pub fn auto_splits(sm_count: u32, num_q_heads: u32) -> u32Expand description
2026-09-25: clamp(ceil(WAVES * sm_count / num_q_heads), 1, MAX_DECODE_SPLITS).
The split-K grid is (num_q_heads, num_splits, num_seqs), so with one
sequence this puts at least WAVES * sm_count CTAs on the device when the
cap allows. A wider batch launches more CTAs with the same count.