Skip to content

vllm.models.kimi_k3.nvidia.specialized_decode.cutedsl.oproj

Modules:

  • consumer –

    O-projection all-reduce consumer (groups 128-thread groups poll disjoint

  • prepared –

    Output projection fused with its TP all-reduce and post-attention AttnRes: rank

  • producer –

    O-projection producer for the split and fused AttnRes paths: 112 CTAs of 64 output