Add challenge 103: Repetition Penalty Logit Processor (Medium)

[claude]

Summary

• Adds a new medium challenge: Repetition Penalty Logit Processor, the in-place sampler step used by LLM inference engines (HuggingFace Transformers, vLLM, TGI) to discourage repeating tokens already seen in the prompt or prior generation.
• Per-batch, scales logits[b, v] by 1/penalty (positive logits) or penalty (negative logits) for every unique token id v present in input_ids[b]. Duplicate ids count only once.
• Tests the solver's ability to avoid the naive O(B·V·T) per-thread scan by building a per-batch presence bitmap (O(B·(V+T))).

Test plan

• Reference impl verified to match the HuggingFace RepetitionPenaltyLogitsProcessor semantics on hand-checked example.
• 10 functional cases cover edge sizes (B=1, V=4–5, T=1), zero logits, mixed signs, multi-batch, non-powers-of-two, and realistic vocabulary sizes (V=32k).
• Performance test uses B=64, V=131,072, T=4,096 (≈33MB logits + 1MB ids) — Llama-class vocab, fits 5× in 16GB Tesla T4 VRAM.
• pre-commit run --all-files passes.
• CUDA solution validated against the live T4 runner via scripts/run_challenge.py --action submit — all tests passed.

🤖 Generated with Claude Code