handler.py

from audiocraft.models import MusicGen
from scipy.io.wavfile import write
import torch
import io

class EndpointHandler:
    def __init__(self, path=""):
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model = MusicGen.get_pretrained("facebook/musicgen-large").to(self.device)
        self.model.set_generation_params(duration=8)

    def __call__(self, data):
        prompt = data.get("inputs", "chill lofi music")
        wav = self.model.generate([prompt])  # internally uses the model device

        sample_rate = 32000
        audio_tensor = wav[0].to("cpu").numpy()  # ensure CPU before saving

        buffer = io.BytesIO()
        write(buffer, sample_rate, audio_tensor)
        buffer.seek(0)

        return {"audio": buffer.read()}