The Wire · Showcase
JAX SHIPS EMIT_PIPELINE PRIMITIVE, FIXES TPU CI FLAKINESS
By RepoJournal · Filed · About Google
JAX landed the emit_pipeline primitive base implementation [ref:1] while switching TPU CI to pinned versions to slash ephemeral test failures.
The emit_pipeline primitive [1] is a major foundation piece for the transform pipeline. JAX also added XLA Transform E2E testing [2] to catch integration issues before they hit production. On the reliability front, switching TPU CI jobs from nightly to pinned versions [3] eliminates the version skew debugging tax that was eating cycles on phantom failures. Pallas/GPU work continues: the kernel interpreter now properly handles collective_axes for run_scoped allocations [4], fixing a gap in distributed GPU kernel support. On the GenAI SDK side, automatic function calling in python-genai is now restricted to Chat.send_message [5], no longer available on direct Model.generate_content calls. This is a breaking change for anyone relying on the direct path. google-cloud-agentidentitycredentials shipped v0.1.0 [6], a new credential type for agent-based authentication workflows. BigFrames picked up Python bytecode translation [8] and better documentation on bqsql cell chaining [9], while Cloud Bigtable expanded sample_row_keys to support row_range filtering [7].
Action items
- → Update python-genai code relying on Model.generate_content function calling - switch to Chat.send_message googleapis/python-genai [plan]
- → Review emit_pipeline integration requirements for transform-heavy workloads google/jax [monitor]
- → Integrate google-cloud-agentidentitycredentials if using agent-based auth patterns googleapis/google-cloud-python [plan]
References
- [1] Add emit_pipeline primitive base implementation ↗ google/jax
- [2] Add XLA Transform E2E Test. ↗ google/jax
- [3] Switch TPU CI jobs to "pinned" instead of "nightly". ↗ google/jax
- [4] [Pallas][GPU kernel interpreter] Properly support `collective_axes` for allocations made by `run_scoped`. ↗ google/jax
- [5] test: for non-stream generate_content, automatic function calling is only available via Chat.send_message. Direct call to Model.generate_content does not support automatic function calling anymore. ↗ googleapis/python-genai
- [6] google-cloud-agentidentitycredentials: v0.1.0 ↗ googleapis/google-cloud-python
- [7] feat: support row_range in sample_row_keys method ↗ googleapis/google-cloud-python
- [8] chore(bigframes): Add simple Python bytecode translation ↗ googleapis/google-cloud-python
- [9] docs(bigframes): add a notebook explaining bqsql magics cell chaining ↗ googleapis/google-cloud-python
FAQ
- What changed in Google on June 10, 2026?
- JAX landed the emit_pipeline primitive base implementation while switching TPU CI to pinned versions to slash ephemeral test failures.
- What should Google teams do about it?
- Update python-genai code relying on Model.generate_content function calling - switch to Chat.send_message • Review emit_pipeline integration requirements for transform-heavy workloads • Integrate google-cloud-agentidentitycredentials if using agent-based auth patterns
- Which Google repositories shipped on June 10, 2026?
- google/jax, googleapis/python-genai, googleapis/google-cloud-python