Add challenge 101: Adaptive Layer Normalization (Medium)

[claude]

Summary

Adds challenge 101 — Adaptive Layer Normalization (AdaLN), the conditioning kernel used in modern Diffusion Transformers (DiT, Stable Diffusion 3, FLUX, OpenSora, SiT). For each (b, n) token, the solver layer-normalizes along the feature dimension and then modulates by a per-batch scale and shift vector that is broadcast across the sequence:

out[b, n, d] = LN(X[b, n, :])[d] * (1 + scale[b, d]) + shift[b, d]

The per-batch (rather than per-channel) modulation is what distinguishes AdaLN from BatchNorm/RMSNorm and is the core mechanism by which time-step / class embeddings condition diffusion transformers.

• Inputs: X (B, N, D), scale (B, D), shift (B, D)
• Output: (B, N, D), fixed eps = 1e-5, float32 throughout
• Performance test: B = 16, N = 4,096, D = 1,152 (DiT-XL/2 inspired, ≈ 600 MB working set)

Validated end-to-end against the live platform (T4) with a hand-written CUDA reference solution — all functional + performance tests pass. Reference impl matches torch.nn.functional.layer_norm to within float32 precision on the example.

Test plan

• pre-commit run --all-files passes
• python scripts/run_challenge.py challenges/medium/101_adaptive_layer_normalization --language cuda --action run → example test passes
• python scripts/run_challenge.py challenges/medium/101_adaptive_layer_normalization --language cuda --action submit → all functional + performance tests pass on T4
• Reference impl produces values matching F.layer_norm + per-batch modulation
• All 6 starter files present, each with a single parameter description comment and an empty solve
• HTML example values match generate_example_test() output
• Performance test working set (~600 MB) fits 5× in 16 GB Tesla T4 VRAM
• Number 101 is unclaimed by any open PR (highest existing/PR number is 100)

🤖 Generated with Claude Code