Who contributed to Google on August 24, 2026?

2 developers shipped this update, including copybara-service and Peter Hawkins.

What were the notable Google updates?

Optimize `planar_snake` calculation, Remove dynamic discovery of unknown rocm wheels, and Remove dynamic discovery of unknown rocm wheels.

Google

JAX, the GenAI SDK, and the Cloud libs - Google's open source layer

Pick a date

Topics: AI / ML Python Full archive →

The Wire · Showcase

JAX SPEEDS UP SNAKE, DROPS ROCM WHEEL LOOKUP

By RepoJournal · Filed 06:58 UTC on August 24, 2026 · About Google

2 people shipped this

copybara-service @copybara-service 2 cited

Peter Hawkins @hawkinsp 2 cited

JAX's planar_snake is now up to 10% faster for power-of-two tiles, and the team is killing a painfully slow ROCm wheel discovery path.

JAX shipped a targeted performance win overnight: the `planar_snake` kernel now uses a `cond` instead of a `select` when `tile_width` is a power-of-two, cutting latency by about 10% in the common case [1]. The same PR also improves speed in most other configurations, so it's a free win for anyone running snake-shaped workloads. Separately, JAX is ripping out dynamic ROCm wheel discovery, a code path that calls `pkgutil.iter_modules()` and can stall for seconds on machines with many installed packages [2]. The maintainers call the mechanism "not really needed" since they no longer support unknown ROCm versions with old releases. A second, near-identical copy of that change landed in PR form, likely the same fix being tracked in both internal and external pipelines [3]. Also in flight: a test fix that stops an unsupported-gather test from accidentally using a gather that's now supported, which could otherwise mask future breakage [4]. All told, 4 commits and 3 PRs across google/jax in the last 24 hours. Nothing here is a security patch, but the snake optimization is the kind of drop-in speedup you want in your next benchmark run.

Action items

→ Pull in the planar_snake optimization if you hit snake layouts with power-of-two tile widths google/jax [plan]
→ Watch the ROCm wheel discovery removal for any custom build scripts that rely on it google/jax [monitor]

References

[1] Optimize `planar_snake` calculation. ↗ google/jax
[2] Remove dynamic discovery of unknown rocm wheels. google/jax
[3] Remove dynamic discovery of unknown rocm wheels. ↗ google/jax
[4] Fix unsupported gather test not to use a (newly supported) gather. google/jax

Quick answers

What shipped in Google on August 24, 2026?: JAX's planar_snake is now up to 10% faster for power-of-two tiles, and the team is killing a painfully slow ROCm wheel discovery path. In total, 4 commits and 3 pull requests landed.
Who contributed to Google on August 24, 2026?: 2 developers shipped this update, including copybara-service and Peter Hawkins.
What were the notable Google updates?: Optimize `planar_snake` calculation, Remove dynamic discovery of unknown rocm wheels, and Remove dynamic discovery of unknown rocm wheels.

JAX Renames Mosaic GPU Parameters

Google's JAX team is renaming a core Mosaic GPU attribute, and it will break your cross-process parameter code.

ai 3 shipped

Google Aug 22

GOOGLE WEB STACK AND SPANNER GET LAST-MINUTE ENGINEERING CATCH-UPS

If you thought Google's Python ecosystem was quiet overnight, you missed a flurry of fixes that touch everything from Bigtable filters to Spanner's DBAPI driver.

ai 49 shipped

Google Aug 21

JAX hardens distributed mode with mTLS

JAX just made its distributed coordination service production-ready for untrusted networks.

ai 49 shipped

Google Aug 20

SPANNER FIXES DATA LOSS IN AUTOCOMMIT MODE

Spanner's Python client silently dropped rows from multi-row DML statements in autocommit mode, and the fix is live.

ai 52 shipped

Elsewhere on the wire

AI Agents Aug 24

CLINE SHIPS V4.1.15 AS TELEMETRY BUGS SURFACE ACROSS THE AGENT STACK

Cline's long-awaited auto-approve fix landed overnight, but it's the telemetry drop in 4.1.11 that should have you checking your session metrics.

ai-agents 44 shipped

Local LLMs Aug 24

VLLM'S P2P WEIGHT SYNC CUTS RL MOE TRANSFER COSTS

Forget all-gathering experts during RL rollout: vLLM's new sharded_rdt backend lets each worker pull only the slice it needs, turning a total_bytes transfer into a total_bytes divided by num_workers one.