app.py

# This code is based on Sanchit Gandhi's MusicGen-Streaming: https://huggingface.co/spaces/sanchit-gandhi/musicgen-streaming

from queue import Queue
from threading import Thread
import numpy as np
import torch
from transformers import MusicgenForConditionalGeneration, MusicgenProcessor, set_seed
import gradio as gr
import spaces

model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")
processor = MusicgenProcessor.from_pretrained("facebook/musicgen-small")
title = "AI Radio"

class MusicgenStreamer:
    def __init__(self, model, device=None, play_steps=10, stride=None, timeout=None):
        self.decoder, self.audio_encoder, self.generation_config = model.decoder, model.audio_encoder, model.generation_config
        self.device = device or model.device
        self.play_steps = play_steps
        self.stride = stride or np.prod(self.audio_encoder.config.upsampling_ratios) * (play_steps - self.decoder.num_codebooks) // 6
        self.token_cache, self.to_yield, self.audio_queue, self.timeout = None, 0, Queue(), timeout
        self.stop_signal = object()

    def apply_delay_pattern_mask(self, input_ids):
        _, mask = self.decoder.build_delay_pattern_mask(input_ids[:, :1], pad_token_id=self.generation_config.decoder_start_token_id, max_length=input_ids.shape[-1])
        input_ids = self.decoder.apply_delay_pattern_mask(input_ids, mask)
        input_ids = input_ids[input_ids != self.generation_config.pad_token_id].reshape(1, self.decoder.num_codebooks, -1)[None, ...]
        return self.audio_encoder.decode(input_ids.to(self.audio_encoder.device), audio_scales=[None]).audio_values[0, 0].cpu().float().numpy()

    def put(self, value):
        if value.shape[0] // self.decoder.num_codebooks > 1:
            raise ValueError("MusicgenStreamer only supports batch size 1")
        self.token_cache = torch.cat([self.token_cache, value[:, None]], dim=-1) if self.token_cache else value
        if self.token_cache.shape[-1] % self.play_steps == 0:
            audio_values = self.apply_delay_pattern_mask(self.token_cache)
            self.on_finalized_audio(audio_values[self.to_yield:-self.stride])
            self.to_yield += len(audio_values) - self.to_yield - self.stride

    def end(self):
        audio_values = self.apply_delay_pattern_mask(self.token_cache) if self.token_cache else np.zeros(self.to_yield)
        self.on_finalized_audio(audio_values[self.to_yield:], stream_end=True)

    def on_finalized_audio(self, audio, stream_end=False):
        self.audio_queue.put(audio, timeout=self.timeout)
        if stream_end:
            self.audio_queue.put(self.stop_signal, timeout=self.timeout)

    def __iter__(self):
        return self

    def __next__(self):
        value = self.audio_queue.get(timeout=self.timeout)
        if value is self.stop_signal:
            raise StopIteration()
        return value

@spaces.GPU()
def generate_audio(text_prompt, audio_length_in_s=10.0, play_steps_in_s=2.0, seed=0):
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    if device != model.device:
        model.to(device)
        if device == "cuda:0":
            model.half()
    max_new_tokens = int(model.audio_encoder.config.frame_rate * audio_length_in_s)
    play_steps = int(model.audio_encoder.config.frame_rate * play_steps_in_s)
    inputs = processor(text=text_prompt, padding=True, return_tensors="pt")
    streamer = MusicgenStreamer(model, device=device, play_steps=play_steps)
    Thread(target=model.generate, kwargs=dict(**inputs.to(device), streamer=streamer, max_new_tokens=max_new_tokens)).start()
    set_seed(seed)
    for new_audio in streamer:
        print(f"Sample of length: {round(new_audio.shape[0] / model.audio_encoder.config.sampling_rate, 2)} seconds")
        yield model.audio_encoder.config.sampling_rate, new_audio

demo = gr.Interface(
    fn=generate_audio,
    inputs=[
        gr.Text(label="Prompt", value="80s pop track with synth and instrumentals"),
        gr.Slider(10, 30, value=15, step=5, label="Audio length in seconds"),
        gr.Slider(0.5, 2.5, value=1.5, step=0.5, label="Streaming interval in seconds", info="Lower = shorter chunks, lower latency, more codec steps"),
        gr.Slider(0, 10, value=5, step=1, label="Seed for random generations"),
    ],
    outputs=[gr.Audio(label="Generated Music", streaming=True, autoplay=True)],
    title=title,
    cache_examples=False,
)

demo.queue().launch()