Skip to content

vllm.models.kimi_k3.nvidia.specialized_decode.cutedsl.common

Modules:

  • embedding –

    Local CuTe vocabulary gather plus the ordinary framework TP all-reduce.

  • host –

    Host-side helpers for the prepared CuTe providers.

  • mailbox –

    Lamport mailboxes: BF16 rows published in 16-byte fragments; a reader polls until no

  • policy –

    Per-batch policy tables.

  • prefetch –

    L2 weight prefetch: each warp walks a persistent device table of (address, bytes),

  • primitives –

    Inline-PTX, memory, synchronization and math helpers of the decode kernels.

  • twoshot –

    Two-shot Lamport all-reduce of BF16 rows [B, width] over TP ranks.