Add mergekit-diff tool for model weight analysis

[juliensimon]

Hi @cg123

I built this while experimenting with Arcee Fusion. If you like it, great. If not, no worries 😃

Add mergekit-diff tool for model weight analysis

Summary

Adds mergekit-diff CLI tool for comparing weights between models with the same architecture.

Changes

• Created mergekit/scripts/diff.py with class-based architecture
• Added mergekit-diff entry point to pyproject.toml
• Updated README.md with tool description and TOC entry
• Supports weight difference analysis and KL divergence computation
• Multiple loading strategies: direct state dict, lazy loading, full loading
• Memory efficient for large models with proper GPU support

Technical Details

• Weight comparison: percentage of differing weights per layer
• KL divergence: histogram-based distributional comparison
• Three loading modes: LazyTensorLoader (most efficient), layer-by-layer, full model
• Type annotations, error handling, progress tracking
• CLI: mergekit-diff BASE_MODEL MODEL [--device cpu|cuda|mps] [--verbose] [--num-bins 100] [--no-lazy]

Use Cases

Model comparison after fine-tuning, merge analysis, quality assurance, research on weight space geometry.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[juliensimon]

I have read the CLA Document and I hereby sign the CLA