Noether v2.0.0 adds key-value caching to AB-UPT, so repeated queries reuse the anchor and geometry encoding

Emmi AI publishes Noether v2.0.0, with two breaking changes. AB-UPT gains key-value caching: the keys and values of the anchors and of the geometry encoding are computed once and reused in later query-only forward passes, so the attention blocks now return an (output, cache) pair and geometry and anchor inputs become optional. The second moves the creation of sub-configurations into the config classes. Mixed attention accepts a mask over keys, which lets anchor sequences of different lengths be padded into one batch. A `noether-init` command generates a starter project for a chosen model and dataset, another validates a configuration and launches a SLURM job, and training can resume without loading data.

Share on LinkedIn, X or Bluesky, or copy the link.