Struct SpecSampling 

pub struct SpecSampling {
    pub temp: f32,
    pub seed: u64,
    pub top_k: i32,
    pub top_p: f32,
    pub min_p: f32,
    pub penalty_last_n: usize,
    pub penalty_repeat: f32,
    pub penalty_freq: f32,
    pub penalty_present: f32,
}

Scratch KV for the MTP block (one full-attn layer).

PERSISTENT MODE (default, 2026-07-03 — the acceptance lever): sized cap = max_ctx and kept in sync with the COMMITTED sequence — slot p holds the MTP block’s K/V for committed token p (roped p+1, the chain’s rope convention), so the draft chain’s self-attention sees the FULL committed history instead of only the current round’s 1..K+1 chain tokens (the reference engine’s “mtp_update” design). Entries come from two sources:

• chain appends: accepted positions KEEP their chain-computed entries (embedding exact, hidden chain-approximate — the reference engine accepts the same);
• mtp_kv_fill batches: prompt positions + the last-draft position on full accept, computed from EXACT trunk hiddens (K/V-only MTP-block pass, no attention/FFN/lm_head). Rejected drafts / p-min extras / pseudo-seed appends are all discarded by the round-start set_len truncation (the KvLayer len mechanism — §C rollback for the draft side). Multi-turn spec-decode session (2026-07-05): trunk Cache + persistent MTP draft scratch + the committed token list, alive across generate_spec_session calls. Turn N+1 primes ONLY its suffix (chunked continuation prime over the quantized past) and mtp_kv_fill’s its suffix rows, then the round loop runs unchanged. last_h carries the pre-output_norm hidden of the last committed row across turns (the predecessor-pairing seed + fill anchor). Per-request sampling config for the sampled-spec serve path.

Fields

temp: f32

seed: u64

top_k: i32

top_p: f32

min_p: f32

penalty_last_n: usize

penalty_repeat: f32

penalty_freq: f32

penalty_present: f32

Trait Implementations

impl Clone for SpecSampling

fn clone(&self) -> SpecSampling

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for SpecSampling

impl Debug for SpecSampling

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

Formats the value using the given formatter.