pub fn paged_mla_decode_attention_f32(
query: &[f32],
query_pe: &[f32],
key_value_cache: &[f32],
key_pe_cache: &[f32],
actual_seq_lens: &[i32],
block_table: &[u32],
batch: usize,
heads: usize,
head_dim: usize,
pe_dim: usize,
block_size: usize,
block_table_batch_stride: usize,
scale: f32,
output_scale: f32,
) -> (Vec<f32>, Vec<f32>)