Tensor0

Tensor0 is an experimental symmetric tensor library for Python, with Rust-backed structural metadata and JAX-backed storage and automatic differentiation. It is not yet a production-ready tensor-network framework. Before 1.0, public API patch releases remain backward-compatible; breaking changes require a minor release and migration notes.

Use it for tensor algebra with conserved charges and other supported symmetries: work with symmetry-allowed blocks, then compose, contract or differentiate them using JAX. It provides tensor primitives rather than a complete simulation algorithm framework.

Documentation · Installation · Usage · API · Support policy

Install from source

0.1.0 is a release candidate, not a published release. Source installation remains the documented distribution route. The candidate wheel target is Linux x86_64 CPU, CPython 3.11–3.14, pending successful matrix verification; this is not a claim of available wheels or PyPI publication.

The numerical baseline is Linux CPU, Python 3.11, JAX and JAXLIB 0.10.1. Source builds require Rust 1.87 or newer and a C++20-capable compiler. The native XLA FFI checks the JAX/JAXLIB versions exactly. Only recorded passing candidate runs establish interpreter coverage.

git clone https://github.com/yitan1/Tensor0.git
cd Tensor0
python3.11 -m venv .venv
source .venv/bin/activate
CARGO_BUILD_JOBS=1 python -m pip install .

Native compilation can use several GiB per heavy C++ unit. Serial compilation limits concurrency, not peak memory; see build prerequisites and memory limits. CPU builds do not require CUDA. Optional CUDA support is experimental and restricted to selected operations, dtypes and AD paths—not the full TensorMap API or full reverse AD.

A small symmetric tensor

import jax.numpy as jnp
from tensor0 import TensorMap, U1Irrep, hom, space, storage_dim

v = space(U1Irrep, {0: 2, 1: 3})
h = hom((v,), (v,))
tensor = TensorMap(h, jnp.arange(storage_dim(h), dtype=jnp.float32))

for sector, block in tensor.blocks():
    print(sector, block.shape)

Tensor0 provides graded spaces and built-in sector families, ordinary and diagonal tensor maps, blockwise linear algebra and truncation, transforms and symmetry-aware contractions, and JAX jit/grad integration. See the usage guide, contraction guide, and executable examples for capabilities and boundaries.

Contributing and benchmarks

See CONTRIBUTING.md for development and verification links, CHANGELOG.md for candidate changes, and the release checklist for installed-artifact acceptance. Reproducible benchmark suites and their reports document measured configurations and limitations; they are not performance guarantees.

Acknowledgments

Tensor0's main design and architecture are based on TensorKit.jl, bringing these ideas to Python with JAX-backed storage and automatic differentiation.

For citation metadata, see CITATION.cff.