pub fn step_tile_ref(
state: &mut AttnState,
q: &[bf16],
k_pool: &[bf16],
v_pool: &[bf16],
tile_blocks: &[i32],
tile_block_counts: &[i32],
num_seqs: usize,
num_q_heads: usize,
num_kv_heads: usize,
head_dim: usize,
block_size: usize,
tile_capacity: usize,
gqa_ratio: usize,
)Expand description
2026-09-25: One tile update over the paged layout [num_blocks, block_size, num_kv_heads, head_dim], with no causal mask. tile_blocks is
[num_seqs, tile_capacity], tile_block_counts is [num_seqs]. It runs
the kernel’s per-token recurrence in fp32, but sums each dot product
sequentially and uses exp where the kernel uses a warp-tree sum and
__expf.