Changelog¶
Release notes for autoray.
v0.8.11 (2026-06-08)¶
Enhancements:
Added initial MLX backend support including
random.default_rng, dtype-aware creation wrappers,to_numpy,count_nonzero,ravel, andlinalg.svdsupport.Added
infer_backend_device_dtypefor shared backend, device, and dtype inference fromlikevalues.Exported
DoFuncfor reusable call-time auto-dispatch, with fasterdoandget_namespacedispatch paths.Refreshed package metadata, pixi environments, docs infrastructure, and CI, including the
testmlxenvironment and MLX CI job.Reworked backend tests around the unified
XFAILSregistry and expanded linear algebra coverage.
Bug Fixes:
Cached failed
.deviceand.dtypeprobes inget_namespaceto avoid repeated slow exceptions for array classes such as JAX tracers (PR #30).Fixed
autojitwith lazylazy_astypedtype evaluation.Added TensorFlow
astypehandling and a NumPy-like Torchnonzerowrapper.Fixed reduced and batched SVD shape handling in
svd_manual_full_matrices_kwarg.Fixed
prime_factorsto preserve integer factors.
Previous release notes can be found on the project releases page.