sutro-problems

Research archive: D4 cold-compilation repair

Research inclusion; D4 qualification remains unresolved. The original author group below passed 3/3 (median 985.074 ms). The separate initial reviewer group passed 1/3: two complete runs missed D4 by 64 and 4 labels, with all 45 timed calls retained. There is no qualifying reviewer-group median and no leaderboard promotion. The runtime source and original author records/report are preserved. Reviewer evidence also separates the earlier source’s warm-up failures from this repaired source. Submitted by @sjbaebae; the complete upstream lineage is not settled by this documentation-only integration.

The earlier source failed two independent review runs at the second-call90-second KMNIST warmup limit. Historical local passes remain preserved in the parent directory. Independent review.

This replacement keeps the exact teacher/student learner and all archived CUDA kernels unchanged. C++ binding headers use tensor types and pybind casters, with host-O0/-g0 flags, to reduce cold compilation cost. The same exhaustive GEMM tuner remains. All three canonical Modal sandbox runs completed warmup and passed accuracy, timing and energy gates. Runtime attribution between compilation and tuning was not independently measured.

Run Error Ranked ms Above-idle mJ
1 2.227273% 985.074 286669.683
2 2.220000% 984.546 283665.268
3 2.228182% 986.252 296628.249

Median 985.074ms, median energy 286669.683mJ, aggregate error 2.225152%. Three distinct boards, unchanged source. No failing qualification run replaced.

Initial no-energy lean/bounded checks passed. The pruned-compilation alternative completed warmup but missed accuracy by two predictions; its failed record is retained and it is not the submitted source.

Source SHA256 c5ed1d8ffb115cc58a6a6b4ac468b07c780d773b3f8d7387ec3f295b6f1b3eb1, 19538bytes; source-only archive without fitted data. CUDA graphs/cuBLAS and custom kernels, not a whole-model megakernel. This PR does not itself change the leaderboard.

python run_modal.py submissions/native-mlp-d4-20261007/leancompile-repair/submission.py:classify --difficulty 4 --runs 3 --json out/d4-leancompile