$ the-wire · showcase
JAX turns silent wrong answers in shard_map into errors
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX's latest batch of commits closes a class of silent numerical bugs in shard_map and on explicit meshes, where mixing reduced, unreduced, and replicated constants produced correct-looking but wrong results.
Raise errors instead of silently computing wrong values when unreduced/reduced arrays are mixed with constants google/jax
pad, concatenate/stack, and select_n now raise instead of returning wrong values when an unreduced or reduced array meets a replicated constant on explicit meshes or inside shard_map. The stated example: with 4 devices, jnp.pad(u, 1, constant_values=1.) returned [4., 10., 4.] because each device held a full copy of the constant and the final sum counted it four times.
Fix _one_vjp under shard_map. Also reject index/pred being unreduced in switch/cond. google/jax
_one_vjp is fixed under shard_map, and switch/cond now reject an index or predicate that is unreduced. while_loop's predicate is also checked so it cannot be more varying than body_jaxpr.out_avals.
Fix sort with reduced/unreduced/varying operands inside shard_map google/jax
sort inside shard_map now handles reduced, unreduced, and varying operands. This is the same fix landed by Yash Katariya in the google/jax mirror.
Fix sort with reduced/unreduced/varying operands inside shard_map google/jax
The commit-side landing of the sort fix for reduced, unreduced, and varying operands inside shard_map, identical in scope to the copybara change.
Add a Pallas SparseCore regression test that starts two push streams out of google/jax
A Pallas SparseCore regression test now starts two push streams out of tile memory to different memories (VMEM_SHARED and HBM) back to back before awaiting them, pinning the ordering behavior in tests only.