, fallback_to_json: bool = True
):
try:
return llama_grammar.LlamaGrammar.from_json_schema(schema, verbose=verbose)
except Exception as e:
if fallback_to_json:
return _grammar_for_json(verbose=verbose)
else:
raise e
def _grammar_for_response_format(
response_format: llama_types.ChatCompletionRequestResponseFormat,
verbose: bool = False,
):
if response_format["type"] != "json_object":
return None
if "schema" in response_format:
return _grammar_for_json_schema(
json.dumps(response_format["schema"]), verbose=verbose
)
else:
return _grammar_for_json(verbose=verbose)
### Chat Formats ###
def register_chat_format(name: str):
def decorator(f: ChatFormatter):
chat_completion_handler = chat_formatter_to_chat_completion_handler(f)
LlamaChatCompletionHandlerRegistry().register_chat_completion_handler(
name, chat_completion_handler
)
return f
return decorator
# see https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/tokenization_llama.py
# system prompt is "embedded" in the first message
@register_chat_format("llama-2")
def format_llama2(
messages: List[llama_types.ChatCompletionRequestMessage],
**kwargs: Any,
) -> ChatFormatterResponse:
_system_template = "[INST] <<SYS>>\n{system_message}\n<</SYS>>"
_roles = dict(user="<s>[INST]", assistant="[/INST]")
_messages = _map_roles(messages, _roles)
system_message = _get_system_message(messages)
if system_message:
system_message = _system_template.format(system_message=system_message)
_prompt = _format_llama2(system_message, _messages, " ", "</s>") + "[/INST]"