local-llm-server/llm_server/llm/llm_backend.py

from typing import Tuple, Union

import flask

from llm_server import opts
from llm_server.database import tokenizer


class LLMBackend:
    _default_params: dict

    def handle_response(self, success, request: flask.Request, response_json_body: dict, response_status_code: int, client_ip, token, prompt, elapsed_time, parameters, headers):
        raise NotImplementedError

    def validate_params(self, params_dict: dict) -> Tuple[bool, str | None]:
        raise NotImplementedError

    # def get_model_info(self) -> Tuple[dict | bool, Exception | None]:
    #     raise NotImplementedError

    def get_parameters(self, parameters) -> Tuple[dict | None, str | None]:
        """
        Validate and return the parameters for this backend.
        Lets you set defaults for specific backends.
        :param parameters:
        :return:
        """
        raise NotImplementedError

    def validate_request(self, parameters: dict) -> Tuple[bool, Union[str, None]]:
        raise NotImplementedError

    @staticmethod
    def validate_prompt(prompt: str) -> Tuple[bool, Union[str, None]]:
        prompt_len = len(tokenizer.encode(prompt))
        if prompt_len > opts.context_size - 10:  # Our tokenizer isn't 100% accurate so we cut it down a bit. TODO: add a tokenizer endpoint to VLLM
            return False, f'Token indices sequence length is longer than the specified maximum sequence length for this model ({prompt_len} > {opts.context_size}). Please lower your context size'
        return True, None