-
Notifications
You must be signed in to change notification settings - Fork 38
/
ctransformers.py
219 lines (194 loc) · 7.33 KB
/
ctransformers.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
import warnings
from functools import cached_property, partial
from typing import Any, Generator, Iterable
from datasets.fingerprint import Hasher
from ..logging import logger as datadreamer_logger
from ..utils import ring_utils as ring
from ..utils.arg_utils import AUTO, Default
from ..utils.background_utils import RunIfTimeout
from ..utils.fs_utils import safe_fn
from ..utils.hf_model_utils import get_model_prompt_template
from ..utils.import_utils import ignore_transformers_warnings
from .hf_transformers import HFTransformers
from .llm import DEFAULT_BATCH_SIZE, LLM
with ignore_transformers_warnings():
from ctransformers import AutoModelForCausalLM, AutoTokenizer
from ctransformers.transformers import CTransformersTokenizer
from transformers import PreTrainedModel, PreTrainedTokenizer
def _add_tokens_patched(*args, **kwargs):
# Makes ctransformers support transformers==4.34.0.dev0
return 0
CTransformersTokenizer._add_tokens = _add_tokens_patched
class CTransformers(HFTransformers):
def __init__(
self,
model_name: str,
model_type: None | str = None,
model_file: None | str = None,
max_context_length: None | int = None,
chat_prompt_template: None | str | Default = AUTO,
system_prompt: None | str | Default = AUTO,
revision: None | str = None,
threads: None | int = None,
gpu_layers: int = 0,
cache_folder_path: None | str = None,
**kwargs,
):
LLM.__init__(self, cache_folder_path=cache_folder_path)
self.model_name = model_name
self.model_type = model_type
self.model_file = model_file
self.max_context_length = max_context_length
if self.max_context_length is None:
warnings.warn(
"CTransformers may not provide an accurate model context length."
" Explicitly set it with CTransformers(..., max_context_length=)"
" to remove this warning.",
stacklevel=2,
)
self.chat_prompt_template, self.system_prompt = get_model_prompt_template(
model_name=self.model_name,
revision=revision,
chat_prompt_template=chat_prompt_template,
system_prompt=system_prompt,
)
self.revision = revision
self.threads = threads
self.gpu_layers = gpu_layers
self.kwargs = kwargs
@cached_property
def _is_encoder_decoder(self) -> bool:
return False
@cached_property
def model(self) -> PreTrainedModel:
# Load model
log_if_timeout = RunIfTimeout(
partial(
lambda self: self.get_logger(
key="model", log_level=datadreamer_logger.level
).info("Loading..."),
self,
),
timeout=10.0,
)
model = AutoModelForCausalLM.from_pretrained(
self.model_name,
revision=self.revision,
model_type=self.model_type,
model_file=self.model_file,
**self.kwargs,
gpu_layers=self.gpu_layers,
hf=True,
)
# Set threads
if self.threads is not None:
model._llm._config.threads = self.threads
# Finished loading
log_if_timeout.stop(
partial(
lambda self: self.get_logger(
key="model", log_level=datadreamer_logger.level
).info("Finished loading."),
self,
)
)
return model
@cached_property
def tokenizer(self) -> PreTrainedTokenizer:
return AutoTokenizer.from_pretrained(self.model)
@ring.lru(maxsize=128)
def get_max_context_length(self, max_new_tokens: int) -> int:
"""Gets the maximum context length for the model. When ``max_new_tokens`` is
greater than 0, the maximum number of tokens that can be used for the prompt
context is returned.
Args:
max_new_tokens: The maximum number of tokens that can be generated.
Returns:
The maximum context length.
"""
if self.max_context_length is None:
return self.model._llm.context_length - max_new_tokens
else:
return self.max_context_length - max_new_tokens
def _is_batch_size_exception(self, e: BaseException) -> bool: # pragma: no cover
return False
def run(
self,
prompts: Iterable[str],
max_new_tokens: None | int = None,
temperature: float = 1.0,
top_p: float = 0.0,
n: int = 1,
stop: None | str | list[str] = None,
repetition_penalty: None | float = None,
logit_bias: None | dict[int, float] = None,
batch_size: int = DEFAULT_BATCH_SIZE,
batch_scheduler_buffer_size: None | int = None,
adaptive_batch_size: bool = False,
seed: None | int = None,
progress_interval: None | int = 60,
force: bool = False,
cache_only: bool = False,
verbose: None | bool = None,
log_level: None | int = None,
total_num_prompts: None | int = None,
return_generator: bool = False,
**kwargs,
) -> Generator[str | list[str], None, None] | list[str | list[str]]:
return super().run(
prompts=prompts,
max_new_tokens=max_new_tokens,
temperature=temperature,
top_p=top_p,
n=n,
stop=stop,
repetition_penalty=repetition_penalty,
logit_bias=logit_bias,
batch_size=batch_size,
batch_scheduler_buffer_size=batch_scheduler_buffer_size,
adaptive_batch_size=adaptive_batch_size,
seed=seed,
progress_interval=progress_interval,
force=force,
cache_only=cache_only,
verbose=verbose,
log_level=log_level,
total_num_inputs=total_num_prompts,
return_generator=return_generator,
**kwargs,
)
@property
def version(self) -> float:
return 1.0
@cached_property
def display_name(self) -> str:
if self.model_file is not None:
return (
LLM.display_name.func(self) + f" ({self.model_name}/{self.model_file})" # type: ignore[attr-defined]
)
else:
return LLM.display_name.func(self) + f" ({self.model_name})" # type: ignore[attr-defined]
@cached_property
def _cache_name(self) -> None | str:
names = [safe_fn(self.model_name, allow_slashes=False)]
if (
self.max_context_length is not None
and self.max_context_length != self.model._llm.context_length
):
names.append(str(self.max_context_length))
if self.revision:
names.append(self.revision)
to_hash: list[Any] = []
if self.model_type is not None:
to_hash.append(self.model_type)
if self.model_file is not None:
to_hash.append(self.model_file)
kwargs_filtered = {
key: self.kwargs[key] for key in ["local_files_only"] if key in self.kwargs
}
if len(kwargs_filtered) > 0:
to_hash.append(kwargs_filtered)
if len(to_hash) > 0: # pragma: no cover
names.append(Hasher.hash(to_hash))
return "_".join(names)
__all__ = ["CTransformers"]