Struct DflashKv 

pub struct DflashKv {
    pub k: Vec<CudaSlice<f32>>,
    pub v: Vec<CudaSlice<f32>>,
    pub len: usize,
    pub cap: usize,
    /* private fields */
}

Draft KV cache (round-cost fix, 2026-07-13): per-layer normed+roped ctx K and raw ctx V, append-only in committed order. Block K/V land TRANSIENTLY at [len..len+b] each round (never committed — the reference crops them identically). Kills the per-round full-ctx projection recompute (first light was O(ctx)/round -> 7 tok/s).

Fields

k: Vec<CudaSlice<f32>>

v: Vec<CudaSlice<f32>>

len: usize

cap: usize

Implementations

impl DflashKv

pub fn new( e: &Engine, cfg: &DflashCfg, cap: usize, ) -> Result<Self, Box<dyn Error>>

impl DflashKv

pub fn export_tail(&self, e: &Engine, upto: usize) -> Option<DflashKvTail>

Copy out the readable tail ending at upto (see DflashKvTail). None when there is nothing to publish or an allocation fails — publication is always optional.

upto IS NOT self.len, and conflating them was the bug the first exactness-gate run caught: publication happens at the scheduler’s drain sweep, by which time the session has committed generated rows, so len had run 35 rows past the capture boundary and every restore was refused with draft KV len 30364 != prompt 30329. The trunk planes are copied at the capture pos for the same reason; the tail must agree with them.

pub fn from_tail( e: &Engine, cfg: &DflashCfg, cap: usize, tail: &DflashKvTail, ) -> Option<Self>

Rebuild a draft KV from a published tail: a fresh allocation at cap, the tail copied back to the SAME absolute rows it came from, and len restored so the next round addresses positions exactly as a cold-primed session would.

Rows below tail.base are ZEROED, not left uninitialised. The clipped SDPA never reads below the block’s window floor, but the legacy full-scan kernel (MEMRA_DFLASH2_SDPA_CLIP=0, the rollback seam) scans EVERY row into the score and the output, relying on masked rows contributing exactly zero — an identity that holds only for finite data (0.0 * NaN = NaN, and an uninit K row can produce a NaN score that poisons the softmax sum). Zeros keep that identity on both kernel arms, so a clip rollback on a restore-armed box stays byte-exact instead of decoding silent garbage (review round 3). Rows above tail.len stay uninit — equally unwritten and unread in the cold path, so restored matches cold there.

This function still REFUSES rather than trusts the window math — if the tail does not cover the window, the caller gets None and must cold-prime.

pub fn cfg_window_rows(&self) -> usize

Rows a restore must carry (see DsparkSpecSession::draft_tail_rows). Stored here because DflashKv owns the row geometry; the value comes from the drafter cfg.

pub fn row_bytes(&self) -> usize

Bytes per row per layer (n_kv * head_dim * 4), the unit both the export and the import address rows in.

pub fn n_layer(&self) -> usize

Number of draft layers, i.e. how many per-layer planes an export produces.