This repository has been archived on 2024-10-27. You can view files and clone it, but cannot push or open issues or pull requests.
local-llm-server/llm_server/llm/vllm/vllm_backend.py

54 lines
2.7 KiB
Python

from typing import Tuple
from flask import jsonify
from vllm import SamplingParams
from llm_server.database.log_to_db import log_to_db
from llm_server.llm.llm_backend import LLMBackend
class VLLMBackend(LLMBackend):
_default_params = vars(SamplingParams())
def handle_response(self, success, request, response_json_body, response_status_code, client_ip, token, prompt: str, elapsed_time, parameters, headers):
if len(response_json_body.get('text', [])):
# Does vllm return the prompt and the response together???
backend_response = response_json_body['text'][0].split(prompt)[1].strip(' ').strip('\n')
else:
# Failsafe
backend_response = ''
log_to_db(ip=client_ip, token=token, prompt=prompt, response=backend_response, gen_time=elapsed_time, parameters=parameters, headers=headers, backend_response_code=response_status_code, request_url=request.url,
response_tokens=response_json_body.get('details', {}).get('generated_tokens'), backend_url=self.backend_url)
return jsonify({'results': [{'text': backend_response}]}), 200
def get_parameters(self, parameters) -> Tuple[dict | None, str | None]:
try:
# top_k == -1 means disabled
top_k = parameters.get('top_k', self._default_params['top_k'])
if top_k <= 0:
top_k = -1
sampling_params = SamplingParams(
temperature=parameters.get('temperature', self._default_params['temperature']),
top_p=parameters.get('top_p', self._default_params['top_p']),
top_k=top_k,
use_beam_search=True if parameters.get('num_beams', 0) > 1 else False,
stop=list(set(parameters.get('stopping_strings') or parameters.get('stop', self._default_params['stop']))),
ignore_eos=parameters.get('ban_eos_token', False),
max_tokens=parameters.get('max_new_tokens') or parameters.get('max_tokens', self._default_params['max_tokens']),
presence_penalty=parameters.get('presence_penalty', self._default_params['presence_penalty']),
frequency_penalty=parameters.get('frequency_penalty', self._default_params['frequency_penalty']),
length_penalty=parameters.get('length_penalty', self._default_params['length_penalty']),
early_stopping=parameters.get('early_stopping', self._default_params['early_stopping'])
)
except ValueError as e:
return None, str(e).strip('.')
# We use max_new_tokens throughout the server.
result = vars(sampling_params)
result['max_new_tokens'] = result.pop('max_tokens')
return result, None