Paper: Conflict-Free Replicated Data Types for Neural Network Model Merging
Posted by Character_Bison5968@reddit | LocalLLaMA | View on Reddit | 5 comments
Paper presenting a two-layer CRDT architecture (CRDTMergeState) that enables conflict-free merging of neural network models across 26 strategies.
Paper:
https://github.com/mgillr/crdt-merge/blob/main/paper/CRDT_Merge_ArXiv.pdf
Silver-Champion-4846@reddit
When can we expect test results?
Character_Bison5968@reddit (OP)
The paper includes full test results across three tiers.. controlled 4×4 tensors (104/104 tests pass), production scale models up to 7.24B parameters (208 strategy-level tests, 43,368 layer-level checks), and multi node convergence with 100 nodes across 20 gossip orderings. See Tables 1–9 and Sections 6.1–6.5. If there are specific additional tests anyone would like to see, please raise them as issues on the repo.
Silver-Champion-4846@reddit
I mean actual open models we can use, I'm blind and can't see the complexities of the tables in the papers
Character_Bison5968@reddit (OP)
Im aware your blind, just posting for the visionaries
Silver-Champion-4846@reddit
Alr