app.py

from flask import Flask, request, jsonify, send_file
import spaces
import torch
import soundfile as sf
from huggingface_hub import login
from diffusers import StableAudioPipeline
import os
import io
import random

# Load Hugging Face token securely
HUGGINGFACE_TOKEN = os.getenv("HF_TOKEN")
if HUGGINGFACE_TOKEN is None:
    raise ValueError("Missing Hugging Face token. Please set it in Hugging Face Secrets.")
login(HUGGINGFACE_TOKEN)

# Set device for PyTorch (GPU or CPU)
device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if device == "cuda" else torch.float32

# Load the StableAudio model from Hugging Face Hub
pipe = StableAudioPipeline.from_pretrained("stabilityai/stable-audio-open-1.0", torch_dtype=torch_dtype)
pipe = pipe.to(device)

# Initialize Flask app
app = Flask(__name__)

# Route to generate audio
@app.route("/generate", methods=["GET"])
@spaces.GPU
def generate_audio():
    
    prompt = request.args.get("prompt")
    seed = request.args.get("seed", random.randint(0, 100000), type=int)
    
    if not prompt:
        return jsonify({"error": "Missing prompt parameter"}), 400

    try:
        
        # Load the StableAudio model from Hugging Face Hub
        #pipe = StableAudioPipeline.from_pretrained("stabilityai/stable-audio-open-1.0", torch_dtype=torch_dtype)
        #pipe = pipe.to(device)
        
        # Generate the audio using StableAudioPipeline
        generator = torch.Generator(device)
        generator.manual_seed(seed) 
        
        audio_output = pipe(
            prompt=prompt,
            negative_prompt='Low Quality',
            num_inference_steps=10,  # Number of diffusion steps
            guidance_scale=14.0,
            audio_end_in_s=1,
            num_waveforms_per_prompt=1,
            generator=generator
        ).audios
        
        # Convert audio to BytesIO in memory
        output_io = io.BytesIO()
        output_io.truncate(0) # Clears any residual data from previous calls
        output_audio = audio_output[0].T.float().cpu().numpy()
        sf.write(output_io, output_audio, pipe.vae.sampling_rate, format="WAV")  # Save as WAV or your preferred format
        output_io.seek(0)  # Reset buffer pointer to beginning

        # Send the file in response as attachment for download
        return send_file(output_io, as_attachment=False, download_name="output.wav", mimetype='audio/wav')
        
        
    except Exception as e:
        return jsonify({"error": str(e)}), 500

# Run the Flask app
if __name__ == "__main__":
    app.run(host="0.0.0.0", port=7860)