$ the-wire · showcase
JAX TPU scan lowering returns, Bigtable batcher closes a flush race
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX relanded the chlo.ScanOp lowering for cumulative reductions on TPU behind a libtpu 0.0.51 guard, while google-cloud-bigtable stopped losing staged mutations when close() races the flush timer.
Roll forward: lower cumulative reductions to chlo.ScanOp on TPU google/jax
cumsum, cumprod, cummax and cummin lower to chlo.ScanOp on TPU again; the downstream issues behind the rollback are reported fixed. Export mode and libtpu older than 0.0.51 keep the reduce_window path, and that guard is scheduled for removal after 2026-11-21.
fix(bigtable): synchronize MutationsBatcher staging and flush scheduling googleapis/google-cloud-python
MutationsBatcher now guards staging, flush scheduling and shutdown with a re-entrant lock (self._stage_lock), because a close() racing the timer flush could drop staged entries silently. close() holds the same lock while setting _closed, cancelling _flush_timer and scheduling the final flush before _sync_flush_executor shuts down.
Fix no_collectives false negatives for tuple sub-axis ordering, convert_element_type sharding, and multi-output primitives. google/jax
no_collectives was missing violations and now tracks (dimension, prefix_product) in _check_no_collectives_dim_change, so P("r") <-> P(("s", "r")) raises ForbiddenCollectiveError while collective-free suffix changes like P("r") <-> P(("r", "s")) still pass. convert_element_type sharding and multi-output primitives are covered by the same fix.
[Pallas:MGPU] Consolidate TMEM transpose and reshape checks into lowering. google/jax
plgpu.transpose_ref is being deprecated in favor of ref.transpose / ref.T, and its eager ref.memory_space check failed on untransformed Ref tracers while being bypassed entirely when calling ref.transpose directly. It now delegates to ref.transpose(permutation), and the TMEM check moves into _handle_transforms as ref_aval.memory_space == gpu_core.TMEM.
chore: log warning if users don't use json schema for FunctionDeclaration.from_callable or FunctionDeclaration.from_callable_with_api_option. Fixes #3105 googleapis/python-genai
Cleanup on the long tail: python-genai warns when FunctionDeclaration.from_callable or from_callable_with_api_option is called without a JSON schema, LiveConnectConfig.http_options gains base_url, api_version and headers, and google-cloud-python drops google-area120-tables as the Area 120 Tables API and its protos are turned down.
Action items