app.py

import gradio as gr
import torch
import numpy as np
from transformers import AutoProcessor, MusicgenForConditionalGeneration

# Load MusicGen model
processor = AutoProcessor.from_pretrained("facebook/musicgen-small")
model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

def generate_music(prompt, duration):
    inputs = processor(
        text=[prompt],
        padding=True,
        return_tensors="pt"
    ).to(device)

    # Generate audio - MusicGen expects `max_new_tokens` to be tied to duration frames
    # 50 audio frames = 1 second, so duration * 50 frames
    generated_audio = model.generate(
        **inputs,
        do_sample=True,
        guidance_scale=3,
        max_new_tokens=duration * 50  # 50 tokens/sec
    )

    # Convert tensor to numpy array
    audio = generated_audio[0].cpu().numpy()
    sampling_rate = 32000  # MusicGen small uses 32kHz output

    return (audio, sampling_rate)

# Gradio UI
demo = gr.Interface(
    fn=generate_music,
    inputs=[
        gr.Textbox(lines=2, label="Describe your vibe (e.g., 'lofi chill with rain and piano')"),
        gr.Slider(5, 30, value=10, step=5, label="Duration (seconds)")
    ],
    outputs=gr.Audio(label="Generated Music"),
    title="🎷 LoFiJazz Agent",
    description="AI agent that composes lo-fi and jazz-style music based on your vibe prompt."
)

demo.launch()