app.py

import streamlit as st
from functools import lru_cache
import requests

# Cache model loading to optimize performance
@lru_cache(maxsize=3)
def load_hf_model(model_name):
    # Use the Hugging Face Inference API directly
    api_url = f"https://api-inference.huggingface.co/models/deepseek-ai/{model_name}"
    return api_url

# Load all models at startup
MODELS = {
    "DeepSeek-R1-Distill-Qwen-32B": load_hf_model("DeepSeek-R1-Distill-Qwen-32B"),
    "DeepSeek-R1": load_hf_model("DeepSeek-R1"),
    "DeepSeek-R1-Zero": load_hf_model("DeepSeek-R1-Zero")
}

# --- Chatbot function ---
def chatbot(input_text, history, model_choice, system_message, max_new_tokens, temperature, top_p):
    history = history or []
    
    # Get the selected model API URL
    api_url = MODELS[model_choice]
    
    # Create payload for the model
    payload = {
        "inputs": {
            "messages": [{"role": "user", "content": input_text}],
            "system": system_message,
            "max_tokens": max_new_tokens,
            "temperature": temperature,
            "top_p": top_p
        }
    }
    
    # Run inference using the selected model
    try:
        headers = {"Authorization": f"Bearer {st.secrets['HUGGINGFACE_TOKEN']}"}
        response = requests.post(api_url, headers=headers, json=payload).json()
        
        # Handle the response format
        if isinstance(response, list) and len(response) > 0:
            # Assuming the response is a list of generated text
            assistant_response = response[0].get("generated_text", "No response generated.")
        elif isinstance(response, dict) and "generated_text" in response:
            # If the response is a dictionary with generated_text
            assistant_response = response["generated_text"]
        else:
            assistant_response = "Unexpected model response format."
    except Exception as e:
        assistant_response = f"Error: {str(e)}"
    
    # Append user and assistant messages to history
    history.append((input_text, assistant_response))
    
    return history

# --- Streamlit App ---
st.set_page_config(page_title="DeepSeek Chatbot", page_icon="🤖", layout="wide")

# Title and description
st.title("DeepSeek Chatbot")
st.markdown("""
    Created by [ruslanmv.com](https://ruslanmv.com/)  
    This is a demo of different DeepSeek models. Select a model, type your message, and click "Submit".  
    You can also adjust optional parameters like system message, max new tokens, temperature, and top-p.
""")

# Sidebar for model selection and parameters
with st.sidebar:
    st.header("Options")
    model_choice = st.radio(
        "Choose a Model",
        options=list(MODELS.keys()),
        index=0
    )
    st.header("Optional Parameters")
    system_message = st.text_area(
        "System Message",
        value="You are a friendly Chatbot created by ruslanmv.com",
        height=100
    )
    max_new_tokens = st.slider(
        "Max New Tokens",
        min_value=1,
        max_value=4000,
        value=200
    )
    temperature = st.slider(
        "Temperature",
        min_value=0.10,
        max_value=4.00,
        value=0.70
    )
    top_p = st.slider(
        "Top-p (nucleus sampling)",
        min_value=0.10,
        max_value=1.00,
        value=0.90
    )

# Initialize chat history
if "chat_history" not in st.session_state:
    st.session_state.chat_history = []

# Display chat history
for user_msg, assistant_msg in st.session_state.chat_history:
    with st.chat_message("user"):
        st.write(user_msg)
    with st.chat_message("assistant"):
        st.write(assistant_msg)

# Input box for user message
user_input = st.chat_input("Type your message here...")

# Handle user input
if user_input:
    # Add user message to chat history
    st.session_state.chat_history = chatbot(
        user_input,
        st.session_state.chat_history,
        model_choice,
        system_message,
        max_new_tokens,
        temperature,
        top_p
    )
    
    # Rerun to update the chat display
    st.rerun()