$ the-wire · showcase
JAX exposes Pallas under jax.pallas and fixes hypot/atan2 and log_ndtr derivative pathologies
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX's numerical core got the day's substantive work: Pallas moves out of experimental onto a stable namespace while three separate fixes repair derivatives and tails that quietly returned zeros, infinities, and NaNs.
Expose the Pallas, TPU, TPU SparseCore, Mosaic GPU, and fuser APIs under `jax.pallas` (`jax.pallas`, `jax.pallas.tpu`, `jax.pallas.tpu_sc`, `jax.pallas.mosaic_gpu`, and `jax.pallas.fuser`). google/jax
Pallas, TPU, TPU SparseCore, Mosaic GPU, and fuser APIs now live at jax.pallas, jax.pallas.tpu, jax.pallas.tpu_sc, jax.pallas.mosaic_gpu, and jax.pallas.fuser, with jax.experimental.pallas retained as a backwards-compatible alias. triton and ops stay in jax.experimental.pallas, so a blanket import rewrite will break those two.
Fix cancellation and overflow in hypot and atan2 derivatives. google/jax
jnp.hypot's custom JVP evaluated the tangent as (dx1 - u1 * dr) / r, cancelling 1 - u1^2 to 0 when one operand dominated and collapsing second and higher derivatives of r to 0; lax.atan2_p's derivative divided by x^2 + y^2 and overflowed to 0 or underflowed to inf/NaN well inside the normal floating-point range. Both are corrected, so code differentiating through these ops gets real gradients i...
Fix precision and premature overflow in jax.scipy.special.log_ndtr and its JVP. google/jax
In float64, _log_ndtr_lower ran a fixed 3-term series (series_order=3) whose omitted 7!! / x^8 term lost up to ~141,500 ULP near x = -20, and below the negative cutoff it computed square(x) and overflowed early to +inf. Precision in the negative tail improves without any call-site change.
Deprecate jax.scipy.special.bessel_jn google/jax
jax.scipy.special.bessel_jn is deprecated in favor of jax.scipy.special.jn, which the same-day implementation covers for static integer orders and real/complex arguments. Migrate call sites now; jn reduces negative orders via J_{-n}(z) = (-1)^n J_n(z) and maps z into the first quadrant.
fix(bigtable): validate DirectRow and table match in MutationsBatcher.mutate (#18874) googleapis/google-cloud-python
Long tail: bigtable's MutationsBatcher.mutate and mutate_rows now reject non-DirectRow inputs and rows bound to a different table, restoring the validation Table.mutate_rows has, which a prior refactor to wrap self.table._table_impl.mutations_batcher dropped; google-api-core aligns HTTP T3 and T4 OpenTelemetry span status and attributes, shipping in gapic-generator 1.44.1; google-cloud-compute ...
Action items