$ the-wire · showcase
JAX adds a legacy escape hatch for random.uniform, google-cloud-python drops Python 3.10
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX is quietly changing the number grid under jax.random.uniform and the numerics tests that police it, while google-cloud-python turns off Python 3.10 support.
jax.random.uniform: Add a new higher-resolution shift-and-scale conversion, protected by a new `jax_legacy_random_uniform` flag google/jax
jax.random.uniform now uses a shift-and-scale conversion that yields one extra bit of randomness in the least-significant place, doubling the grid resolution, at neutral to marginally improved performance depending on hardware and dtype. The change is gated behind a new jax_legacy_random_uniform flag for code that cannot tolerate different sample values.
fix: drop support for end-of-life Python 3.10 googleapis/google-cloud-python
google-cloud-python is dropping support for end-of-life Python 3.10. Environments still pinned to 3.10 will need to move to a supported interpreter before taking new client library releases. The commit confirms the same change.
Add `Domain` support to `numerics_test_util` and bound `AtanhGradTest` to `[-1, 1]`. google/jax
numerics_test_util gains a composable Domain bit-pattern range set (Domain.interval, Domain.points, Domain.infs, Domain.nans, and related constructors), and AtanhGradTest is now bounded to [-1, 1]. The old test evaluated jax.grad(jnp.atanh) across all floats in (-inf, +inf), overflowing to -inf and producing subnormal 1 / (1 - x^2) mismatches on GPU and in float16 on CPU at |x| > sqrt(finfo.max...
Fix small-input `scipy.special.i1e` reference rounding in `bessel_test` and enable `bfloat16` subnormals on GPU. google/jax
bessel_test now uses a 3-term Taylor expansion (0.5 * x * (1 - |x| + 0.625 * x^2)) for tiny |x| instead of scipy.special.i1e, whose float64 Chebyshev sum at 0 lands 1 ULP too high at bfloat16 subnormal midpoints and caused spurious > 0.5 ULP failures on GPU. bfloat16 subnormals are also enabled on GPU.
[partitioning] Lower collective-free dynamic_slice, dynamic_update_slice, and gather. google/jax
Collective-free shard_map partitioning rules are wired for dynamic_slice, dynamic_update_slice, and gather, so these ops no longer need collectives under shard_map. The rest of the day is test and toolchain plumbing: libunwind per-FDE EH frame registration on Linux libc++ builds, plus python-genai picking up usage_metadata and truncated on EmbedContentResponse, proactive_video in ProactivityCon...
Action items
- → Migrate google-cloud-python consumers off Python 3.10 to a supported interpreter googleapis/google-cloud-python [plan]
- → Set jax_legacy_random_uniform if you depend on existing jax.random.uniform sample values google/jax [plan]