vllm.models.kimi_k3.nvidia.specialized_decode.cutedsl.kda

Modules:

• commit –

  RecoverSSM commit of an FP32 KDA state: one CTA per (head, request, layer).

• fused –

  Fused replay verify (KdaFused): the RecoverSSM commit of the previous step and the

• fused_ws –

  Persistent warp-specialized fused replay verify (KdaFusedWS): KdaFused's math

• kernel –

  Cluster-split KDA recurrence; an optional fifth warp projects F_B and multicasts the

• prepared –

  Prepared KDA core. One packed projection supplies Q/K/V, the output gate, F_A and

• specialized –

  Persistent warp-specialized KDA core for wide steps (KdaWS).

• stream –

  Persistent KdaWide: each CTA walks a contiguous range of (request, head) items and

• wide –

  Throughput KDA core: one CTA per (request, head) holds the head's whole FP32 state