app.py

import gradio as gr
from transformers import pipeline

# Carrega o modelo na CPU
pipe = pipeline("text-generation", model="Qwen/Qwen2.5-0.5B-Instruct", device_map="cpu")

def predict(message, history):
    messages = [{"role": "user", "content": message}]
    # Gera a resposta
    results = pipe(messages, max_new_tokens=512)
    # Retorna apenas o texto da resposta
    return results[0]['generated_text'][-1]['content']

# O segredo está aqui: Definimos o nome do endpoint como "chat"
demo = gr.ChatInterface(fn=predict).queue()

if __name__ == "__main__":
    demo.launch()