Module minimax_engine 

MiniMax M2/M3 dedicated stack stub — not generic GQA.

Real checkpoints tag minimax-m2 / minimax-m3 with 256-expert sigmoid MoE and MTP (multi-token prediction) draft heads. Required tensors (llama.cpp minimax*.cpp graph), not implemented here:

• Standard emb/norm/output head
• MoE: ffn_gate_inp, ffn_gate_exps, ffn_up_exps, ffn_down_exps, ffn_exp_probs_b.bias (sigmoid / noaux_tc routing)
• MTP: num_nextn_predict_layers draft-head tensors (nextn.* in llama.cpp) — see docs/CLI.md --mtp (honest fail until loaded)

Fail-closed via [Self::reject] until loader + engine land.

Structs

MinimaxEngine

MinimaxUnavailable