Add challenge 110: Attention with Sinks (Medium)

[claude]

Summary

• Adds challenge 110 (Medium): Attention with Sinks, a StreamingLLM-style attention variant where each query attends to a small set of leading sink tokens plus a causal sliding window of the most recent tokens.
• Real-world inference kernel used to bound the KV cache for long-context streaming while preserving the anchor tokens that stabilize the softmax distribution.
• Distinct from the existing sliding-window, causal, and window-cache challenges: the mask combines a fixed leading prefix (sinks) with a moving window plus a causal constraint, producing a two-region attention pattern that neither sliding-window nor plain causal attention captures.

Files

• challenge.py — reference implementation, ctypes signature, 10 functional tests, perf test (M=5000, d=128, num_sinks=4, window_size=1024).
• challenge.html — problem description, SVG mask visualization, example matching generate_example_test, constraints with performance-test parameters.
• starter/ — CUDA, PyTorch, Triton, JAX, CuTe, Mojo starters that compile but do not solve.

Test plan

• Reference impl runs on all 10 functional tests locally (CPU, all outputs finite).
• Example output values in challenge.html match reference_impl numerically.
• pre-commit run --all-files passes (black, isort, flake8, clang-format, mojo).
• Sample CUDA solution submitted via scripts/run_challenge.py → All tests passed on Tesla T4 (functional + performance).
• Reference impl uses only standard PyTorch ops (works on CUDA and XLA — no torch.cuda.*, no CUDA-only kernels).
• All tensor allocations use device=self.device.

🤖 Generated with Claude Code