app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
from accelerate import init_empty_weights, infer_auto_device_map
import torch

# Nombre del modelo en Hugging Face
model_name = "bushai/sar-i-7b"

# Inicializa un modelo vacío para ahorrar memoria
with init_empty_weights():
    model = AutoModelForCausalLM.from_pretrained(model_name)

# Definir cómo distribuir el modelo entre CPU y GPU
device_map = infer_auto_device_map(model, max_memory={0: "14GB", "cpu": "2GB"})
model = AutoModelForCausalLM.from_pretrained(model_name, device_map=device_map)
tokenizer = AutoTokenizer.from_pretrained(model_name)

def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    # Tokenizar la entrada del usuario
    inputs = tokenizer(message, return_tensors="pt").to(model.device)

    # Generar la respuesta del modelo
    outputs = model.generate(inputs['input_ids'], max_new_tokens=max_tokens, temperature=temperature, top_p=top_p)
    
    # Decodificar la salida generada
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    return response

# Configurar la interfaz de Gradio
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="You are a friendly chatbot.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"),
    ],
)

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