hao-ai-lab-FastVideo/local_scripts/README.md

Model Weight Comparison Tools

This directory contains tools for loading and comparing model weights between different formats.

Files

• weight_compare.py - Comprehensive model comparison tool
• simple_model_loader.py - Simple model inspection tool
• requirements.txt - Required dependencies
• README.md - This file

Installation

Install the required dependencies:

pip install -r requirements.txt

Usage

1. Simple Model Inspection

To quickly inspect the structure of both models:

python simple_model_loader.py

This will:

• Load both models and show their basic structure
• Display the number of parameters
• Show sample parameter keys and shapes
• Calculate total parameter count

2. Comprehensive Model Comparison

To perform a detailed comparison between the models:

python weight_compare.py

This will:

• Load both models
• Compare parameter names and structures
• Identify common and unique parameters
• Calculate differences between matching parameters
• Generate a detailed report saved to model_comparison_results.txt

Model Paths

The scripts are configured to load:

1. DMD PyTorch Model: /mnt/sharefs/users/hao.zhang/DMD/wan_bidirectional_dmd_from_scratch/2025-06-20-08-17-06.607828_seed1024/checkpoint_model_004800/model.pt

2. Wan2.1 Safetensors Model: ../Wan-AI/Wan2.1-T2V-1.3B-Diffusers/transformer/diffusion_pytorch_model.safetensors

Output

The comparison tool generates:

• Console output with summary statistics
• Detailed report file (model_comparison_results.txt) containing:
  • Parameter counts for each model
  • List of common and unique parameters
  • Top parameters with largest differences
  • Shape mismatches
  • Statistical analysis of differences

Features

• Flexible Loading: Handles different PyTorch checkpoint formats
• Memory Efficient: Uses CPU loading to avoid GPU memory issues
• Detailed Analysis: Provides comprehensive comparison metrics
• Error Handling: Graceful handling of missing files and loading errors
• Progress Reporting: Shows loading progress and intermediate results