An educational, ground-up rebuild of DeepEP’s expert-parallelism dispatch — covering variable-length GPU-to-GPU communication, quantized data transfer, and multi-GPU interconnect optimization, validated against a reference implementation at each stage.
Back to projects
May 31, 2026
1 min read
Mini-DeepEP
A from-scratch reimplementation of DeepSeek's expert-parallelism communication library, built up in stages on real multi-GPU hardware.
