Repetition Penalty (CTRL)

Apply the repetition penalty to a logits vector. For each token id that has appeared in prev_tokens:

• If logits[t] > 0: logits[t] /= penalty
• Else: logits[t] *= penalty

This pushes already-seen tokens toward zero either way (when penalty > 1).

Signature: def apply_repetition_penalty(logits: np.ndarray, prev_tokens: list, penalty: float) -> np.ndarray

Return a new array; do not modify the input. Each unique token in prev_tokens is penalized exactly once (duplicate ids do not stack).