pub fn graph_kv_pull_host(
_kv_id: u64,
_layer: usize,
_from: usize,
_to: usize,
_nkv: usize,
_hd: usize,
) -> Option<(Vec<f32>, Vec<f32>, usize)>Expand description
Rows [from, to) of one wgpu exact-attention mirror in the host
cache’s layout (V zero-padded to hd), whatever the mirror’s geometry
(narrow V, a sliding layer’s ring). The third value is the first
position actually read: a ring returns zeros below it. None: no wgpu
mirror holding those rows at (nkv, hd).