RepoJournal
Google

Google

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

Pick a date

The Wire · Showcase

JAX TIGHTENS GPU COMPILATION PIPELINE WHILE GENAI CLIENT GAINS FILE SEARCH

By RepoJournal · Filed · About Google

JAX is consolidating its GPU optimization path by removing experimental flags and shipping dynamic ref selection for data movement, while the Python GenAI client reaches v1.75.0 with multimodal file search support.

JAX's GPU stack is maturing fast. The team removed the xla_gpu_experimental_aot_compiled_thunks flag [1], which is now enabled by default, signaling a move away from experimental compilation modes toward production-ready AOT pathways. Simultaneously, they shipped pl.select_ref [2] — a new primitive that lets you dynamically choose between multiple reference objects for DMAs, supporting complex nested transforms and async copies that were previously awkward to express. This is foundational for building more flexible data movement patterns on TPU. On the visualization front, jaxpr_html [3] gives you actual HTML dumps with full source tracebacks, making it orders of magnitude easier to debug complex computation graphs. Meanwhile, Mosaic GPU got two critical lowering improvements: maskTiledVregs consolidation [4] and n-d transpose support [5], both pushing the compiler closer to handling real workloads. Over in the Python ecosystem, python-genai v1.75.0 [6] landed with multimodal file search capabilities, fixing a stale token caching bug in async mTLS [6] that could bite production deployments. The google-cloud-python monorepo is dropping Python 3.7–3.9 support [7], aligning with end-of-life, and enabling gRPC transport for memorystore clients [8].

Action items

References

  1. [1] Remove usages of xla_gpu_experimental_aot_compiled_thunks google/jax
  2. [2] Create `pl.select_ref` to dynamically choose from multiple refs for DMAs. ↗ google/jax
  3. [3] Add a jaxpr_html IR dump mode that dumps a jaxpr to an HTML document. ↗ google/jax
  4. [4] [Mosaic] Refactor maskTiledVregs and selectWithBounds to consolidate usages of mask and select. ↗ google/jax
  5. [5] [Mosaic GPU] Lower n-d transposes in WG semantics. ↗ google/jax
  6. [6] v1.75.0 ↗ googleapis/python-genai
  7. [7] chore(versions): update Python version support to 3.10+ in gapic-generator (#16910) googleapis/google-cloud-python
  8. [8] feat(memorystore): enable gRPC transport ↗ googleapis/google-cloud-python

FAQ

What changed in Google on May 5, 2026?
JAX is consolidating its GPU optimization path by removing experimental flags and shipping dynamic ref selection for data movement, while the Python GenAI client reaches v1.75.0 with multimodal file search support.
What should Google teams do about it?
Review jaxpr_html dump mode for your complex models — it's production-ready and worth integrating into your debugging workflow • Update to python-genai v1.75.0 if you use async mTLS — the token caching fix is critical for production • Audit your google-cloud-python usage — 3.10+ is now the floor; plan migrations off 3.9 and earlier
Which Google repositories shipped on May 5, 2026?
google/jax, googleapis/python-genai, googleapis/google-cloud-python

Related across the cluster

For your repos

The showcase is a teaser.
Your wire is the product.

Same engine. Different stack. Below: what changes when the wire is yours.

Showcase wire

  • 14 famous open source orgs
  • One wire per day
  • Public, generic
  • Read on the web, when you remember

Your wire

  • Up to 1,500 of your repos - orgs, deps, vendors
  • Morning and evening briefs
  • Action items routed to your team
  • Slack delivery, email, breaking-news CVE alerts

Want a hands-on demo first? Ask a current user for an invite link.