app.py

import os
import subprocess
import sys
from pathlib import Path

# --- 0. Hardcoded Toggle for Execution Environment ---
# Set this to True to use Hugging Face ZeroGPU (recommended)
# Set this to False to use the slower, pure CPU environment
USE_ZEROGPU = True 

# --- 1. Clone the VibeVoice Repository ---
repo_dir = "VibeVoice"
if not os.path.exists(repo_dir):
    print("Cloning the VibeVoice repository...")
    try:
        subprocess.run(
            ["git", "clone", "https://github.com/microsoft/VibeVoice.git"],
            check=True,
            capture_output=True,
            text=True
        )
        print("Repository cloned successfully.")
    except subprocess.CalledProcessError as e:
        print(f"Error cloning repository: {e.stderr}")
        sys.exit(1)
else:
    print("Repository already exists. Skipping clone.")

# --- 2. Install the VibeVoice Package ---
# Note: Other dependencies are installed via requirements.txt
os.chdir(repo_dir)
print(f"Changed directory to: {os.getcwd()}")

print("Installing the VibeVoice package in editable mode...")
try:
    subprocess.run(
        [sys.executable, "-m", "pip", "install", "-e", "."],
        check=True,
        capture_output=True,
        text=True
    )
    print("Package installed successfully.")
except subprocess.CalledProcessError as e:
    print(f"Error installing package: {e.stderr}")
    sys.exit(1)

# --- 3. Modify the demo script to be environment-aware ---
demo_script_path = Path("demo/gradio_demo.py")
print(f"Reading {demo_script_path} to apply environment-specific modifications...")

try:
    file_content = demo_script_path.read_text()

    # Define the original model loading block using a list of lines for robustness.
    # This avoids issues with indentation in multi-line string literals.
    original_lines = [
        '        self.model = VibeVoiceForConditionalGenerationInference.from_pretrained(',
        '            self.model_path,',
        '            torch_dtype=torch.bfloat16,',
        "            device_map='cuda',",
        '            attn_implementation="flash_attention_2",',
        '        )'
    ]
    original_block = "\n".join(original_lines)

    # Check if the block to be patched exists in the file
    if original_block not in file_content:
        print("\033[91mError: The original code block to be patched was not found.\033[0m")
        print("The demo script may have changed, or there might be a whitespace mismatch.")
        print("Please verify the contents of demo/gradio_demo.py.")
        sys.exit(1)

    if USE_ZEROGPU:
        print("Optimizing for ZeroGPU execution...")
        
        # New block for ZeroGPU: We remove the problematic `attn_implementation` line.
        replacement_lines_gpu = [
            '        self.model = VibeVoiceForConditionalGenerationInference.from_pretrained(',
            '            self.model_path,',
            '            torch_dtype=torch.bfloat16,',
            "            device_map='cuda',",
            '        )'
        ]
        replacement_block_gpu = "\n".join(replacement_lines_gpu)
        
        # Add 'import spaces' at the beginning of the file for the @spaces.GPU decorator
        if "import spaces" not in file_content:
             modified_content = "import spaces\n" + file_content
        else:
             modified_content = file_content
        
        # Decorate the main interface class to request a GPU from the Spaces infrastructure
        if "@spaces.GPU" not in modified_content:
            modified_content = modified_content.replace(
                "class VibeVoiceDemo:",
                "@spaces.GPU(duration=120)\nclass VibeVoiceDemo:"
            )
        
        # Replace the model loading block
        modified_content = modified_content.replace(original_block, replacement_block_gpu)
        print("Script modified for ZeroGPU successfully.")

    else: # Pure CPU execution
        print("Modifying for pure CPU execution...")
        
        # New block for CPU: Use float32 and map directly to the CPU.
        replacement_lines_cpu = [
            '        self.model = VibeVoiceForConditionalGenerationInference.from_pretrained(',
            '            self.model_path,',
            '            torch_dtype=torch.float32,  # Use float32 for CPU',
            '            device_map="cpu",',
            '        )'
        ]
        replacement_block_cpu = "\n".join(replacement_lines_cpu)
        
        # Replace the original model loading block with the CPU version
        modified_content = file_content.replace(original_block, replacement_block_cpu)
        print("Script modified for CPU successfully.")

    # Write the dynamically modified content back to the demo file
    demo_script_path.write_text(modified_content)

except Exception as e:
    print(f"An error occurred while modifying the script: {e}")
    sys.exit(1)

# --- 4. Launch the Gradio Demo ---
model_id = "microsoft/VibeVoice-1.5B"

# Construct the command to run the modified demo script
command = [
    "python",
    str(demo_script_path),
    "--model_path",
    model_id,
    "--share"
]

print(f"Launching Gradio demo with command: {' '.join(command)}")
subprocess.run(command)