Tensor0

Tensor0 is an experimental symmetric tensor library with Rust-backed structural metadata and JAX-backed tensor storage.

Use this documentation for the current checkout. The public API is still evolving, and Tensor0 is not yet a production-ready tensor-network framework. Tensor0 0.2.0 is available on PyPI, published on 2026-10-12. Install with pip install tensor0==0.2.0 in a virtual environment; supported prebuilt wheels require no Rust or C++ compiler.

Runtime Compatibility

The released CPU execution baseline supports Linux x86_64 with glibc >=2.28 and macOS 15+ ARM64, with JAX and JAXLIB 0.10.1. The tested 0.2.0 wheel matrix covers CPython 3.11–3.14 on both platforms, built from one shared core sdist; Python 3.11 remains the baseline interpreter. The Python >=3.11 metadata does not imply support for future interpreters. macOS is CPU-only. See Support and compatibility for platform scope.

The native stride handler is compiled against that exact XLA FFI header version and rejects mismatched runtimes. Missing native routes are reported explicitly rather than hidden by an element-address fallback. Native stride operations lower through JAX as stablehlo.custom_call operations. Tensor0 does not ship a separate pure-StableHLO stride backend.

Experimental CUDA support covers selected operations, dtypes and AD paths, not the full TensorMap API or full reverse AD.

Start Here

• Installation: PyPI wheels, source builds and prerequisites.
• Support and compatibility: runtime scope and pre-1.0 API policy.
• Usage Guide: current public API examples and boundaries.
• Contractions: primitive and tensor-network contraction interfaces.
• Examples: executable ordinary, symmetric and contraction scripts.
• Development: contributor setup and detailed verification.
• Release checklist: installed-artifact acceptance before publication.

Current Scope

Tensor0 currently covers typed sector metadata, graded spaces, TensorMap storage, block access, composition, SVD helpers, transforms, symmetry-aware contractions, deterministic morphism constructors, explicit tensor products and random construction, inverse/pseudoinverse, direct solves, QR/LQ- and SVD-backed orthogonalization, full/truncated Hermitian eigendecomposition, numerical structure predicates, monoidal-unit insertion and removal, correctness-first dense conversion for symmetric sectors, direct immutable array execution for Trivial tensors, and JAX jit / grad smoke gates.