ScatterMoE

Triton-based sparse MoE kernels — faster expert routing for open MoE LLMs.