pub const KDA_FRONT: usize = 20;Expand description
kda_mixer split three ways. The parent bucket is 53.5% of prefill — by far the
largest — but it spans three very different things: batched dense projections, a
per-token recurrence that CANNOT be batched in the recurrent arm, and the output
norm + o_proj. Chunking the recurrence measured ~4% SLOWER, which only makes sense
if the recurrence is not the bulk; these say which part to actually attack.