Add challenge 104: Min-P Sampling (Medium)

[claude]

Summary

• Adds challenge 104: Min-P Sampling, a real-world LLM logit-filtering primitive used in vLLM, Hugging Face TGI, and llama.cpp.
• For a batch of logits [B, V], the kernel computes softmax probabilities, masks tokens whose probability is below min_p * max_prob in that row, and renormalizes the survivors.
• Unlike top-k or top-p, no sort is required — the threshold adapts to how peaked each row's distribution is, so the challenge becomes a clean two-pass per-row reduction (max → masked sum → renormalize).

What's included

• challenge.py with reference_impl, get_solve_signature, and example/functional/performance test generators (10 functional tests covering edge cases, min_p=0/0.99/in-between, tied maxima, peaked vs uniform distributions, power-of-2 and non-power-of-2 vocabs, realistic 16×32000).
• challenge.html with description, requirements, worked 2×4 example matching generate_example_test(), and constraints including the performance test sizing bullet.
• Six starter files (CUDA, PyTorch, Triton, JAX, CuTe, Mojo) following the medium-difficulty parameter-comment convention.

Performance test

B = 64, V = 128,000, min_p = 0.05 (~62.5 MB per buffer × 2 buffers; well under the 5× / 16 GB T4 budget).

Test plan

• pre-commit run --all-files passes
• Reference implementation verified locally against the example and all 10 functional tests (every row sums to 1)
• CUDA reference solution validated via scripts/run_challenge.py ... --action submit on T4 — all tests passed

🤖 Generated with Claude Code