Align posterior weights¶

align post-processes neural-network weight samples so that networks which differ only by known, function-preserving symmetries become comparable in weight space. Canonicalization removes exact continuous symmetries; matching aligns the remaining discrete and matrix-valued ones against a common reference.

Start here¶

• Theory and concepts: vocabulary, mathematical picture, supported architectures, canonicalizers, objectives, and solvers.
• Configuration reference: algorithm options.
• Inputs, fitting and artifacts: in-memory inputs, reference fitting, and persistence boundaries.

Architecture reference¶

These pages mirror the recipes in align/architectures:

• MLP
• ConvNet
• Residual ConvNet
• LayerNorm MHA Transformer
• RMSNorm, GQA and RoPE Transformer

Algorithm details that apply across architectures live in Matching objectives and solvers.