Module sampler 

Token sampling strategies for autoregressive generation.

After the model produces logits (raw scores) over the vocabulary, we need to select the next token. This module provides:

• Greedy decoding: Always pick the highest-probability token.
• Temperature scaling: Control randomness by dividing logits by temperature.
• Top-p (nucleus) sampling: Sample from the smallest set of tokens whose cumulative probability exceeds p, preventing unlikely tokens from being chosen.

Example

let mut sampler = Sampler::new(0.7, 0.9, 42);
let next_token = sampler.sample(&mut logits);

Structs

Sampler

Configurable token sampler with temperature and top-p support.