app.py

import os
from typing import Optional
from threading import Thread

import torch
import gradio as gr
from langchain.llms.base import LLM
from langchain.prompts import PromptTemplate
from langchain_community.vectorstores import Pinecone
from langchain.memory import ConversationBufferMemory
from langchain.chains import ConversationalRetrievalChain
from langchain_community.embeddings import HuggingFaceBgeEmbeddings
from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TextIteratorStreamer, pipeline


os.environ["TOKENIZERS_PARALLELISM"] = "false"

def initialize_model_and_tokenizer(model_name="mistralai/Mistral-7B-Instruct-v0.2"):   
    quantization_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.float16,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_use_double_quant=True,
    )
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        trust_remote_code=True,
        torch_dtype=torch.float16,
        device_map='auto',
        quantization_config=quantization_config
    )
    model.eval()
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    tokenizer.pad_token = tokenizer.eos_token
    return model, tokenizer

def init_chain(model, tokenizer, db, embed, temp, max_new_tokens, top_p, top_k, r_penalty):
    class CustomLLM(LLM):

        """Streamer Object"""

        streamer: Optional[TextIteratorStreamer] = None

        def _call(self, prompt, stop=None, run_manager=None) -> str:
            self.streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, Timeout=5)
            inputs = tokenizer(prompt, return_tensors="pt")
            input_ids = inputs["input_ids"].to('cuda')
            generate_kwargs = dict(
                temperature=float(temp),
                max_new_tokens=int(max_new_tokens),
                top_p=float(top_p),
                top_k=int(top_k),
                repetition_penalty=float(r_penalty),
                do_sample=True
                )
            kwargs = dict(input_ids=input_ids, streamer=self.streamer, **generate_kwargs)
            thread = Thread(target=model.generate, kwargs=kwargs)
            thread.start()
            return ""

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

    llm = CustomLLM()
    memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
    questionprompt = PromptTemplate.from_template(
        """<s>[INST] 
        Use the following pieces of context to answer the question at the end. 
        If you don't know the answer, just say that you don't know, don't try to make up an answer.
        Instead of they/them refer to Dmytro Kisil as he/him. 
        CONTEXT: {context} 
        CHAT HISTORY: {chat_history} 
        QUESTION: {question} 
        Helpful Answer: 
        [/INST]
        """
    )
    llm_chain = ConversationalRetrievalChain.from_llm(
      llm=llm,
      retriever=db.as_retriever(search_kwargs={"k": 5}),
      memory=memory,
      condense_question_prompt=questionprompt,
    )
    
    return llm_chain, llm

index_name = "resume-demo"

queries = [["Which masters degree Dmytro Kisil has?"],
           ["Which amount of salary does Dmytro Kisil is looking for?"],
           ["How long does Dmytro Kisil looking for a job?"],
           ["Why Dmytro Kisil moved to the Netherlands?"],
           ["When Dmytro Kisil left Ukraine?"],
           ["Where Dmytro Kisil live now?"],
           ["How many years of working experience in total does Dmytro Kisil have?"],
           ["How fast Dmytro Kisil can start working for my company?"]]

embed = HuggingFaceBgeEmbeddings(model_name='BAAI/bge-small-en-v1.5')

db = Pinecone.from_existing_index(index_name, embed)

model, tokenizer = initialize_model_and_tokenizer(model_name="mistralai/Mistral-7B-Instruct-v0.2")

with gr.Blocks() as demo:
    with gr.Column():
        gr.HTML("""
                <center>
                <h1>Ask about my work experience!<h1>
                </center>
                <center>
                Start typing your question or choose one of the examples below to start from
                </center>
                """)
        with gr.Row():
            gr.Markdown("[Resume]('https://drive.google.com/file/d/1OejkWuQKcjP73_uH6sfnj9u4hfmXQ-Oy/view?usp=sharing')")
            gr.Markdown("[LinkedIn]('https://www.linkedin.com/in/dmytro-kisil/')")
            gr.Markdown("[HuggingFace profile]('https://huggingface.co/Oysiyl')")
    with gr.Column():
        chatbot = gr.Chatbot()
        with gr.Row():
            msg = gr.Textbox(scale=9)
            submit_b = gr.Button("Submit", scale=1)
        with gr.Row():
            retry_b = gr.Button("Retry")
            undo_b = gr.Button("Undo")
            clear_b = gr.Button("Clear")
    examples = gr.Examples(queries, msg)
    with gr.Accordion("Additional options", open=False):
        temp = gr.Slider(
            label="Temperature",
            value=0.0001,
            minimum=0.0001,
            maximum=1.00,
            step=0.0001,
            interactive=True,
            info="Higher values produce more diverse outputs",
        )
        max_new_tokens = gr.Slider(
            label="Max new tokens",
            value=1024,
            minimum=64,
            maximum=8192,
            step=64,
            interactive=True,
            info="The maximum number of new tokens",
        )
        top_p = gr.Slider(
            label="Top-p (nucleus sampling)",
            value=0.95,
            minimum=0.00,
            maximum=1.00,
            step=0.01,
            interactive=True,
            info="Higher values sample more low-probability tokens",
        )
        top_k = gr.Slider(
            label="Top-k",
            value=40,
            minimum=0,
            maximum=100,
            step=1,
            interactive=True,
            info="select from top 0 tokens (because zero, relies on top_p)",
        )
        r_penalty = gr.Slider(
            label="Repetition penalty",
            value=1.15,
            minimum=1.0,
            maximum=2.0,
            step=0.01,
            interactive=True,
            info="Penalize repeated tokens",
        )

    def user(user_message, history):
        return "", history + [[user_message, None]]
    
    def undo(history):
        return history[:-1].copy()

    def retry(user_message, history):
        try:
            prev_user_message = history[-1][0]
        except:
            prev_user_message = ""
        return prev_user_message, history + [[prev_user_message, None]]

    def bot(history, temp, max_new_tokens, top_p, top_k, r_penalty):
        llm_chain, llm = init_chain(model, tokenizer, db, embed, temp, max_new_tokens, top_p, top_k, r_penalty)
        llm_chain.run(question=history[-1][0])
        history[-1][1] = ""
        for character in llm.streamer:
            history[-1][1] += character
            yield history
        llm_chain, llm = init_chain(model, tokenizer, db, embed, temp, max_new_tokens, top_p, top_k, r_penalty)  
    
    msg.submit(user, [msg, chatbot], [msg, chatbot], queue=False).then(bot, [chatbot, temp, max_new_tokens, top_p, top_k, r_penalty], chatbot)
    submit_b.click(user, [msg, chatbot], [msg, chatbot], queue=False).then(bot, [chatbot, temp, max_new_tokens, top_p, top_k, r_penalty], chatbot)
    retry_b.click(retry, [msg, chatbot], [msg, chatbot], queue=False).then(bot, [chatbot, temp, max_new_tokens, top_p, top_k, r_penalty], chatbot)
    clear_b.click(lambda: None, None, chatbot, queue=False)
    undo_b.click(undo, chatbot, chatbot, queue=False)
    
demo.queue()
demo.launch()