LPLB: linear-programming expert-parallel load balancer
Early-stage open library that solves a small linear program per batch to rebalance MoE expert workload across GPUs, building on EPLB.
Reorders experts from workload statistics, adds replicas based on topology and computes optimal token assignment per batch. DeepSeek says performance gains are still under evaluation.
- Date
- Wednesday, 19 November 2025
- Lab
- DeepSeek
- Kind
- infra
- Access
- open weights
Date is GitHub repo creation (2025-11-19). Software, not model weights. README labels it early research.
Sources
This record was checked against its sources on 6 October 2026. How we check