The Wire · Showcase
JAX CONSOLIDATES GPU/TPU TEST INFRASTRUCTURE WHILE GENAI ADDS GEMMA 4 LOCAL TOKENIZER SUPPORT
By RepoJournal · Filed · About Google
JAX shipped test consolidation and Pallas API alignment overnight, while the Python GenAI SDK picked up local Gemma 4 tokenizer support and Google Cloud Python clients got regenerated against the latest API sources.
JAX cleaned house across multiple fronts [1]. The bigger move: Pallas Triton now supports optional `name=` parameters [2], which unblocks removal of the legacy compilation path after the next jaxlib release. This is the infrastructure work that gets you to cleaner abstractions downstream. Meanwhile, `take_along_axis` got smarter about negative indices and dtype casting [3], specifically fixing an overflow risk that was tripping up Mosaic TPU on 8-bit operations. On the Pallas side, the multi-GPU kernel API now mirrors the standard kernel API [4], reducing surface area for developers bouncing between implementations. Over in python-genai, Gemma 4 local tokenizer support landed [5], which means you can now tokenize locally without hitting the API. That matters for cost and latency on the inference side. Google Cloud Python got a full regeneration against the latest googleapis source [6], plus a new google-developer-knowledge service [7]. BigFrames also improved its SQL error messages to catch unescaped braces [8], saving debugging cycles when building queries.
Action items
- → Update JAX if you're using take_along_axis with Mosaic TPU to avoid 8-bit casting issues google/jax [plan]
- → Regenerate Python Cloud clients against latest googleapis source for new developer-knowledge APIs googleapis/google-cloud-python [monitor]
- → Pull python-genai for Gemma 4 local tokenizer if you're doing offline tokenization googleapis/python-genai [plan]
References
- [1] remove a print and delete the GPU test (merged with TPU test now) google/jax
- [2] [pallas:triton] Updated triton_kernel_call_ffi to allow (but not require!) `name=` google/jax
- [3] Add wrap_negative_indices to take_along_axis and avoid upcasting the take_along_dim index when wrap_negative_indices=False. ↗ google/jax
- [4] [pallas:mgpu] Aligned the API of `plgpu.kernel` with `pl.kernel` ↗ google/jax
- [5] feat: Gemma 4 local tokenizer support ↗ googleapis/python-genai
- [6] feat: update API sources and regenerate ↗ googleapis/google-cloud-python
- [7] feat(google/developers/knowledge/v1): add google-developer-knowledge ↗ googleapis/google-cloud-python
- [8] fix(bigframes): improve error message when unescaped `{` are found in SQL cells ↗ googleapis/google-cloud-python
FAQ
- What changed in Google on June 11, 2026?
- JAX shipped test consolidation and Pallas API alignment overnight, while the Python GenAI SDK picked up local Gemma 4 tokenizer support and Google Cloud Python clients got regenerated against the latest API sources.
- What should Google teams do about it?
- Update JAX if you're using take_along_axis with Mosaic TPU to avoid 8-bit casting issues • Regenerate Python Cloud clients against latest googleapis source for new developer-knowledge APIs • Pull python-genai for Gemma 4 local tokenizer if you're doing offline tokenization
- Which Google repositories shipped on June 11, 2026?
- google/jax, googleapis/python-genai, googleapis/google-cloud-python