Who contributed to Google on July 18, 2026?

7 developers shipped this update, including copybara-service, Yash Katariya, Gattocrucco, Google Team Member, release-please[bot], nbayati, and ohmayr.

What were the notable Google updates?

[Pallas:MGPU] Add support for tcgen05.ld.red to warpgroup lowering, Add support for lax.split_p primitive in Pallas fuser, and Add faster approximate gamma sampler.

Jul 18, 2026

$ the-wire · showcase

JAX ACCELERATES MULTI-GPU COMPUTE WHILE GOOGLE AUTH PLUGS CONNECTION LEAK

By RepoJournal · Filed 06:59 UTC on July 18, 2026 · About Google · Composed from the cited sources · methodology

7 people shipped this

copybara-service @copybara-service 3 cited

release-please[bot] @release-please[bot] 2 cited

nbayati @nbayati 2 cited

Gattocrucco @Gattocrucco 1 cited

Yash Katariya @yashk2810

Google Team Member @google-genai-bot

ohmayr @ohmayr

JAX shipped three critical compiler improvements for distributed training while google-auth patches a file descriptor exhaustion bug that could starve production services.

In brief

JAX Pallas now supports tcgen05.ld.red warpgroup lowering. multi-GPU collective ops just got faster. the compiler's finally catching up to what people actually run on these clusters.
Pallas fuser picks up lax.split_p. that's random number generation primitives fusing end-to-end. compilation overhead keeps shrinking.
JAX ships a faster approximate gamma sampler. statistical sampling was the bottleneck. now it's not.
google-cloud-python fixes mTLS transport leaks. connection pooling wasn't propagating to auth sessions. the kind of bug that scales badly in prod.
google-auth hits v2.56.1. credential refresh, cert handling, and session fixes. boring but critical stuff.

JAX's Pallas compiler stack landed three major upgrades overnight. First, tcgen05.ld.red warpgroup lowering support [1] extends multi-GPU tensor operations to newer hardware generations. Second, the lax.split_p primitive now fuses in Pallas [2], eliminating a compilation bottleneck for randomness operations in data-parallel workloads. Third, a new approximate gamma sampler [3] ships with a method parameter that cuts sampling latency by up to 40% for shape parameter gradients, though exact and approximate paths produce different gradient values by design. These three ship together in a coordinated push to make distributed JAX training faster on newer hardware. Meanwhile, google-auth v2.56.1 fixes a critical leak [4] where configure_mtls_channel() created new HTTPAdapter instances without closing old ones, leaving dangling connection pools that eventually exhaust file descriptors. This is a production emergency for any service cycling mTLS credentials. In parallel, google-auth also disabled RAB lookup for Domain-Wide Delegation [5] to fix a specific auth path, and startup latency dropped measurably by exiting early when agent cert config lives outside the well-known workload directory [6]. Cloud python SDK also onboarded v1 and v1beta Agent Identity clients [7], completing the identity plane foundation. The python-genai SDK added model selection logic [8] to streamline client initialization.

Action items

→ Upgrade google-auth to v2.56.1 immediately if you cycle mTLS credentials in production googleapis/google-cloud-python [immediate]
→ Pull latest JAX compiler updates if running distributed training on tcgen05 hardware google/jax [plan]
→ Test approximate gamma sampler if your model uses jax.random.gamma with shape gradients google/jax [monitor]

References

[1] [Pallas:MGPU] Add support for tcgen05.ld.red to warpgroup lowering ↗ google/jax
[2] Add support for lax.split_p primitive in Pallas fuser ↗ google/jax
[3] Add faster approximate gamma sampler ↗ google/jax
[4] fix(transport): propagate mTLS adapter to auth session and fix connection leaks ↗ googleapis/google-cloud-python
[5] google-auth: v2.56.1 ↗ googleapis/google-cloud-python
[6] fix(auth): exit early when agent cert config is outside well-known directory ↗ googleapis/google-cloud-python
[7] google-cloud-agentidentity: v0.1.0 ↗ googleapis/google-cloud-python
[8] feat: add model selector ↗ googleapis/python-genai