Constant ATTN_MMA_QUERIES 

pub const ATTN_MMA_QUERIES: usize = 64;

The query vectors a block of the f16 tensor-core attention (lk_attention_mma) takes, a warp 16 of them.