$ the-wire · showcase
JAX hands DeviceList through the compiler, Spanner partitioned DML learns parameters
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX is replacing internal tuples of devices with xc.DeviceList across compiler entry points and sharding internals, while the Cloud Spanner DB-API driver fixes parameterized partitioned DML that previously failed outright.
fix(spanner): support parameters and preserve safety in partitioned DML (#19035) googleapis/google-cloud-python
Partitioned DML through the Cloud Spanner DB-API driver previously failed whenever query parameters were used, because the driver passed raw positional placeholders and parameter lists straight to Spanner instead of converting them to the named parameters and types it expects. Statement parsing also mutated SQL in place to append a dummy clause; that mutation is now gone, and the commit is flag...
Use xc.DeviceList everywhere i.e. `sharding._device_assignment` now returns a xc.DeviceList instead of a tuple(devices). google/jax
sharding._device_assignment now returns an xc.DeviceList where it used to return a tuple of devices, part of a push to use DeviceList throughout. Indexing and iteration still work as before, but anything type-checking against tuple or unpacking with tuple semantics can break.
Make compiler.py and create_compile_options accept a xc.DeviceList google/jax
compiler.py and create_compile_options now accept a xc.DeviceList, matching the internal representation change above. Callers that passed tuples should keep working, but the canonical path through the compiler now carries a DeviceList end to end.
[jax][hijax] Avoid CondSum for cond backward-pass logs when branches match. google/jax
When every branch of a lax.cond or lax.switch logs the same key with matching pytree structure and types under core.typematch, JAX now emits the logged value directly instead of wrapping it in a tagged CondSum. Branches that differ in presence, structure, or types still get CondSum.
fix(sqlalchemy-bigquery): drop support for end-of-life Python 3.10 (#19057) googleapis/google-cloud-python
sqlalchemy-bigquery drops support for Python 3.10 now that it has reached end of life. Elsewhere in google-cloud-python, the Django Spanner backend enables check constraints unconditionally against the Spanner Emulator, which now supports them, ending the inconsistent DDL that made alter_field migrations emit "None;"; Bigtable gained authorized view, drop_row_range, and drop_column_family sampl...
Action items
- → Migrate sqlalchemy-bigquery consumers off Python 3.10 to a supported interpreter, since the library no longer support... googleapis/google-cloud-python [plan]
- → Audit code that treats sharding._device_assignment as a tuple before upgrading jax jax-ml/jax [plan]