pub fn ffn_gateup_fused_bytes(hidden: usize, inter: usize) -> usizeExpand description
What ONE fused dense-FFN gate+up weight costs: the two [inter, hidden]
E4M3 blocks appended along N, plus their two [inter/128, hidden/128] FP32
block-scale grids appended the same way (#927).
The sum of the two grids and NOT one grid over the fused N, for the reason
predicted_residency::ssm_concat_bytes gives: ceil of a sum is not the
sum of the ceils, and the concat copies the two grids side by side.
RESIDENCY-NEUTRAL: prune_after_load releases the two [inter, hidden]
store tensors this copied, so the same number is also what the checkpoint
gives back. Both sides are priced from this one function.