Constant W8A16_GEMM_M16_N_TILE_WIDE 

pub const W8A16_GEMM_M16_N_TILE_WIDE: u32 = 64;

The wide instantiation’s N tile (w8a16_gemm_m16_n64, kernel M16_N_TILE_WIDE) — opt-in via ATLAS_FFN_M16_TC_NTILE=64. Halves the CTA count for a given N and doubles the reuse of each staged A fragment. WHY it exists and what it is meant to settle: dense_ffn_m16_tc.rs.