-
Notifications
You must be signed in to change notification settings - Fork 0
/
llm_client.py
220 lines (182 loc) · 7.39 KB
/
llm_client.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
import math
import openai
from typing import List
from pydantic import BaseModel
class TokenProbInfo(BaseModel):
token_id: int
token: str
logprob: float
eos: bool
bytes: List[int]
class LlmClient:
def __init__(self,
base_url="http://localhost:6006/v1",
api_key="xxxx",
model_name="glm-4-9b-chat",
):
self.client = openai.AsyncClient(
api_key=api_key,
base_url=base_url
)
self.model_name = model_name
def trim_prob_dist(self, token_dist: List[TokenProbInfo], top_p: float):
token_dist = sorted(token_dist, key=lambda x: x.logprob, reverse=True)
accum_prob = 0
for i, token_prob in enumerate(token_dist):
accum_prob += math.exp(token_prob.logprob)
if accum_prob >= top_p:
break
if accum_prob < top_p:
print(f'warning: accum_prob < top_p, token_num={len(token_dist)} accum_prob={accum_prob}, top_p={top_p}')
return token_dist
return token_dist[:i + 1]
async def calc_next_token_dist(self,
message_list,
prefix: List[int],
top_p: float = 0.9,
temperature: float = 0.8,
):
for i in range(3):
try:
response = await self.client.chat.completions.create(
model=self.model_name,
messages=message_list,
temperature=temperature,
extra_body={
"force_answer_prefix_token_ids": prefix,
"output_log_prob_token_id": True,
},
max_tokens=1,
logprobs=True,
top_logprobs=100,
)
break
except Exception as e:
print(f'error on run {i}: {e}')
choices0 = response.choices[0]
# print(choices0.finish_reason)
token_dist = choices0.logprobs.content[0].top_logprobs
token_dist = [TokenProbInfo(
token_id=token_prob.token_id,
token=token_prob.token,
logprob=token_prob.logprob,
eos=token_prob.eos,
bytes=token_prob.bytes,
) for token_prob in token_dist]
trimmed_token_dist = self.trim_prob_dist(token_dist, top_p)
# print(trimmed_token_dist)
return {
'token_dist': trimmed_token_dist,
'all_token_dist': token_dist,
}
async def sample_trace_stream(self,
message_list,
prefix: List[int],
prefix_logprob: float = 0,
top_p: float = 0.9,
temperature: float = 0.8,
max_tokens=1000,
):
response = await self.client.chat.completions.create(
model=self.model_name,
messages=message_list,
temperature=temperature,
extra_body={
"force_answer_prefix_token_ids": prefix,
"output_log_prob_token_id": True,
},
max_tokens=max_tokens,
logprobs=True,
top_logprobs=50,
stream=True,
timeout=20,
)
accum_prefix = prefix[:]
accum_logprob = prefix_logprob
async with response:
async for message in response:
choice0 = message.choices[0]
if choice0.delta.content is None:
continue
delta_token = choice0.delta.content
logprobs_info = choice0.logprobs.content[0]
delta_token_id = logprobs_info.token_id
delta_token_logprob = logprobs_info.logprob
top_token_dist = logprobs_info.top_logprobs
token_dist = [TokenProbInfo(
token_id=token_prob.token_id,
token=token_prob.token,
logprob=token_prob.logprob,
eos=token_prob.eos,
bytes=token_prob.bytes,
) for token_prob in top_token_dist]
trimmed_token_dist = self.trim_prob_dist(token_dist, top_p)
delta_info = {
'prefix': accum_prefix[:],
'prefix_logprob': accum_logprob,
'token_dist': trimmed_token_dist,
'all_token_dist': token_dist,
'delta_token': delta_token,
'delta_token_id': delta_token_id,
'delta_token_logprob': delta_token_logprob,
'finished_reason': choice0.finish_reason,
}
out_of_top_p = delta_token_id not in [token_prob.token_id for token_prob in trimmed_token_dist]
if out_of_top_p:
delta_info['finished_reason'] = 'out_of_top_p'
yield delta_info
if out_of_top_p:
break
accum_prefix.append(delta_token_id)
accum_logprob += delta_token_logprob
async def sample_trace_to_end_stream(
self,
message_list,
prefix: List[int],
prefix_logprob: float = 0,
top_p: float = 0.9,
temperature: float = 0.8,
max_tokens=1000,
):
while True:
delta_info = None
async for delta_info in self.sample_trace_stream(
message_list,
prefix,
prefix_logprob,
top_p,
temperature,
max_tokens,
):
if delta_info['finished_reason'] == 'out_of_top_p':
delta_token_logprob = delta_info['delta_token_logprob']
pre_sum_prob = sum([math.exp(token_prob.logprob) for token_prob in delta_info['token_dist'] if token_prob.logprob > delta_token_logprob])
# print(f'drop out_of_top_p: {repr(delta_info["delta_token"])} {delta_info["delta_token_id"]} {delta_info["delta_token_logprob"]:.6f} {pre_sum_prob}')
break
yield delta_info
if delta_info['finished_reason'] == 'stop':
break
# print(f'continue from {delta_info["finished_reason"]} ...')
prefix = delta_info['prefix']
prefix_logprob = delta_info['prefix_logprob']
if __name__ == "__main__":
import asyncio
client = LlmClient()
prompt = '''
下列事件中,属于必然事件的是__
A. 任意数的绝对值都是正数
B. 两直线被第三条直线所截,同位角相等
C. 如果a、b都是实数,那么a+b=b+a
D. 抛掷1个均匀的骰子,出现6点朝上
'''.strip()
message_list = [
{"role": "user", "content": prompt},
]
prefix = [
]
async def test_main():
async for delta in client.sample_trace_to_end_stream(message_list, prefix, max_tokens=1000):
print(f'prefix={delta["prefix"][-5:]}, delta={delta["delta_token_id"]}, {repr(delta["delta_token"])}, {delta["delta_token_logprob"]:.6f}, {delta["finished_reason"]}, {delta["token_dist"]}')
asyncio.run(
test_main()
)