app.py

import gradio as gr
import copy
import time
import ctypes #to run on C api directly 
import llama_cpp
from llama_cpp import Llama
from huggingface_hub import hf_hub_download #load from huggingfaces 
from dotenv import load_dotenv
from PyPDF2 import PdfReader
from langchain.text_splitter import CharacterTextSplitter
from langchain.vectorstores import FAISS
from langchain.chat_models import ChatOpenAI
from langchain.embeddings import OpenAIEmbeddings, HuggingFaceInstructEmbeddings
from langchain.memory import ConversationBufferMemory
from langchain.chains import ConversationalRetrievalChain


llm = Llama(model_path= hf_hub_download(repo_id="TheBloke/Dolphin-Llama2-7B-GGML", filename="dolphin-llama2-7b.ggmlv3.q4_1.bin"), n_ctx=2048) #download model from hf/ n_ctx=2048 for high ccontext length

history = []

pre_prompt = " The user and the AI are having a conversation : <|endoftext|> \n "

def get_pdf_text(pdfs):
  text=""
  for pdf in pdfs:
    pdf_reader = PdfReader(pdf)
    for page in pdf_reader.pages:
      text+= page.extract_text()
  return text

def get_text_chunks(text):
  text_splitter = CharacterTextSplitter(separator="\n",
  chunk_size=1000, chunk_overlap = 200, length_function=len)
  chunks = text_splitter.split_text(text)
  return chunks

def get_vectorstore(text_chunks):
    embeddings = OpenAIEmbeddings()
#     embeddings = HuggingFaceInstructEmbeddings(model_name="hkunlp/instructor-xl")
    vectorstore = FAISS.from_texts(texts=text_chunks, embedding=embeddings)
    return vectorstore

def generate_text(input_text, history):
    print("history ",history)
    print("input ", input_text)
    temp =""
    if history == []:
        input_text_with_history = f"SYSTEM:{pre_prompt}"+ "\n" + f"USER: {input_text} " + "\n" +" ASSISTANT:"
    else:
        input_text_with_history = f"{history[-1][1]}"+ "\n"
        input_text_with_history += f"USER: {input_text}" + "\n" +" ASSISTANT:"
    print("new input", input_text_with_history)
    output = llm(input_text_with_history, max_tokens=1024, stop=["<|prompter|>", "<|endoftext|>", "<|endoftext|> \n","ASSISTANT:","USER:","SYSTEM:"], stream=True)
    for out in output:
     stream = copy.deepcopy(out)
     print(stream["choices"][0]["text"])
     temp += stream["choices"][0]["text"]
     yield temp


    history =["init",input_text_with_history]
        


demo = gr.ChatInterface(generate_text,
    title="LLM on CPU",
    description="Running LLM with https://github.com/abetlen/llama-cpp-python. btw the text streaming thing was the hardest thing to impliment",
    examples=["Hello", "Am I cool?", "Are tomatoes vegetables?"],
    cache_examples=True,
    retry_btn=None,
    undo_btn="Delete Previous",
    clear_btn="Clear",)
demo.queue(concurrency_count=1, max_size=5)
demo.launch()