Skip to content

vllm.models.deepseek_v41.nvidia.specialized_decode.attention

Modules:

  • aux_projections –

    Low-batch GEMMs of the attention side projections.

  • candidate_topk –

    Candidate-block selection with cooperative_topk for decode rows.

  • decode_topk_cute –

    Decode top-k of the dense indexers straight from their BF16 logits, in CuTe DSL.

  • indexer –

    The sparse-attention indexers in a specialized decode step.

  • indexer_k_store –

    MXFP4 indexer key store for the FP32 output of the low-batch wk GEMM.

  • mxfp8 –

    MXFP8 attention projections.

  • sparse_mla –

    The FlashInfer sparse-MLA decode of a specialized decode step.

  • sparse_mla_cute –

    The sparse-MLA decode of small batches, in CuTe DSL.

  • step –

    The attention of a specialized decode step.