Skip to content

Tensor0 0.3.0

Latest
Latest

Choose a tag to compare

View all tags
@yitan1 yitan1 released this 14 Oct 01:49
· 4 commits to main since this release
v0.3.0
974153a

Tensor0 0.3.0

Tensor0 0.3.0 upgrades the pinned JAX runtime and fixes homogeneous tensor pytree batches.

Changes

  • Fix homogeneous TensorMap, DiagonalTensorMap, and SectorVector batches returned by jax.vmap, including nested batching and whole-batch scalar-loss gradients (issue #1).
  • Require Python 3.12+ and exact JAX/JAXLIB 0.11.2, with matching native FFI headers and runtime identity checks.
  • Provide ordinary CPython 3.12, 3.13, 3.14, and 3.15 wheels for Linux x86_64 (glibc 2.28+) and macOS 15+ ARM64. Free-threaded Python remains unsupported.
  • Upgrade source-build bindings to PyO3/Rust numpy 0.29; require maturin >=1.15,<2.
  • Retain the optional, exactly version-paired tensor0-cuda12 0.3.0 companion and the existing bounded experimental CUDA12 scope.

Migration

Python 3.11 users must remain on Tensor0 0.2.0 with JAX/JAXLIB 0.10.1, or create a Python 3.12+ environment. Do not independently upgrade JAX underneath an old Tensor0 native extension. Install matching core/addon versions; source users must rebuild native binaries against the new headers.

python -m pip install tensor0==0.3.0
# Experimental Linux CUDA12 installation:
python -m pip install 'tensor0[cuda12]==0.3.0'

Validation and scope

Candidate source: 974153a7ba26d11d5f6f51d3f8d5d01abecc0f30.

CPU full suites run on Python 3.12/3.15 on both platforms; 3.13/3.14 wheels pass installed smoke/examples and bounded native JIT/AD contracts. CUDA publication additionally requires exact-pair A100 acceptance; the original report/log archive is attached as cuda-acceptance.tar.gz. CUDA support remains experimental and limited to the documented operation/dtype/AD subset, not general TensorMap GPU coverage, full reverse AD, multi-GPU, or a speed guarantee.

A bounded single-core desktop comparison found approximately 6–7% higher steady-state latency in two small complex128 contraction cases, with faster compilation; several allocation-sensitive timings were inconclusive. This release does not claim a performance improvement.

All eleven distributions are the original accepted artifacts: eight core wheels, one core sdist, one repaired addon wheel, and one addon sdist. No rebuild or retagging occurs during publication.

Assets 16
Loading