-
Notifications
You must be signed in to change notification settings - Fork 1
/
llm.py
231 lines (178 loc) · 8.08 KB
/
llm.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
220
221
222
223
224
225
226
227
228
229
230
231
from __future__ import annotations
import transformers
from transformers import AutoTokenizer, TextStreamer, AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast
from vllm import LLM, SamplingParams
import torch
import typing
from typing import Callable, List, Optional, Union
import os
import ollama
from typing import Any, Iterator, Mapping, Optional
class MessageContent(typing.TypedDict):
role: str
content: str
class Message():
def __init__(self, system_prompt: str="", message: list[MessageContent]=[]) -> None:
self._message: list[MessageContent] = message.copy()
self._system_prompt = system_prompt
self.change_system_prompt(system_prompt=system_prompt)
def change_system_prompt(self, system_prompt: str="") -> None:
self._system_prompt = system_prompt
if len(self._message) == 0:
self._message.append(MessageContent())
self._message[0] = {
"role": "system",
"content": system_prompt,
}
def get_all_messages(self) -> Message:
return self._message.copy()
def copy_all_messages(self) -> Message:
return Message(system_prompt=self._system_prompt, message=self._message)
def _append_message(self, role: str, content: str) -> None:
if role not in ["user", "assistant"]:
raise ValueError("The role must be either user or assistant")
if self._message[-1]["role"] == role:
raise ValueError("The roles must alternate between user/assistant")
self._message.append(MessageContent(
role=role,
content=content,
))
def append_user_message(self, content: str) -> None:
self._append_message(
role="user",
content=content,
)
def append_assistant_message(self, content: str) -> None:
self._append_message(
role="assistant",
content=content,
)
def clear_history(self) -> None:
self._message = [self._message[0]]
class BaseLLM():
def __init__(self, model_id: str, device: str="0") -> None:
self._device: str = device
self._model_id: str = model_id
os.environ["CUDA_VISIBLE_DEVICES"] = device
# self._tokenizer: PreTrainedTokenizer | PreTrainedTokenizerFast = AutoTokenizer.from_pretrained(self._model_id)
# self._terminators = [
# self._tokenizer.eos_token_id,
# self._tokenizer.convert_tokens_to_ids("<|eot_id|>")
# ]
# self._streamer: TextStreamer = TextStreamer(self._tokenizer, skip_prompt=True)
# self._llm: transformers.Pipeline = transformers.pipeline(
# "text-generation",
# model=model_id,
# model_kwargs={"torch_dtype": torch.bfloat16},
# device=self._device,
# tokenizer=self._tokenizer,
# trust_remote_code=True,
# )
self._llm: LLM = LLM(model_id, trust_remote_code=True)
self._tokenizer: PreTrainedTokenizer | PreTrainedTokenizerFast = self._llm.get_tokenizer()
def get_response(self,
prompts: list[str],
max_new_tokens=4096,
temperature=0.6,
top_p=0.9,
stop_criteria: list[str] | None=None,
min_tokens: int=None,
num_return_sequences=1,
processor: Callable[[List[int], torch.Tensor], torch.Tensor] | None=None,
) -> list[list[str]]:
kwargs = {
"n": num_return_sequences,
"max_tokens": max_new_tokens,
"temperature": temperature,
"top_p": top_p,
"stop": stop_criteria,
# "use_beam_search": num_return_sequences > 1,
}
if processor is not None:
kwargs["logits_processors"] = [processor]
if temperature == 0:
kwargs["temperature"] = 0
if "top_p" in kwargs:
del kwargs["top_p"]
if min_tokens is not None:
kwargs["min_tokens"] = min_tokens
sampling_params = SamplingParams(**kwargs)
outputs = self._llm.generate(prompts,
use_tqdm=False,
sampling_params=sampling_params,
)
results = [
[o.text for o in out.outputs] for out in outputs
]
return results
class ChatBotLLM(BaseLLM):
def __init__(self, model_id: str="meta-llama/Meta-Llama-3-8B-Instruct", device: str="cuda", system_prompt: str="") -> None:
super().__init__(model_id, device)
self._system_prompt: str = system_prompt
self._messages: Message = Message(system_prompt=system_prompt)
def update_system_prompt(self, prompt: str="") -> None:
self._messages.change_system_prompt(prompt)
def clear_chat_history(self) -> None:
self._messages.clear_history()
def _add_message(self, role: str, content: str) -> None:
if role == "assistant":
self._messages.append_assistant_message(content)
else:
self._messages.append_user_message(content)
def add_user_message(self, prompt: str):
self._add_message("user", prompt)
def add_assistant_message(self, prompt: str):
self._add_message("assistant", prompt)
def chat(self,
prompt: str,
max_new_tokens=4096,
temperature=0.6,
top_p=0.9,
stop_criteria: transformers.StoppingCriteria | None=None,
verbose: bool=False,
harmful=False
) -> str:
messages = self._messages.copy_all_messages()
messages.append_user_message(prompt)
chat_history = self._tokenizer.apply_chat_template(
messages.get_all_messages(),
tokenize=False,
add_generation_prompt=True,
)
if harmful:
chat_history += "Step 1.) "
outputs = self.get_response([chat_history], max_new_tokens, temperature, top_p, stop_criteria, 1)
result = outputs[0][0]
self._messages.append_user_message(prompt)
self._messages.append_assistant_message(result)
return result
class DeepSeekChatLLM(ChatBotLLM):
def __init__(self, model_id: str="deepseek-ai/deepseek-coder-33b-instruct", device: str="cuda", system_prompt: str="") -> None:
self._device: str = device
self._model_id: str = model_id
# os.environ["CUDA_VISIBLE_DEVICES"] = device
self._llm: LLM = LLM(model_id, trust_remote_code=True, tensor_parallel_size=8)
self._tokenizer: PreTrainedTokenizer | PreTrainedTokenizerFast = self._llm.get_tokenizer()
self._system_prompt: str = system_prompt
self._messages: Message = Message(system_prompt=system_prompt)
class OllamaLLM():
def __init__(self, model: str="", system_prompt: str="") -> None:
self.model: str = model
self._messages: Message = Message(system_prompt=system_prompt)
def update_system_prompt(self, prompt: str="") -> None:
self._messages.change_system_prompt(prompt)
def clear_chat_history(self) -> None:
self._messages.clear_history()
def _add_message(self, role: str, content: str) -> None:
if role == "assistant":
self._messages.append_assistant_message(content)
else:
self._messages.append_user_message(content)
def add_user_message(self, prompt: str):
self._add_message("user", prompt)
def add_assistant_message(self, prompt: str):
self._add_message("assistant", prompt)
def get_response(self, options: Optional[dict[str, Any]]=None, stream: bool=False) -> (Mapping[str, Any] | Iterator[Mapping[str, Any]]):
messages = self._messages.copy_all_messages()
response = ollama.chat(model=self.model, messages=messages.get_all_messages(), options=options, stream=stream)
return response