Struct DflashCfg 

pub struct DflashCfg {

    pub hidden: usize,
    pub n_head: usize,
    pub n_kv: usize,
    pub head_dim: usize,
    pub n_ff: usize,
    pub n_layer: usize,
    pub eps: f32,
    pub rope_theta: f32,
    pub block_size: usize,
    pub mask_token_id: u32,
    pub target_layer_ids: Vec<usize>,
    pub sliding_window: usize,
    pub layer_sliding: Vec<bool>,
    pub strategy_dspark: bool,
    pub is_causal: Option<bool>,
}

Fields

hidden: usize

n_head: usize

n_kv: usize

head_dim: usize

n_ff: usize

n_layer: usize

eps: f32

rope_theta: f32

block_size: usize

mask_token_id: u32

target_layer_ids: Vec<usize>

sliding_window: usize

layer_sliding: Vec<bool>

true = sliding_attention for that layer (4x true + 1x false on the 31B draft).

strategy_dspark: bool

Checkpoint training-strategy census (dspark_strategy_census over the raw config.json): true = a SpecForge DSPARK-strategy export (shifted labels, ALL rows supervised — the q38 arm-a family). Keys the HARVEST DEFAULT strategy-keyed, never env-keyed (owner-ratified 2026-08-20 after B1 confirmed H1 ×5; DSPARK-POSTMORTEM-20260820.md B0 default-flip plan).

is_causal: Option<bool>

Explicit top-level is_causal from config.json (z-lab reference: an explicit value OVERRIDES the per-layer-type default). The DFlash2 q38 checkpoint carries "is_causal": false — every sliding layer is NON-causal with a symmetric +/-2048 window (model.py _attention_mask). None = key absent (historical exports; the windowless-assert arm keeps handling those byte-identically).