Add challenge 121: Fused Logit Penalties (Medium)

[claude]

Summary

Adds challenge 121: Fused Logit Penalties (Medium) — the logit-processor stage that vLLM, SGLang and HF TGI run on every decode step, right before sampling.

Given a batch of B in-flight requests, each with a V-wide logit row, its prompt/generated token histories (-1-padded), and its own penalty coefficients, the solver produces the penalized logits:

1. Repetition (multiplicative, keyed on prompt ∪ generated tokens): z /= r if z > 0, else z *= r
2. Frequency (keyed on generated counts): z -= frequency_penalty * count
3. Presence (flat, keyed on generated presence): z -= presence_penalty if count > 0

Why this is a worthwhile GPU problem

V (up to 131,072) dwarfs the token histories, so the obvious "scan the history for each vocabulary entry" approach is hopelessly slow. A good solution builds the per-request occurrence table with a scatter (atomics or privatized counters), then streams the logits in a single bandwidth-bound pass. Padding slots and the asymmetry between the three penalties (two different "seen" predicates) add real bookkeeping. It is not an element-wise map.

No overlap with existing sampling challenges (29 top-k, 60 top-p, 104 min-p) — those select/renormalize, this one is a gather/scatter histogram problem — and no overlap with any open PR topic or number.

Contents

• challenge.py — reference impl (standard PyTorch scatter_add_ / where, CUDA + XLA safe), 10 functional tests (edge sizes 1–4, powers of 2, 100/255 non-powers of 2, zeros, negatives, all-padding rows, penalty no-ops, realistic 32k-vocab decode), performance test at B=256, V=128,256, P=1,024, G=512 (~264 MB of tensors)
• challenge.html — description, SVG walk-through of the scatter → count table → apply pipeline, worked example, constraints
• Starters for all 6 frameworks (empty solve, medium convention)

Validation

• python scripts/validate_challenges.py → 0 errors
• pre-commit run --all-files → all hooks pass
• A CUDA solution was validated against the live platform on a Tesla T4: --action run ✓ and --action submit ✓ ("All tests passed"). The solution file is not committed.

🤖 Generated with Claude Code