app.py

import os
import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel, PeftConfig

MAX_NEW_TOKENS = 100
TEMPERATURE = 0.5
TOP_P = 0.95
TOP_K = 50
REPETITION_PENALTY = 1.05
SPECIAL_TOKEN = "->:"

HF_TOKEN = os.getenv('HF_TOKEN')

def load_model():
    base_model_id = "meta-llama/Llama-2-7b-hf"
    peft_model_id = "somosnlp-hackathon-2025/Llama-2-7b-hf-lora-refranes"

    config = PeftConfig.from_pretrained(peft_model_id)

    base_model = AutoModelForCausalLM.from_pretrained(
        base_model_id,
        torch_dtype="auto",
        device_map="auto",
        token=HF_TOKEN
    )

    model = PeftModel.from_pretrained(base_model, peft_model_id)

    tokenizer = AutoTokenizer.from_pretrained(base_model_id)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    return model, tokenizer

model = None
tokenizer = None

def generate_response(input_text, max_tokens, temperature, top_p, repetition_penalty):
    global model, tokenizer

    if model is None or tokenizer is None:
        model, tokenizer = load_model()

    inputs = tokenizer(input_text + SPECIAL_TOKEN, return_tensors="pt").to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_tokens,
            temperature=temperature,
            do_sample=True,
            top_p=top_p,
            top_k=TOP_K,
            repetition_penalty=repetition_penalty
        )

    full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)

    if SPECIAL_TOKEN in full_response:
        response_parts = full_response.split(SPECIAL_TOKEN, 1)
        if len(response_parts) > 1:
            return response_parts[1].strip()

    return full_response.strip()

def chat_interface(message, history, system_message, max_tokens, temperature, top_p, repetition_penalty):
    prompt = f"{message}"
    if system_message:
        prompt = f"{system_message}\n{message}"
        
    response = generate_response(
        prompt, 
        max_tokens, 
        temperature, 
        top_p, 
        repetition_penalty
    )
    return response

demo = gr.ChatInterface(
    chat_interface,
    title="Sabiduría Popular - Refranes",
    description="Esta aplicación explica el significado de refranes en español utilizando un modelo de lenguaje. Escribe un refrán y el modelo te explicará su significado.",
    examples=[
        ["A caballo regalado no le mires el diente"],
        ["Más vale pájaro en mano que ciento volando"],
        ["Quien a buen árbol se arrima, buena sombra le cobija"],
        ["No por mucho madrugar amanece más temprano"]
    ],
    additional_inputs=[
        gr.Textbox(
            value="Eres un experto en sabiduría popular española. Tu tarea es explicar el significado de refranes en español de manera clara y concisa.",
            label="System message"
        ),
        gr.Slider(
            minimum=1,
            maximum=500,
            value=MAX_NEW_TOKENS,
            step=1,
            label="Max new tokens"
        ),
        gr.Slider(
            minimum=0.1,
            maximum=2.0,
            value=TEMPERATURE,
            step=0.1,
            label="Temperature"
        ),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=TOP_P,
            step=0.05,
            label="Top-p (nucleus sampling)"
        ),
        gr.Slider(
            minimum=1.0,
            maximum=2.0,
            value=REPETITION_PENALTY,
            step=0.05,
            label="Repetition penalty"
        ),
    ],
    theme="soft"
)

if __name__ == "__main__":
    print("Iniciando la aplicación. El modelo se cargará con la primera consulta.")
    demo.launch()