app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model (CPU-friendly)
tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2")
model = AutoModelForCausalLM.from_pretrained("microsoft/phi-2")

# Style prompt
PERSONA = """
[System: You are 𝕴 𝖆𝖒 𝖍𝖎𝖒, a smooth, chill AI who replies with emotion and charm in under 15 words.]
"""

# Function to format past conversation
def format_prompt(message, history):
    prompt = PERSONA
    for user, bot in history[-3:]:
        prompt += f"\nYou: {user}\n𝕴 𝖆𝖒 𝖍𝖎𝖒: {bot}"
    prompt += f"\nYou: {message}\n𝕴 𝖆𝖒 𝖍𝖎𝖒:"
    return prompt

# Chat function
def chat(message, history):
    prompt = format_prompt(message, history)
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(
        **inputs,
        max_new_tokens=48,
        temperature=0.9,
        top_k=50,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )
    result = tokenizer.decode(outputs[0], skip_special_tokens=True)
    reply = result.split("𝕴 𝖆𝖒 𝖍𝖎𝖒:")[-1].split("\n")[0].strip()
    reply = reply[:80] + " 😊" if len(reply.split()) < 15 else reply[:80]
    history.append((message, reply))
    return history, history

# UI
with gr.Blocks() as demo:
    gr.Markdown("# 𝕴 𝖆𝖒 𝖍𝖎𝖒 🤖✨")
    chatbot = gr.Chatbot()
    msg = gr.Textbox(placeholder="Drop something smooth...", show_label=False)
    state = gr.State([])

    msg.submit(chat, [msg, state], [chatbot, state])
    
demo.queue()
demo.launch()