vllm.models.kimi_k3.nvidia.specialized_decode.cutedsl.gemm

Modules:

• decode –

  The attention input projection of a decode step, one engine per prepared row count.

• projection –

  Split-K projection on the FlashInfer CuTe core vendored in splitk.py (with an

• splitk –

  FlashInfer 0.7.0.post1 SplitKDenseGemmKernel (dense_bf16_gemm_sm100_splitk)

• warp_splitk –

  FlashInfer's cp.async-pipelined BF16 GEMM for M <= 32 whose four compute warps hold

• ws_gemm –

  Weight-stationary skinny GEMM for the MoE front: y[M, N] = x[M, K] @ W[N, K]^T.