The Wire · Showcase
JAX TIGHTENS SEGMENT SHARDING RULES WHILE PALLAS FIXES CRITICAL TPU HANGS
By RepoJournal · Filed · About Google
JAX is cracking down on conflicting sharding specifications in segment operations while simultaneously shipping a critical fix that prevents TPU Interpret Mode from deadlocking when simulated devices fail.
The JAX team shipped two major enforcement changes to segment operations that will break existing code if you're mixing unreduced and reduced sharding directives [1] [2]. Segment sum, max, min, and prod now validate that you're not simultaneously using both parameter specs and unreduced/reduced output sharding, a pattern that was silently producing incorrect results. This is a hard requirement before the next major version lands. More critically, Pallas TPU Interpret Mode now properly handles device failures [3], implementing failure tracking and propagation so that when one simulated device crashes, others fail cleanly instead of hanging forever waiting for signals that will never arrive. The fix includes test cases that deadlock without it. Over in python-genai, the team shipped v2.7.0 [4] with computer_use field support for Vertex and CSV document mime type support for the Interaction API [5]. On the infrastructure side, google-cloud-python is consolidating CI/CD by moving prerelease and core_deps tests from GitHub Actions to Kokoro [6] [7], removing hardcoded package lists that were causing test coverage gaps [8].
Action items
- → Audit your JAX segment_sum/max/min/prod calls for conflicting sharding specs before next release google/jax [plan]
- → Upgrade python-genai to v2.7.0 if using Vertex computer_use or CSV document processing googleapis/python-genai [plan]
- → Update local Kokoro configuration references if you maintain google-cloud-python forks googleapis/google-cloud-python [monitor]
References
- [1] Error out if `*pspec` or `pspec[...]` is used when unreduced/reduced is non-empty ↗ google/jax
- [2] Add out_sharding to segment_{sum,max,min,prod}. Only allow unreduced out_sharding to `segment_sum`. ↗ google/jax
- [3] [Pallas] Improve error handling in TPU Interpret Mode to avoid hangs. ↗ google/jax
- [4] v2.7.0 ↗ googleapis/python-genai
- [5] docs: update the docs for 2.7 ↗ googleapis/python-genai
- [6] chore: update kokoro prerelease and core_deps jobs (#17286) googleapis/google-cloud-python
- [7] chore: remove prerelease and core_deps from GHA (#17291) googleapis/google-cloud-python
- [8] tests: remove hardcoded packages in system.sh (#17301) googleapis/google-cloud-python
FAQ
- What changed in Google on May 29, 2026?
- JAX is cracking down on conflicting sharding specifications in segment operations while simultaneously shipping a critical fix that prevents TPU Interpret Mode from deadlocking when simulated devices fail.
- What should Google teams do about it?
- Audit your JAX segment_sum/max/min/prod calls for conflicting sharding specs before next release • Upgrade python-genai to v2.7.0 if using Vertex computer_use or CSV document processing • Update local Kokoro configuration references if you maintain google-cloud-python forks
- Which Google repositories shipped on May 29, 2026?
- google/jax, googleapis/python-genai, googleapis/google-cloud-python