Constant MOE_ROW_BATCH_MAX_ROWS 

pub const MOE_ROW_BATCH_MAX_ROWS: usize = 8;

Widest compiled w4a16_gemv_sw_moe_batchm_mR tier. Mirror of the ATLAS_MOE_BATCHM_ENTRY list in kernels/gb10/common/w4a16_gemv.cu and of the [KernelHandle; 7] in Glm5NextMlpKernels.

🔴 Since 2026-09-02 this is a sub-group width, not a caller contract. The prefill sub-chunk is 16 rows wide (glm5next_layer::PREFILL_ROWS) because the DENSE tier widened to 16; the routed experts did not follow, so forward_moe splits any wider row group into even sub-groups of at most this and sweeps each one. Callers may pass any rows their workspace holds.