vllm.models.kimi_k3.nvidia.specialized_decode.cutedsl.mla

Modules:

• dcp –

  MLA decode under decode context parallelism (DCP = TP, rank d owns head block d)

• prepared –

  Prepared MLA core: the packed input projection feeds the query kernel, then either

• query –

  Cluster MLA query RMSNorm, Q_B projection, W_UK absorption and paged FP8 KV insert.

• value –

  MLA value projection (W_UV) and sigmoid gate of the BF16 attention latent.