handler.py

from typing import Dict, List, Any
from transformers import AutoProcessor, MusicgenForConditionalGeneration
import torch

class EndpointHandler:
 def __init__(self, path="pbotsaris/musicgen-small"):
 # load model and processor
 self.processor = AutoProcessor.from_pretrained(path)
 self.model = MusicgenForConditionalGeneration.from_pretrained(path, torch_dtype=torch.float16)
 self.model.to('cuda')

 def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
 """
 Args:
 data (:dict:):
 The payload with the text prompt and generation parameters.
 """

 inputs = data.pop("inputs", data)
 params = data.pop("parameters", None)

 inputs = self.processor(
 text=[inputs],
 padding=True,
 return_tensors="pt"
 ).to('cuda')

 if params is not None:
 with torch.cuda.amp.autocast():
 outputs = self.model.generate(**inputs, **params)
 else:
 with torch.cuda.amp.autocast():
 outputs = self.model.generate(**inputs)

 pred = outputs[0].cpu().numpy().tolist()

 return [{"audio": pred, "sr": self.model.config.sampling_rate}]



if __name__ == "__main__":
 handler = EndpointHandler()