Skip to content

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

Modules:

  • chain –

    Small-batch MoE chains: front -> routed experts -> (shared expert) -> polling tail.

  • experts –

    Small-batch routed experts: ClusterExpertsTP16 (TP16 B1/B2, six-CTA split-K

  • front_tp16 –

    TP16 B1/B2 MoE front: SM placement, early router/latent-down staging, MMA router

  • front_tp8 –

    TP8 B1 small-batch front (full router staged in SMEM, x_moe polling, threshold top16,