vllm.models.kimi_k3.nvidia.specialized_decode.cutedsl.moe.ep

Expert-parallel (DEP) latent MoE: NVLink dispatch / combine, the layer schedule.

Modules:

• comm –

  Expert-parallel dispatch and combine of the latent MoE over NVLink symmetric memory

• layer –

  DEP (TP1 attention, EP over all ranks) latent-MoE layer.