$ cat google/week/2026-09-21.log
the week in review · Sep 21 – Sep 27, 2026
XLA:GPU to flip f32 dot default from TF32 to BF16
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
The change lands "in the nearest future" and aligns GPU precision with TPU, but numerically shifts a number of existing tests.
[XLA:GPU] In the nearest future we are going to flip the default precision of dot(f32,f32) from current TF32 precision to BF16. As a result of that the precision on GPU would match the precision on TP google/jax
The notice in the repo says the flip brings GPU precision in line with TPU, and that it "affects number of tests numerically" which the team wants fixed before it ships. If you compare f32 matmul results against TF32 baselines anywhere, expect drift once BF16 becomes the default.
Improve accuracy of real `jnp.sinc(x)` via additive range reduction. google/jax
Real-input sinc now computes sin(pi*x)/(pi*x) via exact additive range reduction, x = n + r with r in [-0.5, 0.5], plus a Taylor polynomial in r^2, and the ULP precision test was updated to match. Accuracy improves for real x; the complex path is untouched.
Add `jax.lax.polynomial` / `polynomial_p` for polynomial evaluation. google/jax
jax.lax.polynomial and polynomial_p evaluate a degree-N polynomial as a primitive rather than an open-coded Horner loop. The input notes that Horner's recurrence makes autodiff apply the product rule at every step, so this gives differentiation a single op to work with.
[Mosaic GPU] Support partial reductions across warps and lanes google/jax
FragmentedArray.reduce now accepts a target layout with unreduced tiled dimensions, tracked by new Unreduced annotations on TiledLayout, so reductions no longer have to span all warps and lanes. Relevant if you write Mosaic GPU kernels that reduce partially.
[Mosaic:GPU] Use frontend attributes to request collective memory for Mosaic GPU symmetric buffers. google/jax
Symmetric-memory parameters and results now carry operands_memory_spaces and results_memory_spaces frontend attributes instead of the previous backend_config mechanism. Anyone hand-writing Mosaic GPU lowering with backend_config for collective memory needs to move to the attributes.
Clean up references to previously deprecated APIs for v0.12.0 google/jax
by Jake VanderPlas
References to APIs deprecated and removed before v0.11.0 are cleaned up; the only behavioral change is that the custom AttributeError becomes a generic one. If you caught the old error type, that check now fails differently.
[Mosaic TPU] Separate debug info from the Mosaic kernel `body` into a new `debug_locations` in `custom_call_config` with a new configuration flag `jax_mosaic_unzip_debug_locations`. google/jax
Debug info moves out of the Mosaic kernel body into a new debug_locations field in custom_call_config behind the jax_mosaic_unzip_debug_locations flag. The input says this resolves a discrepancy between JAX compilation cache key generation and XLA executable fingerprint generation, which is the kind of mismatch that silently invalidates cache hits.
v2.25.0 googleapis/python-genai
python-genai 2.25.0 adds sample_audio to Voice and a Voices API resource to the GAOS SDKs. Alongside: Google Cloud Python gained Firestore BSONDecimal128 for lossless 128-bit decimal serialization, resumable upload support in google-api-core, a storage_class parameter on AsyncAppendableObjectWriter, and the Python 3.15 classifier in handwritten libraries.
$ 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.
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.