import torch import torchaudio from einops import rearrange import gradio as gr import spaces import os import uuid # Importing the model-related functions from stable_audio_tools import get_pretrained_model from stable_audio_tools.inference.generation import generate_diffusion_cond # Load the model outside of the GPU-decorated function def load_model(): print("Loading model...") model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0") print("Model loaded successfully.") return model, model_config # Function to set up, generate, and process the audio @spaces.GPU(duration=120) # Allocate GPU only when this function is called def generate_audio(prompt, seconds_total=30, steps=100, cfg_scale=7): print(f"Prompt received: {prompt}") print(f"Settings: Duration={seconds_total}s, Steps={steps}, CFG Scale={cfg_scale}") device = "cuda" if torch.cuda.is_available() else "cpu" print(f"Using device: {device}") # Fetch the Hugging Face token from the environment variable hf_token = os.getenv('HF_TOKEN') print(f"Hugging Face token: {hf_token}") # Use pre-loaded model and configuration model, model_config = load_model() sample_rate = model_config["sample_rate"] sample_size = model_config["sample_size"] print(f"Sample rate: {sample_rate}, Sample size: {sample_size}") model = model.to(device) print("Model moved to device.") # Set up text and timing conditioning conditioning = [{ "prompt": prompt, "seconds_start": 0, "seconds_total": seconds_total }] print(f"Conditioning: {conditioning}") # Generate stereo audio print("Generating audio...") output = generate_diffusion_cond( model, steps=steps, cfg_scale=cfg_scale, conditioning=conditioning, sample_size=sample_size, sigma_min=0.3, sigma_max=500, sampler_type="dpmpp-3m-sde", device=device ) print("Audio generated.") # Rearrange audio batch to a single sequence output = rearrange(output, "b d n -> d (b n)") print("Audio rearranged.") # Peak normalize, clip, convert to int16 output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu() print("Audio normalized and converted.") # Generate a unique filename for the output unique_filename = f"output_{uuid.uuid4().hex}.wav" print(f"Saving audio to file: {unique_filename}") # Save to file torchaudio.save(unique_filename, output, sample_rate) print(f"Audio saved: {unique_filename}") # Return the path to the generated audio file return unique_filename # Setting up the Gradio Interface interface = gr.Interface( fn=generate_audio, inputs=[ gr.Textbox(label="Prompt", placeholder="Enter your text prompt here"), gr.Slider(0, 47, value=30, label="Duration in Seconds"), gr.Slider(10, 150, value=100, step=10, label="Number of Diffusion Steps"), gr.Slider(1, 15, value=7, step=0.1, label="CFG Scale") ], outputs=gr.Audio(type="filepath", label="Generated Audio"), title="Pixio Audio", description="Generate variable-length stereo audio from text prompts using Pixio Audio 1.0 ." ) # Pre-load the model to avoid multiprocessing issues model, model_config = load_model() # Launch the Interface interface.launch()