Research map / Graph learning / Graph algorithms and clustering
MGit: A Model Versioning and Management System
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MGit, a git-like model versioning system that stores a lineage graph of model derivatives and uses it for storage compression (hashing plus quantized delta compression), automated update cascades, test traversals and merge-based collaboration.
Problem
Models derived from other models (fine-tuning, quantization, distillation, federated learning, MTL) form ecosystems that are costly to store and hard to debug, update and collaborate on, and existing systems (PyTorch, TensorFlow, HuggingFace) do not record dependencies between models.
Approach
A lineage graph records provenance and versioning edges between models with optional creation functions; this enables content-hash deduplication, delta compression of parent-child parameter differences with accuracy-checked acceptance, automated cascading re-creation of downstream models, and a merge primitive that classifies concurrent edits as conflict, possible conflict, or no conflict.
Key results
- Delta compression with LZMA reduces the lineage graph storage footprint by up to 7x (6.96x on G3 federated learning graph) with max accuracy delta of 0.11.
- Content-based hashing alone saves 9.4%, 16.5% and 79.6% storage on G1, G2 and G5 respectively.
- Automated graph construction correctly inserts 22 of 23 HuggingFace models relative to a gold lineage graph.
- Test bisection finds failing models up to 1.5x faster than linear search in the best case.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Compress model deltas with LZMA and content-based hashing instead of compressing full models | MGit storage · G1 | compression ratio: 1.83 -> 2.14 | Helped |
| Compress model deltas with LZMA and content-based hashing instead of compressing full models | MGit storage · G2 | compression ratio: 1.85 -> 5.35 | Helped |
| Compress model deltas with LZMA and content-based hashing instead of compressing full models | MGit storage · G3 | compression ratio: 2.29 -> 6.96 | Helped |
| Compress model deltas with LZMA and content-based hashing instead of compressing full models | MGit storage · G4 | compression ratio: 2.57 -> 2.57 | No clear effect |
| Use content-based hashing without delta compression (ablation) | MGit storage · G1–G5 | accuracy delta: 0.00 | No clear effect |
| Use LZMA rather than RLE for delta compression (ablation) | MGit storage · G1–G4 | compression ratio | Helped |
| Apply quantization and LZMA to full models rather than model deltas (ablation) | MGit storage · G1–G3 | compression ratio: 2.14 -> 1.83; 5.35 -> 1.85; 6.96 -> 2.29 | Hurt |
| Apply quantization and LZMA to full models rather than model deltas (ablation) | MGit storage · G4 | compression ratio: 2.57 -> 2.57 | No clear effect |
| Remove quantization from full-model compression (ablation) | MGit storage · G1–G4 | compression ratio: 1.83 -> 0.87; 1.85 -> 0.78; 2.29 -> 0.72; 2.57 -> 1.47 | Mixed |
| Add delta compression with quantization and LZMA to content-based hashing (ablation) | MGit storage · G1–G4 | compression ratio: 1.05 -> 2.14; 1.01 -> 5.35; 1.00 -> 6.96; 1.00 -> 2.57 | Helped |
Surprising findings
- Quantization plus LZMA on full models gives worse compression than delta compression except on the pruned edge-device graph G4, and RLE delta compression yields a max accuracy drop of 1.02 on G1.
Limitations
- Automated graph construction made one mis-insertion out of 23 models on G1.
- Higher compression ratios come with longer per-model runtimes; G1 tests run on CPU.
- Delta compression is lossy due to quantization and relies on accuracy tests to bound degradation; pairwise auto-insertion scales quadratically with graph size.
Experimental setup
Datasets: GLUE ImageNet-1K SQuAD 2.0 MNLI
Benchmarks: GLUE ImageNet-1K
Metrics: compression ratio accuracy delta per-model runtime insertion time test bisection speedup
Baselines: Full (quantization + LZMA on full models) Full w/o quantization MGit (Hash) only MGit (RLE + Hash)
Builds on: git ModelHub LZMA run-length encoding delta compression
Closest papers by meaning
- We Should Chart an Atlas of All the World's Models (NeurIPS 2025 · similarity 0.83)
- ModelLens: Finding the Best for Your Task from Myriads of Models (NeurIPS 2026 · similarity 0.82)
- $\texttt{Model-GLUE}$: Democratized LLM Scaling for A Large Model Zoo in the Wild (NeurIPS 2024 · similarity 0.82)
- BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing (ICML 2025 · similarity 0.82)
- RCCDA: Adaptive Model Updates in the Presence of Concept Drift under a Constrained Resource Budget (NeurIPS 2025 · similarity 0.82)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.