handler.py

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch
import json


model_name = "jla25/squareV3"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)


class EndpointHandler:
    def __init__(self, model_dir):
        self.tokenizer = AutoTokenizer.from_pretrained(model_dir)
        self.model = AutoModelForSeq2SeqLM.from_pretrained(model_dir)
        self.model.eval()

    def preprocess(self, data):
        if not isinstance(data, dict) or "inputs" not in data or data["inputs"] is None:
            raise ValueError("La entrada debe ser un diccionario con la clave 'inputs' y un valor válido.")

        # Prompt personalizado para guiar al modelo
        input_text = f"Generate a valid JSON capturing data from this text:{data['inputs']}"
        print(f"Prompt generado para el modelo: {input_text}")
        input_text = input_text.encode("utf-8").decode("utf-8")
        tokens = self.tokenizer(input_text, return_tensors="pt", truncation=True, padding="max_length", max_length=1024)
        return tokens

    def inference(self, tokens):
        generate_kwargs = {
            "max_length": 512,
            "num_beams": 5,
            "do_sample": False,
            "temperature": 0.3,
            "top_k": 50,
            "top_p": 0.8,
            "early_stopping": True,  # Añadir explicitamente esta configuración
            "repetition_penalty": 2.5
        }
        with torch.no_grad():
            outputs = self.model.generate(**tokens, **generate_kwargs)
        return outputs

    def clean_output(self, output):
        try:
            start_index = output.index("{")
            end_index = output.rindex("}") + 1
            return output[start_index:end_index]
        except ValueError:
            return output

    def postprocess(self, outputs):
        decoded_output = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
        cleaned_output = self.clean_output(decoded_output)

        # Imprimir siempre el texto generado para depuración
        print(f"Texto generado por el modelo: {decoded_output}")
        print(f"JSON limpiado: {cleaned_output}")

        return {"response": cleaned_output}

    def __call__(self, data):
        tokens = self.preprocess(data)
        outputs = self.inference(tokens)
        result = self.postprocess(outputs)
        return result


# Crear una instancia del handler
handler = EndpointHandler(model_name)