Contributing GPU scaling results

Thank you for contributing compute. The most valuable contribution is a reproducible run, including a negative result.

Fork the Hugging Face repository and record the commit hash used for training.
Install with pip install -e '.[benchmark]' and run pytest -q.
Do not change a named tier config for an official comparison. New configs belong under configs/community/ with a descriptive name.
Run at least seeds 7, 17, and 27 when compute permits. Use AMP consistently and report whether it was enabled.
Submit generated JSON result files and a short note with GPU count/model, VRAM, CUDA/driver, any OOMs, and code changes.
For new baselines, match data split, augmentation policy, epochs, and evaluation metrics. Report parameter count, throughput, peak memory, wall time, accuracy and NLL.
Do not replace or cherry-pick only favorable seeds. Report all attempted runs.

Suggested PR title: results: CIFAR-10 base on 4x H100, seeds 7/17/27.