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
One tile update. Mirrors the CUDA kernel’s per-token online-softmax recurrence exactly (same accumulation order, fp32 throughout).