Skip to content

vllm.models.kimi_k3.amd.specialized_decode.mono.attention

Modules:

  • back –

    K2a attn_post: one Kimi-K3 KDA layer from the gated core output to the

  • kda –

    K1 kda_pre: one Kimi-K3 KDA layer from the AttnRes seam to the gated core

  • mla –

    M1a mla_pre: one Kimi-K3 MLA layer from the AttnRes seam to AITER's

  • mla_attn –

    M1c mla_attn: Kimi-K3's MLA decode attention at TP8 + DCP1, the bf16

  • mla_attn_geom –

    M1c mla_attn's launch geometry, without the kernel's FlyDSL imports: the

  • mla_back –

    M1b mla_post: one Kimi-K3 MLA layer from AITER's decode output to