XShare: Collaborative in-Batch Expert Sharing for Faster MoE Inference
This addresses efficiency bottlenecks in deploying large MoE models for inference, offering practical improvements for production systems.
The paper tackles the problem of reduced efficiency in Mixture-of-Experts (MoE) architectures during production inference due to request batching and speculative decoding, by proposing XShare, a method that reduces expert activation by up to 30%, cuts peak GPU load by up to 3x, and achieves up to 14% throughput gains.
Mixture-of-Experts (MoE) architectures are increasingly used to efficiently scale large language models. However, in production inference, request batching and speculative decoding significantly amplify expert activation, eroding these efficiency benefits. We address this issue by modeling batch-aware expert selection as a modular optimization problem and designing efficient greedy algorithms for different deployment settings. The proposed method, namely XShare, requires no retraining and dynamically adapts to each batch by maximizing the total gating score of selected experts. It reduces expert activation by up to 30% under standard batching, cuts peak GPU load by up to 3x in expert-parallel deployments, and achieves up to 14% throughput gains in speculative decoding via hierarchical, correlation-aware expert selection even if requests in a batch drawn from heterogeneous datasets.