vllm.models.kimi_k3.nvidia.specialized_decode.drafter

Kimi K3 / K3.1 DSpark drafter decode path (dense MLA draft layers).

Modules:

• attention –

  Dense MLA decode of the DSpark draft block with the query tokens folded into heads.

• comm –

  The DSpark drafter's TP collectives on the specialized decode's Lamport mailboxes.

• linears –

  The DSpark drafter's BF16 linears on the decode GEMM engines.

• markov –

  DSpark's sequential Markov sampling, one launch per draft step.

• model –

  The specialized decode's DSpark drafter (decode_backend=latency_specialized).