Skip to content

vllm.models.kimi_k3.nvidia.recoverssm.cutedsl

Modules:

  • commit –

    KDA RecoverSSM state commit: one CTA per (head chunk, request, layer).

  • verify –

    KDA RecoverSSM verify on the fifth-generation tensor cores: one CTA (4 warps) per

  • verify_ws –

    KDA RecoverSSM verify with the fused short convolution on the fifth-generation