Installation

Install from PyPI

The current PyPI release is Tensor0 0.2.0, published on 2026-10-12.

Published 0.2.0 platform availability

Platform	Prebuilt core wheels	Status
Linux x86_64, glibc >=2.28	CPython 3.11–3.14, manylinux_2_28	Available and tested
macOS 15+ ARM64	CPython 3.11–3.14, macosx_15_0_arm64	Available and tested on hosted macOS 15
Windows	Not provided	Outside the current support scope

All eight core wheels were built from one shared sdist and passed installed-wheel acceptance. macOS is CPU-only: do not install the Linux-only cuda12 extra there. The macOS wheels are ARM64, not Intel or universal2, and require macOS 15 or newer. See Releasing for the artifact acceptance policy.

Other platforms have not been validated. Lack of a wheel is not proof that source builds are impossible; neither is it a guarantee that they work.

Use a virtual environment:

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install tensor0==0.2.0
python -m pip check

Installing these prebuilt wheels requires no Rust or C++ compiler and no CUDA toolkit. JAX and JAXLIB 0.10.1 are installed through Tensor0's pinned dependencies. Native handlers use matching vendored XLA FFI headers and reject mismatched runtime versions. Do not independently upgrade JAX/JAXLIB.

Python >=3.11 package metadata does not establish support for future interpreters or other platforms. Documentation on main describes the current checkout; consult the matching release documentation when using an older artifact.

Source build prerequisites

These tools are needed only when building from source, not when installing a supported wheel:

• Rust/Cargo 1.87 or newer, and a linker and standard system build tools (including ar).
• A C++20-capable compiler and its standard library. The build defaults to /usr/bin/c++; set CXX to select another compiler. CXXFLAGS is not used.
• A supported Python interpreter and access to the pinned Python dependencies and Cargo crates, or suitable local caches. The source-build procedure below has been validated on Linux x86_64 and macOS ARM64.

Native C++ template compilation is memory-intensive: heavy units can each use several GiB. Cargo's default parallelism follows CPU count, not free memory or container limits. Start with CARGO_BUILD_JOBS=1 on constrained machines and increase only within the available memory budget. Even a serial build may exceed a small container's limit; there is no universal minimum-RAM or successful-build guarantee. A killed compiler can indicate an out-of-memory limit—inspect system or container logs before retrying. Compiler jobs are unrelated to runtime worker threads. See Development for profiles, optimization overrides, cache behavior and jobserver details.

Source install

Use a virtual environment with the baseline interpreter:

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

This is a non-editable source build. It requires access to Python build/runtime dependencies and Cargo crates, or suitable local caches. No CUDA toolkit is needed for the default CPU-only extension. For editable contributor setup, follow Local Setup instead.

Check the installed package

Run outside the checkout, with the virtual environment still active, to avoid accidentally testing source-tree imports:

cd /tmp
JAX_PLATFORMS=cpu python - <<'PY'
from importlib.metadata import version
import jax
import tensor0
from tensor0 import U1Irrep, identity, permute, space

print("Tensor0:", version("tensor0"), tensor0.__file__)
print("JAX/JAXLIB:", version("jax"), version("jaxlib"))
print("Devices:", jax.devices())
v = space(U1Irrep, {0: 2, 1: 1})
t = identity(v)
r = jax.jit(lambda x: permute(x, ((1,), (0,))))(t)
r.storage.data.block_until_ready()
print("Installed CPU smoke OK:", r.storage.data.shape)
PY

This is a smoke check, not comprehensive numerical acceptance. See the release checklist for artifact validation and the support policy for execution and API boundaries.

Experimental CUDA

Version 0.2.0 provides a separate tensor0-cuda12 companion with an exact matching core version. Install the published pair using:

python -m pip install "tensor0[cuda12]==0.2.0"
python -m pip check

CUDA support currently targets Linux x86_64 and remains experimental. Installing GPU-enabled JAX alone does not supply the Tensor0 companion. Prebuilt companion wheels do not require a local CUDA toolkit; compatible NVIDIA hardware, a driver and the specified runtime dependencies are needed for execution. Building the companion from source additionally requires the CUDA toolkit and host compiler described in the development guide. See CUDA support for the tested operations, dtypes and AD limits.