app.py

import streamlit as st
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
import torch
import os

def initialize_model():
    """Initialize a small and fast model for CPU"""
    # Using a tiny model optimized for CPU
    model_id = "facebook/opt-125m"  # Much smaller model (125M parameters)
    model_id ="GEB-AGI/geb-1.3b"
    
    try:
        # Initialize the pipeline directly - more efficient than loading model separately
        pipe = pipeline(
            "text-generation",
            model=model_id,
            device_map="cpu",
            model_kwargs={"low_cpu_mem_usage": True}
        )
        tokenizer = AutoTokenizer.from_pretrained(model_id)
        
        return pipe, tokenizer
    except Exception as e:
        print(f"Error loading model: {str(e)}")
        raise e

def generate_response(pipe, tokenizer, prompt, conversation_history):
    """Generate model response"""
    try:
        # Format conversation context
        context = ""
        for turn in conversation_history[-3:]:  # Only use last 3 turns for efficiency
            context += f"Human: {turn['user']}\nAssistant: {turn['assistant']}\n"
        
        # Create the full prompt
        full_prompt = f"{context}Human: {prompt}\nAssistant:"
        
        # Generate response with conservative parameters
        response = pipe(
            full_prompt,
            max_new_tokens=50,  # Limit response length
            temperature=0.7,
            top_p=0.9,
            num_return_sequences=1,
            do_sample=True,
            pad_token_id=tokenizer.pad_token_id if tokenizer.pad_token_id else tokenizer.eos_token_id
        )[0]['generated_text']
        
        # Extract only the assistant's response
        try:
            assistant_response = response.split("Assistant:")[-1].strip()
            if not assistant_response:
                return "I apologize, but I couldn't generate a proper response."
            return assistant_response
        except:
            return response.split(prompt)[-1].strip()

    except Exception as e:
        return f"An error occurred: {str(e)}"

def main():
    st.set_page_config(page_title="LLM Chat Interface", page_icon="🤖")
    
    st.title("💬 Quick Chat Assistant")

    # Initialize session state
    if "chat_history" not in st.session_state:
        st.session_state.chat_history = []
    
    if "model_loaded" not in st.session_state:
        st.session_state.model_loaded = False

    # Initialize model (only once)
    if not st.session_state.model_loaded:
        with st.spinner("Loading the model... (this should take just a few seconds)"):
            try:
                pipe, tokenizer = initialize_model()
                st.session_state.pipe = pipe
                st.session_state.tokenizer = tokenizer
                st.session_state.model_loaded = True
            except Exception as e:
                st.error(f"Error loading model: {str(e)}")
                return

    # Display chat messages
    for message in st.session_state.chat_history:
        with st.chat_message("user"):
            st.write(message["user"])
        with st.chat_message("assistant"):
            st.write(message["assistant"])

    # Chat input
    if prompt := st.chat_input("Ask me anything!"):
        # Display user message
        with st.chat_message("user"):
            st.write(prompt)

        # Generate and display assistant response
        with st.chat_message("assistant"):
            with st.spinner("Thinking..."):
                current_turn = {"user": prompt, "assistant": ""}
                st.session_state.chat_history.append(current_turn)
                
                response = generate_response(
                    st.session_state.pipe,
                    st.session_state.tokenizer,
                    prompt,
                    st.session_state.chat_history
                )
                
                st.write(response)
                st.session_state.chat_history[-1]["assistant"] = response

        # Keep only last 5 turns
        if len(st.session_state.chat_history) > 5:
            st.session_state.chat_history = st.session_state.chat_history[-5:]

    # Sidebar
    with st.sidebar:
        if st.button("Clear Chat"):
            st.session_state.chat_history = []
            st.rerun()
        
        st.markdown("---")
        st.markdown("""
        ### Chat Info
        - Using OPT-125M model
        - Optimized for quick responses
        - Best for short conversations
        """)

if __name__ == "__main__":
    main()