Skip to content

vllm.models.kimi_k3.nvidia.specialized_decode.cutedsl.dsa

Modules:

  • attn_ds –

    fp8_ds_mla sparse core of a K3.1 DSA layer (FLASHMLA_SPARSE), as the stock decode:

  • attn_fi –

    FlashInfer sparse MLA core of a K3.1 DSA layer (FLASHINFER_MLA_SPARSE): the FP8

  • attn_sparse –

    Specialized sparse MLA decode of a DCP K3.1 DSA layer on a per-tensor FP8 cache

  • dcp_stage –

    DCP stage of a K3.1 DSA layer (DCP = TP; each rank's cache holds 1 / DCP of every

  • front –

    Fused DSA indexer front: the q_a RMSNorm, the wq_b projection (N 4096, K 1536) and

  • front_mma –

    Fused DSA indexer front on tcgen05: the wq_b projection (N 4096, K 1536) and

  • index –

    DSA indexer inputs of one decode step, one 256-thread CTA per token:

  • indexer –

    Fused decode indexer of a K3.1 DSA layer. TP (every rank scores the whole

  • indexer_stage –

    Indexer stage of a K3.1 DSA layer: the stock decode scoring and top-k

  • overlap –

    Query overlap of a K3.1 DSA layer. The indexer path (wq_b -> DsaIndex -> scoring

  • prepared –

    Prepared DSA core (Kimi K3.1 sparse MLA layers): the orchestration of one layer's

  • query_stage –

    Query stage of a K3.1 DSA layer: MlaQuery (dsa) -> wq_b -> DsaIndex.

  • topk_exchange –

    DCP top-k candidate exchange over NVLink and its exact merge (K3.1 DSA decode).