Skip to content

vllm.models.minimax_m3.amd.mono.drafter

Modules:

  • attention –

    Attention of the DSpark drafter kernels.

  • device –

    Device helpers of the DSpark drafter kernels: memory, bounded waits, the MXFP8

  • dispatch –

    Which MiniMax-M3.1 DSpark draft steps run on the drafter kernels

  • layer –

    The DSpark drafter's 5 layers and final norm as one persistent launch.

  • layout –

    Shapes, CTA roles and memory layout of the DSpark drafter kernels.

  • runner –

    Host side of the DSpark drafter kernels: one rank's weights, scratch, peers and

  • wide –

    The DSpark drafter's 5 layers and final norm as one launch for 3..8 requests.