project_kv_block_ref

Function project_kv_block_ref 

Source
pub fn project_kv_block_ref(
    k_block: &[bf16],
    p: &[bf16],
    block_size: usize,
    num_kv_heads: usize,
    head_dim: usize,
    r: usize,
) -> Vec<bf16>
Expand description

Per-token projection. Output layout [num_kv_heads, block_size, r].