Struct DflashConfig 

pub struct DflashConfig {

    pub hidden_size: usize,
    pub num_hidden_layers: usize,
    pub intermediate_size: usize,
    pub num_attention_heads: usize,
    pub num_key_value_heads: usize,
    pub head_dim: usize,
    pub vocab_size: usize,
    pub draft_vocab_size: Option<usize>,
    pub tie_word_embeddings: bool,
    pub block_size: usize,
    pub dflash_config: Option<DflashSubConfig>,
    pub rope_theta: f32,
    pub rope_scaling: Option<DflashRopeScaling>,
    pub markov_rank: usize,
    pub markov_head_type: Option<String>,
    pub enable_confidence_head: bool,
    pub confidence_head_with_markov: bool,
}

Drafter HF config.json (subset Atlas consumes). Mirrors z-lab/Qwen3.6-35B-A3B-DFlash/config.json field names verbatim so serde_json::from_str works directly on the raw file.

Fields

hidden_size: usize

num_hidden_layers: usize

intermediate_size: usize

num_attention_heads: usize

num_key_value_heads: usize

head_dim: usize

vocab_size: usize

draft_vocab_size: Option<usize>

tie_word_embeddings: bool

block_size: usize

Block size γ. Qwen3.6-DFlash ships block_size: 16.

dflash_config: Option<DflashSubConfig>

DFlash-specific nested config object.

rope_theta: f32

Drafter base RoPE θ. Defaults to 10M (matches Qwen3.6-DFlash).

rope_scaling: Option<DflashRopeScaling>

HF-style rope_scaling block. None ⇒ plain RoPE (the v2 2026-04-27 Qwen3.6-DFlash drafter ships rope_scaling: null). When present and rope_type == "yarn", the drafter’s YaRN parameters are used to build the inv_freq table at construction time. alias = "rope_parameters": newer transformers releases (RadixArk DSpark, incoai DFlash2) ship the block under that key — without the alias Atlas silently drops the scaling (the RadixArk config.json had to be hand-patched before this).

markov_rank: usize

DSpark Markov head rank. 0 (default) ⇒ plain DFlash drafter with no Markov head. RadixArk Qwen3.8-27B-DSpark ships markov_rank: 256 top-level (SpecForge DSparkConfig convention — see the checkpoint’s dspark.py: DSpark fields are declared as top-level config attrs).

markov_head_type: Option<String>

DSpark Markov head flavor. Only "vanilla" (low-rank learned bigram bias) is defined by the reference; anything else is rejected at load.

enable_confidence_head: bool

DSpark confidence head (AcceptRatePredictor): a Linear(input, 1) predicting per-draft-position acceptance probability, used for adaptive block length. Loaded when present; consumed by the dynamic-K scheduling phase (the Markov fixup works without it).

confidence_head_with_markov: bool

When true the confidence head’s input is [hidden ‖ markov_embed] (input_dim = hidden_size + markov_rank); when false, hidden only.

Implementations

impl DflashConfig

pub fn effective_block_size(&self) -> usize

Resolved block size γ: the drafter’s own trained value when the checkpoint states it, else the top-level field.

The top-level block_size defaults to 16, and serde fills that default happily for a checkpoint that never mentioned it — so a DFlash2 drafter trained at 8 comes up as 16 unless the sub-config is consulted first. That is not a cosmetic mismatch: the serve then runs num_drafts=15 against an 8-block drafter, which measured 0% accept on EVERY verify step, and sizes the drafter’s per-sequence pools for twice the block it will ever use. --dflash-gamma still overrides both.

Trait Implementations

impl Clone for DflashConfig

fn clone(&self) -> DflashConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DflashConfig

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for DflashConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.