from typing import  Dict, List, Any
from transformers import pipeline
import soundfile as sf
import torch
import logging
import io

logger = logging.getLogger(__name__)

class EndpointHandler():
    def __init__(self, path=""):
        # load the optimized model
        # create inference pipeline
        self.pipeline = pipeline("text-to-audio", "facebook/musicgen-stereo-large", device="mps", torch_dtype=torch.float16)

    def generate_audio(self, text: str):
        # Here you can implement your audio generation logic
        # For demonstration purposes, let's use your existing code
        logger.info("Generating audio for text: %s", text)
        try:
            music = self.pipeline(text, forward_params={"max_new_tokens": 256})
            return music["audio"][0].T, music["sampling_rate"]
        except Exception as e:
            logger.error("Error generating audio for text: %s", text, exc_info=True)
            raise e

    def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
        """
        Args:
            data (:obj:):
                includes the input data and the parameters for the inference.
        Return:
            A :obj:`list`:. The object returned should be a list of one list like [[{"label": 0.9939950108528137}]] containing :
                - "label": A string representing what the label/class is. There can be multiple labels.
                - "score": A score between 0 and 1 describing how confident the model is for this label/class.
        """
        input = data.pop("input", data)

        audio_data, sampling_rate = self.generate_audio(input)

        
        with io.BytesIO() as buffer:
            sf.write(buffer, audio_data, sampling_rate, format="WAV")
            buffer.seek(0)
            return buffer.getvalue()