musicgen-small / handler.py
pbotsaris's picture
removed float16 types from tensor
224fe62
from typing import Dict, List, Any
from scipy.io import wavfile
from transformers import AutoProcessor, MusicgenForConditionalGeneration
import torch
import io
import base64
def create_params(params, fr):
# default
out = { "do_sample": True,
"guidance_scale": 3,
"max_new_tokens": 256
}
has_tokens = False
if params is None:
return out
if 'duration' in params:
out['max_new_tokens'] = params['duration'] * fr
has_tokens = True
for k, p in params.items():
if k in out:
if has_tokens and k == 'max_new_tokens':
continue
out[k] = p
return out
class EndpointHandler:
def __init__(self, path="pbotsaris/musicgen-small"):
self.processor = AutoProcessor.from_pretrained(path)
self.model = MusicgenForConditionalGeneration.from_pretrained(path)
self.model.to('cuda:0') #type: ignore
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"
)
params = create_params(params, self.model.config.audio_encoder.frame_rate) #type: ignore
outputs = self.model.generate(**inputs.to('cuda:0'), **params) #type: ignore
pred = outputs[0, 0].cpu().numpy()
sr = self.model.config.audio_encoder.sampling_rate #type: ignore
wav_buffer = io.BytesIO()
wavfile.write(wav_buffer, rate=sr, data=pred)
wav_data = wav_buffer.getvalue()
base64_encoded_wav = base64.b64encode(wav_data).decode('utf-8')
return [{"audio": base64_encoded_wav, "sr": sr}]
if __name__ == "__main__":
handler = EndpointHandler()