handler.py

from audiocraft.data.audio_utils import convert_audio
from audiocraft.data.audio import audio_write
from audiocraft.models import MusicGen

from typing import Dict, List, Any
import logging

class EndpointHandler:
    def __init__(self, path=""):
        self.model = MusicGen.get_pretrained("musicgen-medium")

    def __call__(self, data: Dict[str, Any]) -> Dict[str, str]:
        """
        Args:
            data (:dict:):
                The payload with the text prompt and generation parameters.
        """
        # process input
        logging.info(f"data: {data}")
        inputs = data.pop("inputs", data)
        parameters = data.pop("parameters", None)
        self.model.set_generation_params(**parameters)

        outputs = self.model.generate(**inputs)

        # postprocess the prediction
        prediction = outputs[0].cpu().numpy().tolist()

        return [{"generated_audio": prediction}]