Constant W8A16_GEMM_M16_N_TILE 

pub const W8A16_GEMM_M16_N_TILE: u32 = 32;

N columns one CTA owns on the DEFAULT instantiation. SSOT for the launch geometry AND for the dispatch rule’s “does this shape have enough CTAs” reasoning, so the two cannot drift: the kernel’s M16_N_TILE must equal this.