Enum DsparkHarvest 

pub enum DsparkHarvest {
    Dflash,
    Dspark,
}

Draft-row harvest convention for DFlash-family block drafters (darklanes research/deepseek-flash-20260818/DSPARK-POSTMORTEM-20260820.md).

The DFlash and DSpark SpecForge training strategies supervise DIFFERENT rows of the same [anchor, MASK x b-1] block, so the row -> trunk-position mapping is a property of the CHECKPOINT’s training strategy, not of the loader:

• Dflash (mask-fill; z-lab dflash / SpecForge OnlineDFlashModel): row k is trained to predict the token AT position anchor+k — “Labels: same-position prediction”, weight_mask *= (pos_in_block > 0) excludes the anchor row (SpecForge specforge/algorithms/common/dflash_family_model.py:453-472). Drafts = rows 1..b-1; the anchor row’s output is untrained.
• Dspark (shifted; SpecForge OnlineDSparkModel, training.strategy: dspark — the q38 arm-a export): row k is trained to predict the token at anchor+k+1, ALL rows supervised INCLUDING the anchor row (label_offsets = arange(1, block_size+1), dflash_family_model.py:816). sglang’s DSPARK worker — the stack every arm-a bank number was measured on — harvests gamma = block_size drafts with the anchor row’s output as draft 1 (verified on the v0.5.17 eval-pin tag: dspark_components/dspark_draft.py:248,260,318; dspark_config.py:269).

Mismatching the convention verifies every slot against a position the row was never trained for — the q38 accept collapse (2.9 -> 1.43) in the postmortem.

Variants

Dflash

mask-fill: drafts = rows 1..b-1, row k fills position anchor+k.

Dspark

shifted: drafts = rows 0..b-1, row k predicts position anchor+k+1.

Implementations

impl DsparkHarvest

pub fn resolve(cfg: &DflashCfg) -> Self

The served resolution: explicit MEMRA_DSPARK_HARVEST={dflash|dspark} wins (unknown values REFUSE loudly — a typo silently reverting the convention would re-open the postmortem’s misalignment); UNSET defers to the CHECKPOINT’s own training-strategy census — the owner-ratified default flip (2026-08-20, after B1 confirmed H1 interleaved ×5 on serving-class hardware: accept 1.38→2.41 agentic / 1.53→3.66 math, E2E ALL EXACT both arms). Strategy-keyed, not env-keyed, per the B0 plan: a DSPARK-strategy export harvests shifted (all-rows), a mask-fill export keeps the historical dflash arm byte-identical.

pub fn resolve_value(v: Option<&str>, strategy_dspark: bool) -> Self

pub fn from_env_value(v: Option<&str>) -> Self

ENV-ONLY parser (no checkpoint census): unset = Dflash, the historical arm. Kept for the explicit-value path of Self::resolve_value and the seam tests; round arms resolve through Self::resolve so the default stays strategy-keyed.

pub fn name(self) -> &'static str

Manifest/serialized name (the oracle geometry manifest’s harvest field).

pub fn from_name(v: &str) -> Option<Self>

pub fn first_row(self) -> usize

First drafter OUTPUT row consumed as a draft candidate.

pub fn n_drafts(self, b: usize) -> usize

Drafted tokens harvested per round from a b-row block.

pub fn trained_offset_of_row(self, row: usize) -> usize

The position offset (relative to the round anchor at the block’s row 0) that drafter output row row is TRAINED to predict under this convention.

Trait Implementations

impl Clone for DsparkHarvest

fn clone(&self) -> DsparkHarvest

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for DsparkHarvest

impl Debug for DsparkHarvest

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for DsparkHarvest

impl PartialEq for DsparkHarvest

fn eq(&self, other: &DsparkHarvest) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for DsparkHarvest