$ the-wire · showcase
jax.reshard slices before reducing, SparseCore NaN comparisons corrected
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX removed a NotImplementedError from jax.reshard while fixing a NaN comparison bug in TPU Pallas kernels, a small productivity pass on SparseCore and python-genai helpers.
Slice free mesh axes before reducing unreduced axes in `jax.reshard` google/jax
When jax.reshard reduces unreduced axes and the target sharding introduces mesh axes the input does not use yet, it now slices on those axes first instead of raising NotImplementedError. The slice needs no communication, and the all_reduce or reduce_scatter that follows moves less data.
[Pallas:TPU] Lower lax.ne on floats to arith.cmpf une google/jax
Pallas on TPU mapped lax.ne_p on floating point operands to arith.CmpFPredicate.ONE, ordered inequality, which is false whenever an operand is NaN. Since x != y, jnp.isnan(x), and jnp.nan_to_num(x) all lower through lax.ne_p, SparseCore kernels returned false on NaN inputs; they now lower to arith.CmpFPredicate.UNE, matching IEEE 754 and Pallas interpret mode.
Issue a single AG and A2A when multiple trailing dims don't exist in dst_spec google/jax
When several trailing dims are absent from dst_spec, the reshard now issues a single all-gather and all-to-all rather than one per missing dim.
chore: Support display_name and Files API URIs in FunctionResponsePart helpers googleapis/python-genai
The FunctionResponsePart helpers in python-genai accept display_name and Files API URIs, so callers no longer have to assemble those fields by hand. The same change lands as a separate commit authored by Wanlin Du.
Prepare ForEachBufferSlice fore new Slice field google/jax
Elsewhere in the tree, ForEachBufferSlice gains preparation for a new Slice field, groundwork with no behavior visible from the API yet.