clifra

Layout-first Clifford algebra tools for PyTorch.

Clifra exposes one planner-owned algebra host. Full-lane tensors and compact grade layouts share the same algebra, while planning builds static executors for products, metrics, exponentials, actions, layers, and analysis utilities.

Install

uv sync
uv sync --extra dev

Docs:

uv sync --group docs
uv run --group docs mkdocs serve

Minimal Use

import torch

from clifra import make_algebra

algebra = make_algebra(3, 0, device="cpu")
vectors = algebra.layout((1,))
products = algebra.plan_product(
    op="gp",
    left_layout=vectors,
    right_layout=vectors,
    output_layout=algebra.layout((0, 2)),
)

left = torch.randn(8, vectors.dim)
right = torch.randn(8, vectors.dim)
out = products(left, right)

Checks

uv run pytest tests/ -m unit -q --tb=short
uv run pytest tests/ -m "not slow" -q --tb=short
uv run ruff check .
uv run --group docs mkdocs build

Docs

The docs are intentionally small: API pages come from live docstrings, and docs/core-design.md keeps the process snippets.

Research

research/continuum_solver is a research engine built on clifra for fitting stable, fast continuum deformations. It uses diffeomorphism-style coordinate morphing over clifra charts and invertible bivector fields, inheriting clifra's planned algebra execution.

Contribution

Found a problem or want to propose a change? Please open an Issue first, especially before a PR, so the scope is clear.

License

Apache License 2.0. See LICENSE and NOTICE.

Citation

@software{kim2026clifra,
  author  = {Kim, Eunkyum},
  title   = {clifra: Clifford Algebra Layers for PyTorch},
  url     = {https://github.com/Concode0/clifra},
  version = {1.1.0},
  year    = {2026},
  doi     = {10.5281/zenodo.18939518},
  license = {Apache-2.0}
}

References

RotorGadget Layer Architectures

This project implements the equivariant layer architectures derived from the irreducible decomposition of Clifford multivectors.

• RotorGadget implementations found in clifra/layers/primitives/rotor_gadget.py are based on:
  • Pence, T., Yamada, D., & Singh, V. (2025). "Composing Linear Layers from Irreducibles." arXiv:2507.11688.

Optimization on Manifolds

The core solvers and adaptive optimizers in clifra/optimizers/ leverage Riemannian optimization techniques to handle the non-linear constraints of multivector rotors and versor transformations:

• Absil, P.-A., Mahony, R., & Sepulchre, R. (2008). Optimization Algorithms on Matrix Manifolds. Princeton University Press.
• Boumal, N. (2023). An Introduction to Optimization on Smooth Manifolds. Cambridge University Press.