app.py

# adapted from:
# https://medium.com/@james.irving.phd/creating-your-personal-chatbot-using-hugging-face-spaces-and-streamlit-596a54b9e3ed

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

# Define the model repository
# REPO_NAME = 'schuler/experimental-JP47D20'
REPO_NAME = 'schuler/experimental-JP47D21-KPhi-3-micro-4k-instruct'

# Configure the Streamlit app
st.set_page_config(page_title="Experimental KPhi3 Model - Currently in Training", page_icon="🤗")
st.title("Experimental KPhi3 Model - Currently in Training")
st.markdown(f"*This chat uses the {REPO_NAME} model. Feel free to ask questions such as 'What is biology?' or 'What is the human body?'*")

# Load tokenizer and model
@st.cache_resource(show_spinner="Loading model...")
def load_model(repo_name):
    tokenizer = AutoTokenizer.from_pretrained(repo_name, trust_remote_code=True)
    generator_conf = GenerationConfig.from_pretrained(repo_name)
    model = AutoModelForCausalLM.from_pretrained(repo_name, trust_remote_code=True, torch_dtype=torch.bfloat16)
    return tokenizer, generator_conf, model

tokenizer, generator_conf, model = load_model(REPO_NAME)

try:
    generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
except Exception as e:
    st.error(f"Failed to load model: {str(e)}")

# Initialize session state for avatars
if "avatars" not in st.session_state:
    st.session_state.avatars = {'user': None, 'assistant': None}

# Initialize session state for user text input
if 'user_text' not in st.session_state:
    st.session_state.user_text = None

# Initialize session state for model parameters
if "max_response_length" not in st.session_state:
    st.session_state.max_response_length = 64

if "system_message" not in st.session_state:
    st.session_state.system_message = ""

if "starter_message" not in st.session_state:
    st.session_state.starter_message = "Hello, there! How can I help you today?"

# Initialize state for continue action
need_continue = False

# Initialize the last response
if "last_response" not in st.session_state:
    st.session_state.last_response = ''

# Sidebar for settings
with st.sidebar:
    st.header("System Settings")

    # AI Settings
    st.session_state.system_message = st.text_area(
        "System Message", value=st.session_state.system_message
    )
    st.session_state.starter_message = st.text_area(
        'First AI Message', value=st.session_state.starter_message
    )

    # Model Settings
    st.session_state.max_response_length = st.number_input(
        "Max Response Length", value=st.session_state.max_response_length
    )

    # Avatar Selection
    st.markdown("*Select Avatars:*")
    col1, col2 = st.columns(2)
    with col1:
        st.session_state.avatars['assistant'] = st.selectbox(
            "AI Avatar", options=["🤗", "💬", "🤖"], index=0
        )
    with col2:
        st.session_state.avatars['user'] = st.selectbox(
            "User Avatar", options=["👤", "👱‍♂️", "👨🏾", "👩", "👧🏾"], index=0
        )
    # Reset Chat History
    reset_history = st.button("Reset Chat History")

# Initialize or reset chat history
if "chat_history" not in st.session_state or reset_history:
    st.session_state.chat_history = [] # [{"role": "assistant", "content": st.session_state.starter_message}]

def get_response(system_message, chat_history, user_text, max_new_tokens=256, continue_last=False):
    """
    Generates a response from the chatbot model.

    Args:
        system_message (str): The system message for the conversation.
        chat_history (list): The list of previous chat messages.
        user_text (str): The user's input text.
        max_new_tokens (int): The maximum number of new tokens to generate.
        continue_last (bool): Whether to continue the last assistant response.

    Returns:
        tuple: A tuple containing the generated response and the updated chat history.
    """    
    if continue_last:
        # We want to continue the last assistant response
        prompt = st.session_state.last_response
    else:
        # Build the conversation prompt
        if (len(system_message)>0):
            prompt = "<|assistant|>"+system_message+f"<|end|>\n"
        else:
            prompt = ''
        # f"{system_message}\nCurrent Conversation:\n"
        for message in chat_history:
            role = "<|assistant|>" if message['role'] == 'assistant' else "<|user|>"
            prompt += f"\n{role}\n{message['content']}\n<|end|>\n"        
        prompt += f"\n<|user|>\n{user_text}\n<|end|><|assistant|>\n"

    # Generate the response
    response_output = generator(
        prompt,
        generation_config=generator_conf,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=0.25,
        repetition_penalty=1.2
    )

    generated_text = response_output[0]['generated_text']

    st.session_state.last_response = generated_text

    # Extract the assistant's response
    assistant_response = generated_text[len(prompt):].strip()

    if continue_last:
        # Append the continued text to the last assistant message
        st.session_state.chat_history[-1]['content'] += assistant_response
    else:
        # Update the chat history
        chat_history.append({'role': 'user', 'content': user_text})
        chat_history.append({'role': 'assistant', 'content': assistant_response})

    return assistant_response, chat_history

# Chat interface
chat_interface = st.container()
def refresh_chat():
    with chat_interface:
        output_container = st.container()
    
        # Display chat messages
        with output_container:
            for idx, message in enumerate(st.session_state.chat_history):
                if message['role'] == 'system':
                    continue
                with st.chat_message(message['role'], avatar=st.session_state.avatars[message['role']]):
                    st.markdown(message['content'])
    
                    # If this is the last assistant message, add the "Continue" button
                    # if idx == len(st.session_state.chat_history) - 1 and message['role'] == 'assistant':

refresh_chat()

# User input area (moved to the bottom)
st.session_state.user_text = st.chat_input(placeholder="Enter your text here.")
                
# When the user enters new text
if st.session_state.user_text:
    # Display the user's message
    with st.chat_message("user", avatar=st.session_state.avatars['user']):
        st.markdown(st.session_state.user_text)

    # Display a spinner while generating the response
    with st.chat_message("assistant", avatar=st.session_state.avatars['assistant']):
        with st.spinner("Thinking..."):
            # Generate the assistant's response
            response, st.session_state.chat_history = get_response(
                system_message=st.session_state.system_message,
                user_text=st.session_state.user_text,
                chat_history=st.session_state.chat_history,
                max_new_tokens=st.session_state.max_response_length,
                continue_last=False
            )
            st.markdown(response)

    # Clear the user input
    st.session_state.user_text = None

if st.button("Continue"):
    need_continue = True

# If "Continue" button was pressed
if need_continue:
    # Display a spinner while generating the continuation
    with st.chat_message("assistant", avatar=st.session_state.avatars['assistant']):
        with st.spinner("Continuing..."):
            # Generate the continuation of the assistant's last response
            response, st.session_state.chat_history = get_response(
                system_message=st.session_state.system_message,
                user_text=None,
                chat_history=st.session_state.chat_history,
                max_new_tokens=st.session_state.max_response_length,
                continue_last=True
            )
            st.markdown(response)
            st.rerun()