vllm.models.minimax_m3.amd.specialized_decode.runtime

Side-effect-free MiniMax-M3 specialized-decode configuration.

Modules:

• caches –

  Named cache ABI validation and identity tracking.

• decode –

  Layer-mode specialized-decode lifecycle and per-forward plan.

• executor –

  FFN-launch executor: vLLM runs each sparse layer's input norm and attention,

• guards –

  Aggregated loaded-model and device refusals.

• routing –

  Rank-uniform per-step routing and cache identity tracking.

• spec –

  Side-effect-free configuration for MiniMax-M3 specialized decode.

• weights –

  Checked zero-copy views of MiniMax-M3 post-load tensors.