Struct DflashWeights 

pub struct DflashWeights {

    pub config: DflashConfig,
    pub fc: DenseWeight,
    pub hidden_norm: DenseWeight,
    pub norm: DenseWeight,
    pub layers: Vec<DflashLayerWeights>,
    pub draft_id_to_target_id: Option<Vec<i64>>,
    pub markov_w1: Option<DenseWeight>,
    pub markov_w2: Option<DenseWeight>,
    pub confidence_proj: Option<DenseWeight>,
    pub confidence_bias: Option<DenseWeight>,
    pub selector_pred: Option<DenseWeight>,
    pub selector_succ: Option<DenseWeight>,
    pub selector_hidden_proj: Option<DenseWeight>,
}

Raw weight bundle for the DFlash drafter, post-load.

Verified against z-lab/Qwen3.6-35B-A3B-DFlash (commit 42d3b34, May 2026): the checkpoint ships 91 BF16 tensors — fc.weight, hidden_norm.weight, norm.weight, plus 11 weights per drafter layer × 8 layers. No embed_tokens or lm_head are in the checkpoint — the drafter shares the target’s embedding and LM head at construction time. This matches the vLLM PR #40898 flow: when those keys are absent, vLLM’s AutoWeightsLoader adds them to skip_substrs, leaving the runtime to slot in the target’s pointers.

Fields

config: DflashConfig

fc: DenseWeight

[draft_hidden, len(target_layer_ids) * target_hidden]. Qwen3.6-35B-A3B-DFlash: [2048, 10240].

hidden_norm: DenseWeight

[draft_hidden] — RMSNorm applied to the projected target context before mixing with token embeddings.

norm: DenseWeight

[draft_hidden] — final RMSNorm before LM head.

layers: Vec<DflashLayerWeights>

draft_id_to_target_id: Option<Vec<i64>>

Present iff the drafter has a draft-id → target-id mapping (i.e. draft_vocab_size != target_vocab_size). Absent for Qwen3.6-35B-A3B-DFlash (both vocabs = 248320).

markov_w1: Option<DenseWeight>

Markov head markov_w1: [vocab, markov_rank] BF16 embedding table (prev-token → latent). Present iff config.markov_rank > 0 and the checkpoint carries the tensor.

markov_w2: Option<DenseWeight>

Markov head markov_w2: [vocab, markov_rank] BF16 (nn.Linear(rank, vocab, bias=False).weight, i.e. [N, K] for the GEMV convention). Projects the latent back to a full-vocab bias.

confidence_proj: Option<DenseWeight>

Confidence head weight: [1, hidden(+rank)] BF16.

confidence_bias: Option<DenseWeight>

Confidence head bias: [1] BF16.

selector_pred: Option<DenseWeight>

candidate_selector.predecessor_codebook [vocab, selector_rank] BF16.

selector_succ: Option<DenseWeight>

candidate_selector.successor_codebook [vocab, selector_rank] BF16.

selector_hidden_proj: Option<DenseWeight>

candidate_selector.hidden_projection.weight [selector_rank, hidden].