A CONCEPT · PROJECTED FROM THE VINDEX KNOWLEDGE GRAPH
MAMBA2 — THE SSD MIXER
the mixer is the block
A pure state-space operator: one fused five-way input projection, a causal conv over the x|B|C channels, per-head scalar decay, a gated RMSNorm — and no attention and no FFN anywhere, because the mixer is the whole layer.
Nine operand roles, closure needs all: the fused in-projection, the conv and its bias, A_log, D, dt_bias, the gated norm, the out-projection, and the layer's single pre-mixer norm. Its continuation is TWO declared state regions, not one: an SSM state (head_dim × state_size per head) and a convolution history — both constant in sequence length, and the history is load-bearing at the first single-token continuation. The declared geometry must close over the tensor estate exactly, an unbounded dt-clamp side is carried as a declared fact, and the generic executor is the reference mixer transcribed stage by stage — paritied against the fp32 oracle with argmax exact at all 430 scored positions and every greedy trajectory token-for-token.
IN THE GRAPH
mamba2 — the ssd mixer— sibling of →linear attention
mamba2 — the ssd mixer— carries →continuation state
mamba2 — the ssd mixer— verified by →operand closure
mamba2 — the ssd mixer— admitted by →the layer operator