pub fn apply_lz_penalty(logits: &mut [f32], history: &[u32], penalty: f32)Expand description
2026-09-25: LZ penalty over the last LZ_WINDOW (256) tokens of history.
For each token in that window and each n-gram length 3, 4 and 5, count the
earlier n-grams that equal the last n-1 tokens followed by that token, and
subtract penalty * (ngram_len - 2) * count from its logit. Token ids at
or above logits.len() are skipped.