utils/customLLM.py

from typing import Any, List, Mapping, Optional

from langchain.llms.base import LLM
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline


class CustomLLM(LLM):
    
    # Create the pipeline for question answering
    def __init__(self, model: AutoModelForCausalLM, tokenizer: AutoTokenizer):
        self.pipeline = pipeline(
        model=model,
        tokenizer=tokenizer,
        task="text-generation",
        # device=0, # GPU device number
        # max_length=512,
        do_sample=True,
        top_p=0.95,
        top_k=50,
        temperature=0.7
    )

    def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:
        prompt_length = len(prompt)
        response = self.pipeline(prompt, max_new_tokens=525)[0]["generated_text"]

        # only return newly generated tokens
        return response[prompt_length:]

    @property
    def _identifying_params(self) -> Mapping[str, Any]:
        return {"name_of_model": self.model_name}

    @property
    def _llm_type(self) -> str:
        return "custom"