Changelog

Release notes for autoray.

v0.11.0 (2026-08-22)

Enhancements:

  • compose now supplies an AutoNamespace to any default implementation taking a namespace parameter: the namespace itself when called as xp.my_func(...), so its dtype and device defaults apply, otherwise one from the first argument if it matches the dispatched backend. Implementations registered for a specific backend keep their own signatures.

  • Namespaces from get_namespace are now stable objects: registering a function drops their cached lookups rather than discarding the namespaces, so one already held picks up the new function.

  • autoray.lazy "sum", "prod", "min" and "max" now accept keepdims, matching the eager and array API signatures. The reduced axes are kept as size 1 in the inferred lazy shape.

  • Added the lazy reductions "mean", "std", "var", "all", "any", "count_nonzero", "argmin" and "argmax", which also take axis and keepdims, and "cumsum", which takes axis and accumulates over the flattened array when it is None. "argmin" and "argmax" follow the eager convention of only accepting a scalar axis. Backend support for keepdims varies: see XFAILS in tests/conftest.py.

  • The lazy reductions and "cumsum" pass any further keyword arguments, such as ddof or dtype, straight through to the backend function. A LazyArray supplied this way, for example as where, is tracked as a dependency of the result.

Bug Fixes:

  • autoray.lazy reductions now pass axis on as supplied, rather than always normalizing it to a tuple. do("prod", x, axis=0) on a lazy torch array therefore works, since torch.prod only accepts a scalar dim.

v0.10.1 (2026-08-07)

This is a re-release of v0.10.1, see below.

v0.10.0 (2026-08-06)

Enhancements:

  • Added random_array, i.e. do("random.array", ...), for backend-agnostic normal, uniform and rademacher random arrays. It accepts backend-specific generators or seeds, inherits dtype and device from like, and generates complex normal samples with unit total variance. A complex rademacher sample is a choice from the four roots of unity, so it also has modulus one. jax "random.default_rng" with seed=None warns one time, because tracing fixes its generated seed.

  • jax, mlx and tensorflow "random.default_rng" now default to seed=None, and jax also accepts a key. torch "random.default_rng" accepts a torch.Generator, and raises a descriptive error if a device is also requested that the generator cannot supply.

  • Added a random numbers documentation guide, covering random_array, the generator objects and their per backend method coverage, the shared random state, the complex normal convention, and the rules for jax.jit and torch.compile.

Bug Fixes:

  • torch "random.normal" and "random.uniform" now generate at the requested dtype rather than casting afterwards. A complex dtype therefore gets a complex sample rather than a real one cast up, which previously left the imaginary part zero, and float64 now carries full double precision entropy.

  • torch "random.seed" is now registered, and seeds the shared random state. Previously it resolved to torch.random.seed, which takes no seed and thus raised TypeError.

  • tensorflow "random.seed" is now registered, as an alias of tf.random.set_seed. Previously it raised ImportError.

  • mlx now uses mlx’s own shared random state, rather than a separate module level generator of autoray’s. do("random.seed", seed, like="mlx") therefore reaches mx.random.seed, and mixing autoray and raw mlx.core.random calls is now reproducible in both directions.

v0.9.1 (2026-08-03)

Bug Fixes:

  • Registration now removes the cache entries that depend on it. A registration applies immediately, also to functions and classes that autoray used before. Previously, register_function with module=, alias= or wrapper= had no effect if the code called the function first, because only the first import of a function reads these three arguments.

  • register_backend and register_backend_alias now clear the caches that hold the backend of each class. Autoray then finds the new backend, also for a class that it used before.

  • tree_register_container now clears the pytree dispatch caches. Autoray then uses the new functions for a container class, also if it used that class as a leaf before, or if it used the functions of a parent class.

  • Registration also clears the namespaces from get_namespace and the dtype cache of to_backend_dtype. A namespace object that you already have keeps the functions that it found. This is the correct behavior: get a new namespace to use a later registration.

  • autoray no longer shows the internal exception chain in the ImportError that it raises when it cannot find a function.

  • torch "random.default_rng" now defaults to seed=None

v0.9.0 (2026-07-15)

Enhancements:

  • Added to for converting arrays, or nested collections (“pytrees”) of arrays, to a target backend, dtype and/or device, all specifiable in a single string such as "torch-float32-cuda:0", via explicit kwargs, or an example array. Repeated references to the same input array are converted once and share the same output array. Matching torch.nn.Module.to semantics, only floating point and complex arrays are cast when a dtype is given, so e.g. integer index arrays are preserved.

  • Added to_device composed function for moving arrays between devices, with "gpu" accepted as an alias for "cuda" where relevant, and a bare device type such as "gpu" meaning ‘ensure on this type of device’, without migrating arrays between device indices.

  • Added from_numpy composed creation routine for converting a numpy array into a backend array, directly with a given dtype and on a given device where possible, e.g. via a single torch.as_tensor call. An example array supplied as like supplies its backend, dtype and device as defaults.

  • Made to_numpy a composed function with default implementation np.asarray, so that unknown backends are handled automatically.

  • Reworked register_function into the single entry point for all function-level registration, taking module=, alias=, wrapper=, inject_dtype= and inject_device=, deprecating register_creation_routine.

  • Dtype name resolution now handles scalar types such as np.float32 and builtins such as float and complex, anywhere a dtype is specified.

  • MLX: array and asarray now preserve the input array’s dtype rather than applying mlx defaults, e.g. no longer silently downcasting float64 to float32.

  • MLX: count_nonzero now uses the native mlx implementation where available (v0.32+), falling back to a manual version which also supports the axis and keepdims kwargs.

  • The "backend[alt]" fallback, used when a function is missing from an older version of a library, now also finds directly registered implementations, and caches its result against the original backend so the lookup only happens once.

  • MLX: requesting a device warns that mlx arrays live in unified memory with per-op computation placement, rather than failing generically.

  • Lazy: to_numpy and to_device on LazyArrays now raise explicit errors, while from_numpy creates a lazy leaf node.

  • Added the testcupy pixi environment, and refreshed the contributing and developer guides, including a code of conduct and AI policy.

Bug Fixes:

  • Fixed get_namespace caching for backends with unhashable device objects such as cupy.cuda.Device.

  • Added cupy linalg.cholesky support for the upper kwarg via the new generic cholesky_manual_upper wrapper.

  • Torch: random.default_rng now only inherits floating point dtypes from the like argument and creates generated arrays on the same device as its generator.

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, and linalg.svd support.

  • Added infer_backend_device_dtype for shared backend, device, and dtype inference from like values.

  • Exported DoFunc for reusable call-time auto-dispatch, with faster do and get_namespace dispatch paths.

  • Refreshed package metadata, pixi environments, docs infrastructure, and CI, including the testmlx environment and MLX CI job.

  • Reworked backend tests around the unified XFAILS registry and expanded linear algebra coverage.

Bug Fixes:

Previous release notes can be found on the project releases page.