Struct EngramHasher 

pub struct EngramHasher { /* private fields */ }

The n-gram hash for one sequence: keeps the compressed ids of every position so decode steps see the prefill’s look-back, exactly as the reference’s cache does.

Implementations

impl EngramHasher

pub fn new(t: Arc<EngramHashTables>, max_seq: usize) -> Self

pub fn tables(&self) -> &Arc<EngramHashTables>

pub fn reset(&mut self)

pub fn hash(&mut self, input_ids: &[u32], start_pos: usize) -> Result<Vec<i64>>

Hash input_ids placed at start_pos... Returns [seqlen, n_layers, n_hash_cols] row ids.

pub fn layer_row_ids( &self, hashes: &[i64], tokens: usize, hash_index: usize, ) -> Vec<u64>

The row ids one engram layer needs, in [token][col] order, from a hash() result of tokens positions.