vllm.models.kimi_k3.nvidia.ops.cute_dsl ¶

Modules:

• gemm_rs_ar –

  SM100 BF16 GEMM with fused tensor-parallel reduce-scatter/all-reduce.

• kda_skinny_gemm –

  Kimi-K3 TP8 skinny GEMMs for the KDA F_A/beta and F_B projections.

• latent_moe_tail –

  CuTe DSL kernels for KimiK3LatentMoETailOp.