Struct V2State 

pub struct V2State { /* private fields */ }

Per-sequence mutable state: KV cache, token history, Engram ring, and every scratch buffer a step needs. Allocated once so a decode step does no heap work.

Implementations

impl V2State

pub fn pos(&self) -> usize

Absolute position of the next token.

pub fn history(&self) -> &[u32]

Tokens fed so far, oldest first.

pub fn kv_window(&self) -> usize

The attention window this sequence runs at.

pub fn set_kv_window( &mut self, window: Option<usize>, model: &V2Model, ) -> Result<(), &'static str>

Override the attention window. None restores the model’s own.

Needed because the two heads upstream ships disagree: the LM path runs the checkpoint’s kv_window (256), while architecture.encode_contrastive and forward_confidence default to window=0, full causal. Below 256 tokens the two are bit-identical; above it they diverge — at 642 tokens the contrastive embeddings differ by cosine 3.5e-4 and the confidence logit by 0.096 — so this is a real fork, not a rounding detail.

Takes effect from the next step; it does not rewrite the existing cache.

Fails when the requested window needs more cached positions than this state allocated — a window of 0 over a ring-sized cache would read stale slots and quietly return wrong embeddings. Use V2Model::make_state_full_causal for that case.

pub fn cache_len(&self) -> usize

Positions the KV ring holds.

pub fn kv_bytes(&self) -> usize

Bytes the KV cache occupies — the dominant part of a session.

pub fn reset(&mut self)

Start a new sequence, keeping the allocations.