$ the-wire · showcase
JAX adds a polynomial primitive, Mosaic GPU grows partial reductions
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX ships two changes that touch how gradients and GPU reductions are expressed, while python-genai lands labels across live and GenerateContent calls and google-cloud-python opens storage_class on async appends.
Add `jax.lax.polynomial` / `polynomial_p` for polynomial evaluation. google/jax
The new jax.lax.polynomial / polynomial_p primitive gives differentiation a rule of its own instead of the open-coded Horner loop (acc = acc * x + c_k), where the product rule recurses at every step and k-th order derivatives blow up to Theta(2^k N) operations. The motivation, as the PR describes it, is that those derivatives "never simplify to zero even for orde[r]"; code that differentiates t...
[Mosaic GPU] Support partial reductions across warps and lanes google/jax
FragmentedArray.reduce now accepts a target layout with unreduced tiled dimensions, and TiledLayout gains Unreduced annotations to mark dimensions not yet fully reduced across the warp or lane hierarchy. Partially reduced arrays can then take part in pointwise ops and a later reduce, so reductions no longer have to complete in a single call.
feat(storage): support storage_class in AsyncAppendableObjectWriter googleapis/google-cloud-python
AsyncAppendableObjectWriter.__init__ takes a storage_class parameter defaulting to None and propagates it to _AsyncWriteObjectStream at open() time, so async appendable writes can pick a storage class without a separate copy or rewrite. The same PR updates system test configuration and zonal tests for RCU buckets and preprod environments.
feat: include labels for LiveClientSetup googleapis/python-genai
LiveClientSetup now carries labels, matching what the REST surface already allowed on generate requests; the GenerateContent side followed in a separate change. If you route traffic through the Live API and were tagging sessions elsewhere to segment them, that bookkeeping can move into the client.
ci(storage): add GCS read microbenchmark runner and Cloud Build config googleapis/google-cloud-python
A new benchmarks-cloudbuild.yaml and run_benchmark_tests.sh drive high-bandwidth GCS DirectPath read microbenchmarks against rapid zonal buckets, orchestrating VM lifecycle and private SSH runs and printing a formatted results table straight into Cloud Build logs. Also landing: unit and cover nox sessions that enforce 100% coverage for google-crc32c, and documented Python 3.15 wheel support in ...
