DeepSeek Open-Sources DeepEP V2 and TileKernels for MoE Training and Inference Infrastructure
GitHub·medium signal
DeepSeek released DeepEP V2 (expert-parallel communication library with high-throughput all-to-all GPU kernels for Mixture-of-Experts models) and TileKernels (a kernel library written in tilelang) on April 23. Both are open-source infrastructure components that power DeepSeek's own training and inference stack. For builders running local MoE models on multi-GPU setups, DeepEP V2's optimized expert routing directly improves inference throughput.