Release notes

[v2.11.0] - 2026-02-24

Added

Serve mode

eOn can now serve any potential over Cap’n Proto RPC using the rgpot protocol. Run eonclient -p lj --serve-port 12345 for the simplest case, or use --serve "lj:12345,eam_al:12346" for multi-model serving. Replicated mode (--replicas N) starts N independent servers on sequential ports; gateway mode (--gateway) places a round-robin dispatcher in front of the pool so clients only need one address.

All options are also available through a [Serve] INI config section for fully config-driven operation. The feature requires -Dwith_serve=true at build time and a Cap’n Proto installation (provided by the serve pixi environment).

See Serve Mode for full documentation.

rgpycrumbs configuration examples

The user guide now includes dictionary-style configuration examples using rgpycrumbs, demonstrating programmatic config generation alongside INI files.

Developer

asv-perch benchmark CI

The benchmark PR comment workflow now uses the asv-perch GitHub Action instead of hand-rolled shell/JS scripts. Benchmark execution for main and PR HEAD is parallelized via a matrix strategy.

Serve mode build infrastructure

Added rgpot subproject wrap, with_serve meson option, serve pixi environment, a CI workflow for serve mode builds, and Catch2 unit tests for serve spec parsing.

Fixed

Windows torch_global_deps

Skip torch_global_deps on Windows where the conda-forge libtorch package does not ship it.

Serve mode AtomMatrix collision

Fixed a segfault caused by a type collision between eOn’s Eigen-based AtomMatrix and rgpot’s custom AtomMatrix. The rgpot::PotentialBase virtual interface was replaced with a flat-array ForceCallback, and the serve code now only links the capnp schema dependency from rgpot.