Spaces:
Sleeping
Sleeping
| title: CRDT-Merge Convergence Lab | |
| emoji: 🔬 | |
| colorFrom: gray | |
| colorTo: blue | |
| sdk: gradio | |
| sdk_version: 5.29.0 | |
| python_version: "3.12" | |
| app_file: app.py | |
| pinned: true | |
| license: other | |
| license_name: BUSL-1.1 | |
| license_link: https://github.com/mgillr/crdt-merge/blob/main/LICENSE | |
| tags: | |
| - crdt | |
| - model-merging | |
| - distributed-systems | |
| - convergence | |
| - neural-network | |
| - federated-learning | |
| short_description: "CRDT convergence proof: 100 nodes, 26 strategies" | |
| # CRDT-Merge Multi-Node Convergence Laboratory | |
| **Empirical proof that the two-layer CRDTMergeState architecture guarantees identical merged models across distributed nodes — regardless of merge ordering, network partitions, or strategy choice.** | |
| > **Patent** — UK Application No. 2607132.4, GB2608127.3 | **E4 Trust-Delta Architecture** | |
| > **Paper**: *Conflict-Free Replicated Data Types for Neural Network Model Merging* | |
| > **Library**: [crdt-merge](https://pypi.org/project/crdt-merge/) v0.9.5 | |
| ## Experiments | |
| ### 1. Multi-Node Convergence | |
| Simulates N distributed nodes (up to 100), each contributing a unique model tensor. Nodes merge in multiple random orderings. Verifies that **all orderings produce bitwise-identical Merkle roots and resolved tensors**. | |
| ### 2. Network Partition & Healing | |
| Splits nodes into isolated partitions. Each partition gossips internally and converges to its own state. Partitions are then healed (full gossip resumes). Verifies that **all nodes converge to the same state post-healing** — the core CRDT guarantee. | |
| ### 3. Cross-Strategy Sweep | |
| Tests **every merge strategy** (weight averaging, SLERP, TIES, DARE, Fisher, evolutionary, and 7+ more) for convergence on the same node set. Verifies that the two-layer architecture provides **universal CRDT compliance regardless of strategy**. | |
| ### 4. Scalability Benchmark | |
| Measures gossip and resolve overhead from 2 to 100 nodes. Confirms that the CRDT merge operation (set union on metadata) remains **sub-millisecond regardless of model size**, while resolve time scales linearly with contributions. | |
| ## Key Results | |
| | Metric | Result | | |
| |--------|--------| | |
| | Max nodes tested | 100 | | |
| | Strategies verified | 13/13 (no-base) | | |
| | Convergence rate | 100% across all orderings | | |
| | Partition healing | Always converges | | |
| | CRDT merge overhead | < 0.5ms | | |
| | Bitwise reproducibility | Guaranteed | | |
| | E4 trust convergence | 0.000 max divergence, 3.84ms | | |
| | E4 proof-carrying ops | 167K build/s, 101K verify/s | | |
| ## How It Works | |
| The two-layer architecture separates concerns: | |
| - **Layer 1 (CRDTMergeState)**: Manages a *set* of model contributions using OR-Set CRDT semantics. Merge = set union — trivially commutative, associative, idempotent. | |
| - **Layer 2 (Strategy)**: Applies any merge strategy as a deterministic pure function over the canonically-ordered contribution set. Same inputs → same outputs. | |
| Since Layer 1 guarantees all replicas converge to the same set of inputs, and Layer 2 is deterministic, **all replicas compute identical merged models**. | |
| ## Links | |
| - **GitHub**: [mgillr/crdt-merge](https://github.com/mgillr/crdt-merge) | |
| - **PyPI**: [crdt-merge](https://pypi.org/project/crdt-merge/) | |
| - **Architecture**: [CRDT_ARCHITECTURE.md](https://github.com/mgillr/crdt-merge/blob/main/docs/CRDT_ARCHITECTURE.md) | |
| --- | |
| Copyright 2026 Ryan Gillespie / Optitransfer | |