app.py

# MusicGen + Gradio + GPT Demo App (CPU-Optimized with MCP Server)

import gradio as gr
import os
import numpy as np
import torch
from transformers import AutoProcessor, MusicgenForConditionalGeneration
from openai import OpenAI
import scipy.io.wavfile

# Force CPU device (no GPU required)
device = torch.device("cpu")

# Load MusicGen model onto CPU
model_name = "facebook/musicgen-small"
model = MusicgenForConditionalGeneration.from_pretrained(model_name).to(device)
processor = AutoProcessor.from_pretrained(model_name)

# Initialize OpenAI client (set OPENAI_API_KEY in HF Spaces Secrets)
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))

# Refine user prompt via GPT
def refine_prompt(user_input):
    completion = client.chat.completions.create(
        model="gpt-4",
        messages=[
            {"role": "system", "content": "You are a music assistant. Make the user's input more descriptive for an AI music generator."},
            {"role": "user", "content": user_input}
        ]
    )
    return completion.choices[0].message.content.strip()

# Generate music (shorter tokens for CPU speed)
def generate_music(prompt, max_new_tokens: int = 128):
    inputs = processor(text=[prompt], return_tensors="pt").to(device)
    audio_values = model.generate(**inputs, max_new_tokens=max_new_tokens)
    sampling_rate = model.config.audio_encoder.sampling_rate
    audio = audio_values[0].cpu().numpy()

    # Normalize to float32 in -1.0 to 1.0 range for Gradio
    audio = audio / np.max(np.abs(audio))
    audio = audio.astype(np.float32)

    # Prepare int16 version and ensure 1D for WAV
    int_audio = (audio * 32767).astype(np.int16)
    int_audio = np.squeeze(int_audio)
    if int_audio.ndim > 1:
        int_audio = int_audio[:, 0]

    # Save as .wav file (in /tmp for Spaces)
    scipy.io.wavfile.write("/tmp/output.wav", sampling_rate, int_audio)
    return sampling_rate, audio

# Combined Gradio function
def main(user_input, max_new_tokens):
    detailed_prompt = refine_prompt(user_input)
    sampling_rate, audio = generate_music(detailed_prompt, max_new_tokens)
    return detailed_prompt, (sampling_rate, audio), "/tmp/output.wav"

# Build Gradio UI
demo = gr.Blocks()
with demo:
    gr.Markdown("""# 🎵 AI Music Generator  
Enter a music idea or mood and get a short AI-generated track. (CPU mode)""")

    user_input = gr.Textbox(label="Describe the mood or style of music")
    max_tokens = gr.Slider(32, 256, value=128, step=32, label="Length (tokens) for CPU")
    generate_btn = gr.Button("Generate Music")

    refined_output = gr.Textbox(label="Enhanced Prompt by GPT")
    audio_output = gr.Audio(label="Generated Audio", type="numpy")
    download_wav = gr.File(label="Download .wav file")

    generate_btn.click(
        main,
        inputs=[user_input, max_tokens],
        outputs=[refined_output, audio_output, download_wav]
    )

# Launch with Gradio MCP Server
from gradio.mcp_server import MCPServer

if __name__ == "__main__":
    server = MCPServer(demo, host="0.0.0.0", port=7860)
    server.run()