app.py

import os
import torch
import soundfile as sf
from huggingface_hub import login
from diffusers import StableAudioPipeline
import gradio as gr
import spaces

# 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 Spaces Secrets.")
login(HUGGINGFACE_TOKEN)

# Set device for PyTorch (only CPU, if no GPU is available)
device = "cpu"
torch_dtype = torch.float32  # Use float32 for CPU by default

# Check for GPU availability
if torch.cuda.is_available():
    device = "cuda"
    torch_dtype = torch.float16  # Use float16 for GPU to optimize memory usage

# Load the pipeline
pipe = StableAudioPipeline.from_pretrained(
    "stabilityai/stable-audio-open-1.0",
    torch_dtype=torch_dtype
)
pipe = pipe.to(device)

# Function to generate audio
@spaces.GPU
def generate_audio(prompt, negative_prompt, duration, diffusion_steps, seed):
    generator = torch.Generator(device).manual_seed(seed)
    audio_output = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=int(diffusion_steps),  # Number of diffusion steps
        audio_end_in_s=duration,
        num_waveforms_per_prompt=1,
        generator=generator
    ).audios
    output_audio = audio_output[0].T.float().cpu().numpy()
    output_file = "output.wav"
    sf.write(output_file, output_audio, pipe.vae.sampling_rate)
    return output_file

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("## 🎧 Stable Audio Open - Audio Generation 🎼")
    gr.Markdown("### Adjust prompts, duration, and diffusion steps to control the generation!")
    
    # Input Section
    with gr.Row():
        prompt_input = gr.Textbox(label="Prompt", value="The sound of a hammer hitting a wooden surface.")
        negative_input = gr.Textbox(label="Negative Prompt", value="Low quality.")
    with gr.Row():
        duration_input = gr.Slider(minimum=1, maximum=10, step=0.5, value=1, label="Duration (seconds)")
        diffusion_steps_input = gr.Slider(minimum=1, maximum=500, step=10, value=10, label="Diffusion Steps")
    with gr.Row():
        seed_input = gr.Number(label="Random Seed", value=42)
    
    # Output Section
    generate_button = gr.Button("Generate Audio")
    output_audio = gr.Audio(label="Generated Audio", type="filepath")
    
    # Connect the function to the button click
    generate_button.click(
        generate_audio, 
        inputs=[prompt_input, negative_input, duration_input, diffusion_steps_input, seed_input], 
        outputs=output_audio
    )

# Launch the app
demo.launch()