Struct LogitBias 

pub struct LogitBias { /* private fields */ }

Logit-bias stage — applies per-token additive biases and hard bans.

This stage must be positioned before temperature scaling, repetition penalty, and any filtering stages so that bans and biases influence all downstream steps uniformly.

Processing order (matches sampling::mod::apply_logit_bias_and_banned_tokens):

1. Banned tokens → f32::NEG_INFINITY (hard ban, cannot be overridden by bias).
2. Logit biases are added to surviving logits.

Implementations

impl LogitBias

pub fn new(biases: HashMap<u32, f32>, banned: Vec<u32>) -> Self

Create a new logit-bias stage.

biases maps token IDs to additive values (positive = boost, negative = suppress). banned is the list of tokens to hard-ban.

pub fn banned_only(banned: Vec<u32>) -> Self

Create a stage with only hard-banned tokens and no biases.

pub fn biases_only(biases: HashMap<u32, f32>) -> Self

Create a stage with only biases and no bans.

Trait Implementations

impl SamplerStage for LogitBias

fn apply(&self, logits: &mut Vec<f32>, _recent_tokens: &[u32])

Apply this stage to the logit vector in place.

fn name(&self) -> &'static str

Human-readable name for logging / debugging.