[DAC'25] Official implement of "HybriMoE: Hybrid CPU-GPU Scheduling and Cache Management for Efficient MoE Inference"
Python
105
14
Apache License 2.0
HybriMoE is a system designed to improve the efficiency of Mixture-of-Experts (MoE) inference by optimizing CPU-GPU scheduling and cache management. It introduces hybrid scheduling to balance workloads, impact-driven prefetching to prioritize expert loading, and MoE-specialized caching to reduce cache misses. This project is intended for researchers and developers working on large-scale MoE models, particularly those seeking to optimize inference performance and resource utilization.
Total donated
Undistributed
Share with your subscribers:
Recipients
How the donated funds are distributed
Support the dependencies
Support the repos that depend on this repository
Top contributors
shuzhangzhong
6 contributions
Recent events
Kivach works on the Obyte network, and therefore you can track all donations.
No events yet