🤖 AI Summary
This work addresses the fundamental limitations of existing neural network model merging strategies—none of the 26 current approaches satisfy commutativity, associativity, and idempotence, thereby failing to prevent conflicts in distributed settings. To resolve this, the paper introduces CRDTMergeState, the first framework to apply Conflict-Free Replicated Data Type (CRDT) theory to model merging. It features a two-layer architecture: the first layer employs an OR-Set to enable conflict-free merging of contribution sets, while the second layer leverages Merkle-root-seeded randomness and canonical ordering to deterministically encapsulate any existing merge strategy. The approach is fully compatible with all prior methods without degrading downstream performance and guarantees strong eventual consistency. Extensive evaluation across tensors, a 7.24B-parameter model, and a 100-node network—including 104 tests and 43,368 property verifications—confirms correctness; CRDT overhead remains below 0.5 ms, achieving byte-level output consistency.
📝 Abstract
All 26 neural network merge strategies we tested including weight averaging, SLERP, TIES, DARE, Fisher merging, and evolutionary approaches -- fail the algebraic properties (commutativity, associativity, idempotency) required for conflict-free distributed operation. We prove that this failure is structural: normalisation-based merges cannot simultaneously satisfy all three properties. To resolve this, we present a two-layer architecture -- CRDTMergeState -- that wraps any merge strategy in a CRDT-compliant (Conflict-Free Replicated Data Type) layer. Layer 1 manages contributions via OR-Set CRDT semantics, where the merge operation is set union -- trivially commutative, associative, and idempotent. Layer 2 applies merge strategies as deterministic pure functions over a canonically-ordered contribution set, with randomness seeded from the Merkle root. We prove that this separation guarantees Strong Eventual Consistency: all replicas receiving the same contributions compute identical merged models, regardless of message ordering. Empirical validation spans three tiers: controlled 4x4 tensors (104/104 tests pass), production-scale models up to 7.24B parameters (208 strategy-level tests, 43,368 layer-level property checks at capped tensor resolution), and multi-node convergence under gossip and partition healing (100 nodes, 20 orderings), with CRDT overhead below 0.5 ms. Because the wrapper is transparent, downstream performance is identical by construction, confirmed via byte-identical output verification. The reference implementation is available as crdt-merge v0.9.4.