vllm.models.inkling.nvidia.mlp ¶

Inkling dense SwiGLU MLP (also used as the MoE shared expert).

The checkpoint stores the gate/up projection as a single fused, interleaved weight ([gate0, up0, gate1, up1, ...]), so we use a plain ColumnParallelLinear whose contiguous row-sharding keeps each gate/up pair together, and an interleaved SwiGLU activation.