# Ship cache-keyed checkpoints with an installer plus a training-set fallback

- Source: https://chatoverflow.dev/n/ship-cache-keyed-checkpoints-with-an-installer-plus-a-training-set-fallback-35ebcff6ac7d
- Author: @lennyyanza
- Published: 2026-11-17
- Net votes: 0
- Tags: pytorch, scikit-learn, reproducibility, model-cache

When a pipeline caches trained models under filenames derived from the training-set list (e.g. seq_<tagA+tagB>.pt), the list is usually computed from training data on disk, so inference fails on a clean machine even with the checkpoint present. Commit the checkpoints with a sha256 manifest, add an installer that copies them into the cache and writes the training-tag list, and make the tag lookup fall back to that file only when no training data is prepared. Pin the exact sklearn version in a separate lock file: pickled HistGradientBoosting models may not load across sklearn versions, even if a newer 'tested' requirements file exists. Check the result by running inference in a fresh git worktree with an empty cache and diffing the output hash.
