$ the-wire · showcase
JAX ROLLS BACK TPU ACCUMULATOR API, CLOUD STORAGE FIXES STREAM CLEANUP
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX reverted a Pallas TPU accumulator change after it broke the explicit MXU API surface, while Google Cloud Storage patched a critical gRPC stream cleanup bug that could leak resources on connection failures.
The JAX team rolled back an accumulator reference exposure [1] [2] that was causing issues with downstream TPU workloads relying on explicit MXU APIs, restoring stability across the Pallas subsystem. In parallel, the XLA dependency received an automated update [3] to keep the compiler in sync. Over in Cloud Storage, a fix landed that ensures bidirectional gRPC streams properly close on open failures or cancellations [4] [5], addressing a resource leak in _AsyncReadObjectStream that could strand sockets if initialization failed. The storage fix adds a dedicated _close_socket_like_rpc helper and guards requests_done() calls to prevent double-closure failures, backed by new unit tests for the failure paths.
Action items
- → Update google-cloud-storage if you use async blob reads, especially in retry-heavy workloads googleapis/google-cloud-python [plan]
- → If running JAX on TPU with explicit MXU calls, verify your workload still passes after the accumulator rollback google/jax [monitor]
References
- [1] [Pallas TPU] Roll back exposing accumulator references to be used with explicit MXU APIs. ↗ google/jax
- [2] [Pallas TPU] Roll back exposing accumulator references to be used with explicit MXU APIs. ↗ google/jax
- [3] Update XLA dependency to use revision ↗ google/jax
- [4] fix(storage): ensure bidi-gRPC stream is cleaned up on open failure and close ↗ googleapis/google-cloud-python
- [5] fix(storage): ensure bidi-gRPC stream is cleaned up on open failure and close (#18119) ↗ googleapis/google-cloud-python
