vllm.models.minimax_m3.amd.mono

MiniMax-M3 fused per-layer decode of up to config.MAX_TOKENS tokens; entered through dispatch.MonoDecode (VLLM_ROCM_MINIMAX_M3_MONO=1).

Modules:

• check –

  Debug tooling for the mono path.

• config –

  Shapes the mono kernels are compiled for: one MiniMax-M3 TP4 rank.

• dispatch –

  Which MiniMax-M3 forwards take the mono path (VLLM_ROCM_MINIMAX_M3_MONO=1).

• envs –

  Environment switches of the mono decode path (read once at import).

• kernels –

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

• layout –

  Where the mono layer kernel (K4) puts its tasks and data, per decode token count.

• peer_buffer –

  Symmetric peer memory for the mono kernels' in-kernel all-reduces.

• runner –

  One decode step of 1..MAX_TOKENS tokens on the mono kernels.

• weights –

  Zero-copy views (and two converted copies) of one vLLM MiniMaxM3DecoderLayer