Skip to content

vllm.models.minimax_m3.amd.specialized_decode.mono

gfx950 kernel stages for MiniMax-M3 specialized decode.

Modules:

  • a16w4 –

    BF16 x MXFP4 (A16W4) MFMA operands over AITER's post-load layouts.

  • ffn –

    MiniMax-M3 FFN launch: one persistent kernel from the unreduced o_proj

  • residency –

    The HIP driver's occupancy of a compiled FlyDSL kernel: a persistent grid

  • runner –

    Host side of the FFN launch: one compiled kernel, scratch region and the

  • sources –

    This package's kernel sources and the DS V4.1 primitives they import, in

  • sync –

    Bounded tagged hand-offs and the sticky error word.