19 lines
475 B
Python
19 lines
475 B
Python
|
import traceback
|
||
|
|
|
||
|
|
import requests
|
||
|
|
import tiktoken
|
||
|
|
|
||
|
|
from llm_server import opts
|
||
|
|
|
||
|
|
tokenizer = tiktoken.get_encoding("cl100k_base")
|
||
|
|
|
||
|
|
|
||
|
|
def tokenize(prompt: str) -> int:
|
||
|
|
try:
|
||
|
|
r = requests.post(f'{opts.backend_url}/tokenize', json={'input': prompt}, verify=opts.verify_ssl, timeout=opts.backend_generate_request_timeout)
|
||
|
|
j = r.json()
|
||
|
|
return j['length']
|
||
|
|
except:
|
||
|
|
print(traceback.format_exc())
|
||
|
|
return len(tokenizer.encode(prompt)) + 10
|