RepoJournal

$ cat google/week/2026-10-05.log

Google

Google

the week in review · Oct 5 – Oct 11, 2026

jnp.hypot accuracy rework lands as jaxlib preps stable ABI

By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology

♥

JAX replaced hypot's division-based formula with power-of-two scaling and a custom JVP rule, while jaxlib's C API calls moved off the tuple-and-list setters.

215 commits 197 PRs merged 14 releases 7 briefings covered

all google reviews →

Improve `jnp.hypot` numerical accuracy and add n-ary precision tests. google/jax

by copybara-service

The division-based formula is gone, replaced by power-of-two scaling plus a custom JVP rule that the PR says avoids rounding errors for large inputs. The derivative at (0, 0) no longer returns NaN. If you call hypot on large-magnitude inputs or differentiate through it, your results shift.

Fix vmap of a scan closing over a Ref batched on a non-leading axis google/jax

by adityasingh2400

A `Ref` passed to `vmap` with a non-leading `in_axes` and then read or written inside a `lax.scan` no longer crashes. The root cause was `batching.batch_jaxpr`: its interface is per-input booleans, so it could not express the non-leading axis. Both failures from #39288 are fixed.

Replace `PyTuple_SET_ITEM` and `PyList_SET_ITEM` with nanobind builders. google/jax

by copybara-service

Raw `PyTuple_New`/`PyTuple_SET_ITEM` and `PyList_New`/`PyList_SET_ITEM` are not part of `Py_LIMITED_API`, so they were swapped for nanobind builders. This is groundwork for building jaxlib and xla/python against the stable ABI, a prerequisite for free-threaded Python support.

Default float16 to non-FTZ in numerics tests and fix subnormal underflow in lax.hypot. google/jax

by Peter Hawkins

float16 now preserves subnormals across CPU, GPU, and TPU; `numerics_test_util` defaults `input_ftz` and `output_ftz` to False for float16, and the per-test overrides that had worked around the old behavior are removed. The subnormal underflow in `lax.hypot` is fixed.

[lax] Removed an opt-in Pallas Triton lowering for ragged dot google/jax

by Sergei Lebedev

If this breaks you, the change description points to two exits: Tokamax, which it calls "a better-supported collection of ragged dot implementations for various GPUs", or plain XLA, which can lower ragged dot to cuDNN. Anyone who opted into the Pallas Triton path needs to migrate.

feat(auth): support bound access and ID tokens for GKE googleapis/google-cloud-python

by nbayati

The auth library now supports bound access and ID tokens for GKE, merged to main for release after review in two earlier change requests. Workloads that need identity tokens bound to a specific audience no longer have to assemble them by hand.

fix(spanner): correctly wrap async streaming calls in metrics interceptor googleapis/google-cloud-python

by olavloite

Async streaming RPCs were being wrapped in `_AsyncUnaryResponseWrapper` instead of `_AsyncStreamingResponseWrapper`, so metrics went unrecorded. In `grpc.aio`, streaming call objects implement `__aiter__`, and the interceptor's checks missed that case. Spanner metrics from streaming calls were the ones that vanished.

[typing] Use Self annotation for unary jax.Array methods google/jax

by copybara-service

Unary methods on `jax.Array` now use the `Self` annotation, which the change says unblocks annotation work specific to jax PRNG keys. Cheap cleanup unless you maintain a type-checked wrapper around Array methods.

$ ls google/week/ # the briefings behind this review

Keep up with Google in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

all google reviews →