models/llamaCustom.py

import os
import pickle
from json import dumps, loads
from typing import Any, List, Mapping, Optional

import numpy as np
import openai
import pandas as pd
import streamlit as st
from dotenv import load_dotenv
from huggingface_hub import HfFileSystem
from langchain.llms.base import LLM
from llama_index import (
    Document,
    GPTVectorStoreIndex,
    LLMPredictor,
    PromptHelper,
    ServiceContext,
    SimpleDirectoryReader,
    StorageContext,
    load_index_from_storage,
)
from llama_index.llms import CompletionResponse, CustomLLM, LLMMetadata
from llama_index.prompts import Prompt
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

from assets.prompts import custom_prompts

load_dotenv()
# openai.api_key = os.getenv("OPENAI_API_KEY")
fs = HfFileSystem()

# define prompt helper
# set maximum input size
CONTEXT_WINDOW = 2048
# set number of output tokens
NUM_OUTPUT = 525
# set maximum chunk overlap
CHUNK_OVERLAP_RATION = 0.2

text_qa_template = Prompt(custom_prompts.text_qa_template_str)

refine_template = Prompt(custom_prompts.refine_template_str)


@st.cache_resource
def load_model(model_name: str):
    # llm_model_name = "bigscience/bloom-560m"
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForCausalLM.from_pretrained(model_name, config="T5Config")

    pipe = pipeline(
        task="text-generation",
        model=model,
        tokenizer=tokenizer,
        # device=0, # GPU device number
        # max_length=512,
        do_sample=True,
        top_p=0.95,
        top_k=50,
        temperature=0.7,
    )

    return pipe


class OurLLM(CustomLLM):
    def __init__(self, model_name: str, model_pipeline):
        self.model_name = model_name
        self.pipeline = model_pipeline

    @property
    def metadata(self) -> LLMMetadata:
        """Get LLM metadata."""
        return LLMMetadata(
            context_window=CONTEXT_WINDOW,
            num_output=NUM_OUTPUT,
            model_name=self.model_name,
        )

    def complete(self, prompt: str, **kwargs: Any) -> CompletionResponse:
        prompt_length = len(prompt)
        response = self.pipeline(prompt, max_new_tokens=NUM_OUTPUT)[0]["generated_text"]

        # only return newly generated tokens
        text = response[prompt_length:]
        return CompletionResponse(text=text)

    def stream_complete(self, prompt: str, **kwargs: Any) -> CompletionResponse:
        raise NotImplementedError()


class LlamaCustom:
    def __init__(self, model_name: str) -> None:
        self.vector_index = self.initialize_index(model_name=model_name)

    def initialize_index(self, model_name: str):
        index_name = model_name.split("/")[-1]

        file_path = f"./vectorStores/{index_name}"

        if os.path.exists(path=file_path):
            # rebuild storage context
            storage_context = StorageContext.from_defaults(persist_dir=file_path)

            # local load index access
            index = load_index_from_storage(storage_context)

            # huggingface repo load access
            # with fs.open(file_path, "r") as file:
            #     index = pickle.loads(file.readlines())
            return index
        else:
            prompt_helper = PromptHelper(
                context_window=CONTEXT_WINDOW,
                num_output=NUM_OUTPUT,
                chunk_overlap_ratio=CHUNK_OVERLAP_RATION,
            )

            # define llm
            pipe = load_model(model_name=model_name)
            llm = OurLLM(model_name=model_name, model_pipeline=pipe)

            llm_predictor = LLMPredictor(llm=llm)
            service_context = ServiceContext.from_defaults(
                llm_predictor=llm_predictor, prompt_helper=prompt_helper
            )

            # documents = prepare_data(r"./assets/regItems.json")
            documents = SimpleDirectoryReader(input_dir="./assets/pdf").load_data()

            index = GPTVectorStoreIndex.from_documents(
                documents, service_context=service_context
            )

            # local write access
            index.storage_context.persist(file_path)

            # huggingface repo write access
            # with fs.open(file_path, "w") as file:
            #     file.write(pickle.dumps(index))
            return index

    def get_response(self, query_str):
        print("query_str: ", query_str)
        # query_engine = self.vector_index.as_query_engine()
        query_engine = self.vector_index.as_query_engine(
            text_qa_template=text_qa_template, refine_template=refine_template
        )
        response = query_engine.query(query_str)
        print("metadata: ", response.metadata)
        return str(response)