app.py

import os
import sys
import torch
import uuid

# 1. Force PyTorch to use a single thread to prevent Python 3.13 async loop collisions
torch.set_num_threads(1)

# Fix for PyTorch 2.6+ strict weights_only default for legacy checkpoints
_original_torch_load = torch.load
def trusted_torch_load(*args, **kwargs):
    if 'weights_only' not in kwargs:
        kwargs['weights_only'] = False
    return _original_torch_load(*args, **kwargs)
torch.load = trusted_torch_load

# 2. Clone the repository dynamically if it doesn't exist
if not os.path.exists('nuwave'):
    os.system('git clone https://github.com/mindslab-ai/nuwave.git')

# Add the cloned directory to the system path
sys.path.append('./nuwave')

import gradio as gr
import librosa
import soundfile as sf
from huggingface_hub import hf_hub_download
from lightning_model import NuWave

# 3. Force CPU Device Configuration
device = 'cpu'

ckpt_path = hf_hub_download(repo_id='nateraw/nu-wave-x2', filename='lit_model.ckpt')

# 🔥 Explicitly map the checkpoint structure directly to the CPU
model = NuWave.load_from_checkpoint(ckpt_path, train=False, map_location='cpu').to(device)
model.eval()

# 4. Inference Function
def upsample_audio(audio_path, steps, noise_schedule):
    if audio_path is None:
        return None
    
    # Load audio
    wav, sr = librosa.load(audio_path, sr=None)
    
    # Process on CPU
    wav_tensor = torch.FloatTensor(wav).to(device)
    
    with torch.no_grad():
        # Replace this with your model's exact generation call when ready:
        # e.g., upsampled_wav = model.generate(wav_tensor, steps=steps, schedule=noise_schedule)
        upsampled_wav = wav_tensor  # Placeholder
    
    output_wav = upsampled_wav.cpu().numpy()
    
    # Generate unique output file to prevent caching/file-locking errors between runs
    unique_filename = f"output_upsampled_{uuid.uuid4().hex[:8]}.wav"
    sf.write(unique_filename, output_wav, 48000) 
    
    return unique_filename

# 5. Build Gradio Web UI
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown(
        """
        # 🌊 NU-Wave X2: Audio Bandwidth Extension (CPU Mode)
        Upload a low-resolution audio file to upsample and reconstruct missing high-frequency details.
        
        *Note: Running on CPU will take longer to process higher step counts than a GPU Space.*
        """
    )
    
    with gr.Row():
        with gr.Column():
            input_audio = gr.Audio(
                type="filepath", 
                label="Input Low-Res Audio"
            )
            
            gr.Markdown("### ⚙️ Inference Fine-Tuning")
            
            steps_slider = gr.Slider(
                minimum=4, 
                maximum=50, # Lowered maximum recommended steps for CPU performance
                value=6, 
                step=1, 
                label="Inference Steps",
                info="⚡ **Impact:** Higher steps improve clarity but increase CPU processing time significantly."
            )
            
            schedule_dropdown = gr.Dropdown(
                choices=["linear", "fibonacci", "cosine"], 
                value="linear", 
                label="Noise Schedule"
            )
            
            submit_btn = gr.Button("Upsample Audio", variant="primary")
            
        with gr.Column():
            output_audio = gr.Audio(
                label="Upsampled High-Res Audio (48kHz)"
            )

    submit_btn.click(
        fn=upsample_audio,
        inputs=[input_audio, steps_slider, schedule_dropdown],
        outputs=output_audio
    )

if __name__ == "__main__":
    demo.launch()