Struct DflashKvTail 

pub struct DflashKvTail {
    pub layers: Vec<(CudaSlice<f32>, CudaSlice<f32>)>,
    pub base: usize,
    pub rows: usize,
    pub len: usize,
    pub row_bytes: usize,
}

A DFlash draft-KV tail, per drafter layer, ready to ride a cross-request prefix-cache entry: (k, v) f32 rows covering absolute positions [base, base + rows).

Only the tail travels, and that is a fact about this export rather than an optimisation: every DFlash2 draft layer is sliding_attention (the port asserts it at load), so a round at context pos reads rows [pos - window + 1, pos + block) and nothing older. Storing the whole history for a 30k-token prompt would be ~1,229 MB — MORE than the ~1,057 MB of trunk planes it would ride with; the tail is ~85 MB.

Fields

layers: Vec<(CudaSlice<f32>, CudaSlice<f32>)>

base: usize

Absolute position of the first stored row.

rows: usize

Rows stored per layer.

len: usize

Logical length the KV had when exported (= pos), so an import can restore the same absolute row addressing the rope positions were baked against.

row_bytes: usize

Bytes per row per layer, carried so an import cannot disagree about the geometry.

Implementations

impl DflashKvTail

pub fn bytes(&self) -> usize