pub const BA_GATES_GROUPS: u32 = 8;Expand description
BAH_GROUPS in the kernel: output groups a thread accumulates at once, and
therefore how many BA outputs one fetch of the activation row serves. The
row is read ceil(ceil(N/4)/8) * 4 times per token — 12 at N=96, against
the parent’s 96, which reads it once per output. Chosen on ptxas: 64
registers and no spill at 8; 12 spills and 16 costs a quarter of the
occupancy for 3.3% of the instructions.