Script 'mail_helper' called by obssrc Hello community, here is the log from the commit of package python-autoray for openSUSE:Factory checked in at 2026-11-18 18:30:34 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Comparing /work/SRC/openSUSE:Factory/python-autoray (Old) and /work/SRC/openSUSE:Factory/.python-autoray.new.1251 (New) ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "python-autoray" Wed Nov 18 18:30:34 2026 rev:16 rq:1398254 version:0.11.0 Changes: -------- --- /work/SRC/openSUSE:Factory/python-autoray/python-autoray.changes 2026-06-15 19:45:04.651521693 +0200 +++ /work/SRC/openSUSE:Factory/.python-autoray.new.1251/python-autoray.changes 2026-11-18 18:31:14.325049789 +0100 @@ -1,0 +2,67 @@ +Tue Nov 17 18:50:23 UTC 2026 - Dirk Müller <[email protected]> + +- update to 0.11.0: + * `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. + * 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`. + * `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. + +------------------------------------------------------------------- Old: ---- autoray-0.8.11.tar.gz New: ---- autoray-0.11.0.tar.gz ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Other differences: ------------------ ++++++ python-autoray.spec ++++++ --- /var/tmp/diff_new_pack.nZEs6l/_old 2026-11-18 18:31:15.248088894 +0100 +++ /var/tmp/diff_new_pack.nZEs6l/_new 2026-11-18 18:31:15.250088979 +0100 @@ -17,13 +17,13 @@ Name: python-autoray -Version: 0.8.11 +Version: 0.11.0 Release: 0 Summary: A lightweight python automatic-array library License: Apache-2.0 URL: https://github.com/jcmgray/autoray Source: https://files.pythonhosted.org/packages/source/a/autoray/autoray-%{version}.tar.gz -BuildRequires: %{python_module base >= 3.10} +BuildRequires: %{python_module base >= 3.11} BuildRequires: %{python_module dask-array} BuildRequires: %{python_module hatch-vcs} BuildRequires: %{python_module hatchling} ++++++ autoray-0.8.11.tar.gz -> autoray-0.11.0.tar.gz ++++++ ++++ 36254 lines of diff (skipped)
