Who contributed to Google on September 15, 2026?

6 developers shipped this update, including vfdev-5, Sergei Lebedev, copybara-service, mattjj, Mark Daoust, and release-please[bot].

Sep 15, 2026

$ the-wire · showcase

JAX lands batched host copies, Pallas MGPU deprecations

Q: What were the notable Google updates?

C++-side support for batched_copy_to_host_async, [pallas:mgpu] Deprecated `pl.program_id` in Pallas MGPU kernels, and [pltpu.emit_pipeline] Support async pipeline prefetch.

By RepoJournal · Filed 06:59 UTC on September 15, 2026 · About Google · Composed from the cited sources · methodology

6 people shipped this

copybara-service @copybara-service 2 cited

vfdev-5 @vfdev-5 1 cited

Sergei Lebedev @superbobry 1 cited

release-please[bot] @release-please[bot] 1 cited

mattjj @mattjj

Mark Daoust @MarkDaoust

JAX's multi-device and pipeline layers moved most: a new single-call host copy path, async prefetch in emit_pipeline, and a backward-incompatible deprecation in Pallas MGPU kernels, while the google-cloud-python client libraries shipped one regenerated API batch.

C++-side support for batched_copy_to_host_async. google/jax

by copybara-service

C++-side support for batched_copy_to_host_async, a new JAX API that copies multiple multi-device arrays' data to the host in one call. The implementation groups arrays by (ifrt::Client*, DeviceList) before dispatching; the C++ changes were split from the JAX/jaxlib work so the two can be versioned independently, which matters if you build against jaxlib rather than tracking JAX main.

[pallas:mgpu] Deprecated `pl.program_id` in Pallas MGPU kernels google/jax

by Sergei Lebedev

pl.program_id is now deprecated in Pallas MGPU kernels. The commit notes pl.num_programs is always resolved statically there, so no lowering rule exists. Kernels that reference program_id need to move off it before the deprecation becomes a removal.

[pltpu.emit_pipeline] Support async pipeline prefetch. google/jax

by copybara-service

pltpu.emit_pipeline now supports async pipeline prefetch, extending the pipelining primitive's scheduling for Pallas TPU kernels.

Updated requirement locks before adding torch dependency google/jax

by vfdev-5

Elsewhere in JAX, updated requirement locks landed ahead of adding a torch dependency, and vjp3 now errors on duplicate backward-pass log keys rather than silently accepting them.

google-cloud-visionai: v0.5.2 googleapis/google-cloud-python

by release-please[bot]

The google-cloud-python client batch landed on one regenerated API source update: google-cloud-visionai v0.5.2, google-cloud-vectorsearch v0.11.3, google-cloud-retail v2.13.0, google-cloud-geminidataanalytics v0.13.3, and google-cloud-compute-v1beta v0.12.5 all carry the same generate-and-release change. Generic Gemini SDK README docs gained a Gemini 3.8 Flash update, and premature dotnet examp...