Add challenge 87: Speculative Decoding Verification (Medium)

[claude]

Summary

• Adds Challenge 87: Speculative Decoding Verification (Medium difficulty)
• Implements the core token acceptance/rejection step used in every modern LLM serving framework (vLLM, TensorRT-LLM, etc.)
• Given a batch of B draft sequences with T candidate tokens, the solver must: accept tokens whose target/draft probability ratio passes a uniform random test, resample from the adjusted distribution clamp(q−p, 0) on the first rejection, or sample a bonus token if all T tokens are accepted

GPU learning moments:

• The acceptance chain has a sequential dependency (each position depends on the previous), so naive per-position parallelism doesn't work — solvers must think about parallel scan for finding the first rejection
• The inverse-CDF resampling step is O(V) and naturally parallelizable across the vocabulary dimension
• Batch dimension B provides the main axis for parallelism across sequences

Test cases: 10 functional tests covering single-token edge cases, all-accept (bonus token), forced first-rejection, mixed acceptance, varying V (4–1,000), and a realistic batch. Performance test: B=64, T=8, V=32,768.

Test plan

• All 6 starter files present (.cu, .pytorch.py, .triton.py, .jax.py, .cute.py, .mojo)
• pre-commit run --all-files passes (Black, isort, flake8, clang-format)
• Validated with run_challenge.py --action submit — all functional + performance tests pass on NVIDIA Tesla T4
• Checklist in CLAUDE.md verified: HTML starts with <p>, <h2> sections, example matches generate_example_test(), SVG visualization included, performance bullet matches generate_performance_test()

🤖 Generated with Claude Code

[shxjames]