Lumabri – P2P Swarm-Based MoE Model Execution Using Colibri

Lumabri supports distributed execution of large-scale Mixture-of-Experts (MoE) models in P2P Swarm environments via the Colibri engine.

Core Mechanisms:

• On-demand Data Loading: There is no need to download the entire model in advance. During inference, only the bytes actually accessed are fetched from peers and stored in a local mirror, enabling high-speed execution from the local disk on subsequent calls.
• File System Interception: Uses an LD_PRELOAD shim to intercept libc calls via liblumabri.so. This allows model files to behave like sparse local mirrors with actual sizes, while missing blocks are immediately replenished over the network.
• MoE Distributed Processing: The execution host (Chatter) retains only dense weights, routers, and KV caches. It sends only 4KB-sized Activation values to peers holding each Expert to perform computations, ensuring Expert weights are never transferred to the execution host.

Reliability and Security:

• Integrity Verification: All data blocks are verified using SHA256, and peer data is validated via ed25519 signatures.
• Computing Verification: To ensure the reliability of remote computing, a certain percentage of calls are re-executed on other replicas to verify that the results match.

This engine guarantees identical outputs in both CPU and GPU environments and maintains data consistency even when the network configuration changes.