File size: 16,790 Bytes
256a159 |
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 |
import json
import os
import re
import time
from concurrent.futures import ThreadPoolExecutor
from threading import Lock
from typing import Dict, List, Optional, Union
import jieba
import requests
from opencompass.registry import MODELS
from opencompass.utils.prompt import PromptList
from .base_api import BaseAPIModel
PromptType = Union[PromptList, str]
OPENAI_API_BASE = 'https://api.openai.com/v1/chat/completions'
@MODELS.register_module()
class OpenAI(BaseAPIModel):
"""Model wrapper around OpenAI's models.
Args:
path (str): The name of OpenAI's model.
max_seq_len (int): The maximum allowed sequence length of a model.
Note that the length of prompt + generated tokens shall not exceed
this value. Defaults to 2048.
query_per_second (int): The maximum queries allowed per second
between two consecutive calls of the API. Defaults to 1.
retry (int): Number of retires if the API call fails. Defaults to 2.
key (str or List[str]): OpenAI key(s). In particular, when it
is set to "ENV", the key will be fetched from the environment
variable $OPENAI_API_KEY, as how openai defaults to be. If it's a
list, the keys will be used in round-robin manner. Defaults to
'ENV'.
org (str or List[str], optional): OpenAI organization(s). If not
specified, OpenAI uses the default organization bound to each API
key. If specified, the orgs will be posted with each request in
round-robin manner. Defaults to None.
meta_template (Dict, optional): The model's meta prompt
template if needed, in case the requirement of injecting or
wrapping of any meta instructions.
openai_api_base (str): The base url of OpenAI's API. Defaults to
'https://api.openai.com/v1/chat/completions'.
mode (str, optional): The method of input truncation when input length
exceeds max_seq_len. 'front','mid' and 'rear' represents the part
of input to truncate. Defaults to 'none'.
temperature (float, optional): What sampling temperature to use.
If not None, will override the temperature in the `generate()`
call. Defaults to None.
"""
is_api: bool = True
def __init__(self,
path: str = 'gpt-3.5-turbo',
max_seq_len: int = 4096,
query_per_second: int = 1,
rpm_verbose: bool = False,
retry: int = 2,
key: Union[str, List[str]] = 'ENV',
org: Optional[Union[str, List[str]]] = None,
meta_template: Optional[Dict] = None,
openai_api_base: str = OPENAI_API_BASE,
mode: str = 'none',
temperature: Optional[float] = None):
super().__init__(path=path,
max_seq_len=max_seq_len,
meta_template=meta_template,
query_per_second=query_per_second,
rpm_verbose=rpm_verbose,
retry=retry)
import tiktoken
self.tiktoken = tiktoken
self.temperature = temperature
assert mode in ['none', 'front', 'mid', 'rear']
self.mode = mode
if isinstance(key, str):
self.keys = [os.getenv('OPENAI_API_KEY') if key == 'ENV' else key]
else:
self.keys = key
# record invalid keys and skip them when requesting API
# - keys have insufficient_quota
self.invalid_keys = set()
self.key_ctr = 0
if isinstance(org, str):
self.orgs = [org]
else:
self.orgs = org
self.org_ctr = 0
self.url = openai_api_base
self.path = path
def generate(
self,
inputs: List[str or PromptList],
max_out_len: int = 512,
temperature: float = 0.7,
) -> List[str]:
"""Generate results given a list of inputs.
Args:
inputs (List[str or PromptList]): A list of strings or PromptDicts.
The PromptDict should be organized in OpenCompass'
API format.
max_out_len (int): The maximum length of the output.
temperature (float): What sampling temperature to use,
between 0 and 2. Higher values like 0.8 will make the output
more random, while lower values like 0.2 will make it more
focused and deterministic. Defaults to 0.7.
Returns:
List[str]: A list of generated strings.
"""
if self.temperature is not None:
temperature = self.temperature
with ThreadPoolExecutor() as executor:
results = list(
executor.map(self._generate, inputs,
[max_out_len] * len(inputs),
[temperature] * len(inputs)))
return results
def _generate(self, input: str or PromptList, max_out_len: int,
temperature: float) -> str:
"""Generate results given a list of inputs.
Args:
inputs (str or PromptList): A string or PromptDict.
The PromptDict should be organized in OpenCompass'
API format.
max_out_len (int): The maximum length of the output.
temperature (float): What sampling temperature to use,
between 0 and 2. Higher values like 0.8 will make the output
more random, while lower values like 0.2 will make it more
focused and deterministic.
Returns:
str: The generated string.
"""
assert isinstance(input, (str, PromptList))
# max num token for gpt-3.5-turbo is 4097
context_window = 4096
if '32k' in self.path:
context_window = 32768
elif '16k' in self.path:
context_window = 16384
elif 'gpt-4' in self.path:
context_window = 8192
# will leave 100 tokens as prompt buffer, triggered if input is str
if isinstance(input, str) and self.mode != 'none':
context_window = self.max_seq_len
input = self.bin_trim(input, context_window - 100 - max_out_len)
if isinstance(input, str):
messages = [{'role': 'user', 'content': input}]
else:
messages = []
for item in input:
msg = {'content': item['prompt']}
if item['role'] == 'HUMAN':
msg['role'] = 'user'
elif item['role'] == 'BOT':
msg['role'] = 'assistant'
elif item['role'] == 'SYSTEM':
msg['role'] = 'system'
messages.append(msg)
# Hold out 100 tokens due to potential errors in tiktoken calculation
max_out_len = min(
max_out_len, context_window - self.get_token_len(str(input)) - 100)
if max_out_len <= 0:
return ''
max_num_retries = 0
while max_num_retries < self.retry:
self.wait()
with Lock():
if len(self.invalid_keys) == len(self.keys):
raise RuntimeError('All keys have insufficient quota.')
# find the next valid key
while True:
self.key_ctr += 1
if self.key_ctr == len(self.keys):
self.key_ctr = 0
if self.keys[self.key_ctr] not in self.invalid_keys:
break
key = self.keys[self.key_ctr]
header = {
'Authorization': f'Bearer {key}',
'content-type': 'application/json',
}
if self.orgs:
with Lock():
self.org_ctr += 1
if self.org_ctr == len(self.orgs):
self.org_ctr = 0
header['OpenAI-Organization'] = self.orgs[self.org_ctr]
try:
data = dict(
model=self.path,
messages=messages,
max_tokens=max_out_len,
n=1,
stop=None,
temperature=temperature,
)
raw_response = requests.post(self.url,
headers=header,
data=json.dumps(data))
except requests.ConnectionError:
self.logger.error('Got connection error, retrying...')
continue
try:
response = raw_response.json()
except requests.JSONDecodeError:
self.logger.error('JsonDecode error, got',
str(raw_response.content))
continue
try:
return response['choices'][0]['message']['content'].strip()
except KeyError:
if 'error' in response:
if response['error']['code'] == 'rate_limit_exceeded':
time.sleep(1)
continue
elif response['error']['code'] == 'insufficient_quota':
self.invalid_keys.add(key)
self.logger.warn(f'insufficient_quota key: {key}')
continue
self.logger.error('Find error message in response: ',
str(response['error']))
max_num_retries += 1
raise RuntimeError('Calling OpenAI failed after retrying for '
f'{max_num_retries} times. Check the logs for '
'details.')
def get_token_len(self, prompt: str) -> int:
"""Get lengths of the tokenized string. Only English and Chinese
characters are counted for now. Users are encouraged to override this
method if more accurate length is needed.
Args:
prompt (str): Input string.
Returns:
int: Length of the input tokens
"""
enc = self.tiktoken.encoding_for_model(self.path)
return len(enc.encode(prompt))
def bin_trim(self, prompt: str, num_token: int) -> str:
"""Get a suffix of prompt which is no longer than num_token tokens.
Args:
prompt (str): Input string.
num_token (int): The upper bound of token numbers.
Returns:
str: The trimmed prompt.
"""
token_len = self.get_token_len(prompt)
if token_len <= num_token:
return prompt
pattern = re.compile(r'[\u4e00-\u9fa5]')
if pattern.search(prompt):
words = list(jieba.cut(prompt, cut_all=False))
sep = ''
else:
words = prompt.split(' ')
sep = ' '
l, r = 1, len(words)
while l + 2 < r:
mid = (l + r) // 2
if self.mode == 'front':
cur_prompt = sep.join(words[-mid:])
elif self.mode == 'mid':
cur_prompt = sep.join(words[:mid]) + sep.join(words[-mid:])
elif self.mode == 'rear':
cur_prompt = sep.join(words[:mid])
if self.get_token_len(cur_prompt) <= num_token:
l = mid # noqa: E741
else:
r = mid
if self.mode == 'front':
prompt = sep.join(words[-l:])
elif self.mode == 'mid':
prompt = sep.join(words[:l]) + sep.join(words[-l:])
elif self.mode == 'rear':
prompt = sep.join(words[:l])
return prompt
class OpenAIAllesAPIN(OpenAI):
"""Model wrapper around OpenAI-AllesAPIN.
Args:
path (str): The name of OpenAI's model.
url (str): URL to AllesAPIN.
key (str): AllesAPIN key.
query_per_second (int): The maximum queries allowed per second
between two consecutive calls of the API. Defaults to 1.
max_seq_len (int): Unused here.
meta_template (Dict, optional): The model's meta prompt
template if needed, in case the requirement of injecting or
wrapping of any meta instructions.
retry (int): Number of retires if the API call fails. Defaults to 2.
"""
is_api: bool = True
def __init__(self,
path: str,
url: str,
key: str,
temperature: float = 1.0,
query_per_second: int = 1,
rpm_verbose: bool = False,
max_seq_len: int = 2048,
meta_template: Optional[Dict] = None,
retry: int = 2):
super().__init__(path=path,
max_seq_len=max_seq_len,
query_per_second=query_per_second,
rpm_verbose=rpm_verbose,
meta_template=meta_template,
retry=retry)
self.url = url
self.temperature = temperature
self.headers = {
'alles-apin-token': key,
'content-type': 'application/json',
}
def _generate(self, input: str or PromptList, max_out_len: int,
temperature: float) -> str:
"""Generate results given an input.
Args:
inputs (str or PromptList): A string or PromptDict.
The PromptDict should be organized in OpenCompass'
API format.
max_out_len (int): The maximum length of the output.
temperature (float): What sampling temperature to use,
between 0 and 2. Higher values like 0.8 will make the output
more random, while lower values like 0.2 will make it more
focused and deterministic.
Returns:
str: The generated string.
"""
assert isinstance(input, (str, PromptList))
if isinstance(input, str):
messages = [{'role': 'user', 'content': input}]
else:
messages = []
for item in input:
msg = {'content': item['prompt']}
if item['role'] == 'HUMAN':
msg['role'] = 'user'
elif item['role'] == 'BOT':
msg['role'] = 'assistant'
elif item['role'] == 'SYSTEM':
msg['role'] = 'system'
messages.append(msg)
# model can be response with user and system
# when it comes with agent involved.
assert msg['role'] in ['user', 'system']
data = {
'model': self.path,
'messages': messages,
'temperature': temperature
}
for _ in range(self.retry):
self.wait()
raw_response = requests.post(self.url,
headers=self.headers,
data=json.dumps(data))
try:
response = raw_response.json()
except requests.JSONDecodeError:
self.logger.error('JsonDecode error, got',
str(raw_response.content))
time.sleep(1)
continue
if raw_response.status_code == 200 and response[
'msgCode'] == '10000':
data = response['data']
choices = data['choices']
if choices is None:
self.logger.error(data)
else:
return choices[0]['message']['content'].strip()
try:
match = re.match(r'Error code: \d+ - (.*)', response['data'])
err = eval(match.group(1))['error']
if err['code'] == 'content_filter' and err['status'] == 400:
return err['message']
except Exception:
pass
self.logger.error(response['msg'])
self.logger.error(response)
time.sleep(1)
raise RuntimeError('API call failed.')
def get_token_len(self, prompt: str) -> int:
"""Get lengths of the tokenized string. Only English and Chinese
characters are counted for now. Users are encouraged to override this
method if more accurate length is needed.
Args:
prompt (str): Input string.
Returns:
int: Length of the input tokens
"""
enc = self.tiktoken.encoding_for_model(self.path)
return len(enc.encode(prompt))
|