The Wire · Showcase
JAX ACCELERATES MULTI-GPU COMPUTE WHILE GOOGLE AUTH PLUGS CONNECTION LEAK
By RepoJournal · Filed · About Google
JAX shipped three critical compiler improvements for distributed training while google-auth patches a file descriptor exhaustion bug that could starve production services.
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