Add sliding windows to the off-graph KV-cache reference

[kiymetakdemir]

Summary

Adds a sliding window to the eager KV-cache reference. CacheConfig.window takes an int for every layer or one entry per layer, so a model that alternates policies (gemma-style) is expressible.

Both reference caches honour it. ContiguousReferenceCache adds a lower bound to the band it already emits, and windowing alone is now enough to make a step EXPLICIT. CellReferenceCache memoizes the mask per window rather than per layer: occupancy, membership and causality are common to every layer of a step, and only the lower bound moves, so a mixed model pays two masks per step instead of one per layer.

The reference masks; it does not evict. Cells stay allocated under a window, because reclaiming them when layers disagree needs one cache per policy group.

Files

• extension/llm/cache/reference_cache.py — CacheConfig.window + window_for, the windowed band in ContiguousReferenceCache._spec, and _CellStepPlan split into a shared base plus a per-window memo.
• extension/llm/cache/test_update_and_attend.py — 8 new tests.

Testing

pytest extension/llm/cache/test_update_and_attend.py

New coverage: a window narrowing each query without crossing sequences, layers windowing independently, layers sharing a window sharing one mask, windowed decode attending only retained cells, and a window wider than the history staying fused.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21901

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit befbca3 with merge base c461421 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.