Skip to content

vllm.models.minimax_m3.amd.mono.drafter.layout

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

The MiniMax-M3.1 DSpark draft model (5 layers) at TP4 on gfx950: 256 CTAs of 512 threads, one CTA per CU. A draft step of up to 2 requests (7 query rows each) is one layer launch; 3..8 requests are one wide launch. Roles are fixed per CTA for the whole launch:

  • G, CTAs 0..191: the MXFP8 GEMVs of every layer (QKV 12 rows, o_proj 32, gate 16 + up 16, down 32 a CTA).
  • L, CTAs 192..239: attention partitions and the merge tasks.
  • N, CTAs 240..255: the row norms; a CTA keeps its row's residual in LDS.

Scratch regions are per layer and written once a launch; flag regions and the symmetric (IPC) regions are shared by the layers and told apart by their tags (step * 8 + layer + 1).