Top-k & Top-p (Nucleus) Filtering

Before sampling the next token, LLM decoders usually filter the probability distribution.

Implement filter_probs(probs, top_k=None, top_p=None):

1. If top_k is set, keep only the k most probable tokens.
2. If top_p is set, keep the smallest set of most-probable tokens whose cumulative probability is $\ge p$.
3. Zero out everything else and renormalise so the result sums to 1.

Apply top_k first, then top_p on the result. Return a NumPy array the same shape as probs.

Example

filter_probs(np.array([0.5, 0.3, 0.15, 0.05]), top_p=0.8)
# array([0.625, 0.375, 0.0, 0.0])