Implement Top-p (Nucleus) Sampling in PyTorch
Mediumllm-decodingauto-graded
Sort logits, compute cumulative probabilities, mask tokens below the nucleus threshold, and sample from the filtered distribution.
Solve it
Check your answer
The grader verifies properties of your implementation, so a correct solution written differently from ours still passes.
pip install torchleet
from torchleet import check
check("top-p-sampling", top_p_sample)