vllm.models.minimax_m3.amd.mono.kernels

FlyDSL kernels of the mono decode path: K1 pre_attn and K4 post_attn.

Modules:

• common –

  Device-side building blocks shared by the mono kernels.

• index_score –

  Indexer block scores for K4's sparse selection (select_blocks polls them).

• post_attn –

  K4 mono_post: one MiniMax-M3 sparse MoE layer, attention through FFN reduce.

• pre_attn –

  K1 mono_pre: one MiniMax-M3 sparse layer, input norm through cache insert.