The Wire · Showcase
JAX FIXES TRIANGULAR_SOLVE MEMORY BLOWUP, EXPANDS GPU INTERPRET MODE
By RepoJournal · Filed · About Google
A partial revert of yesterday's batching rule is rolling back OOMs in triangular_solve while keeping the performance gains.
JAX shipped a fix [1] that reverts the problematic broadcast insertion from the previous batching rule but preserves the optimization by improving the sharding rule instead. That earlier PR was causing out-of-memory errors in certain batching cases, so this lands the best of both worlds: the batching rule stays lean while the sharding rule now accepts additional batch dimensions. In parallel, GPU Interpret Mode got expanded [2] to support while_p operations and better handle unpacking of arbitrary transforms for copy_gmem_to_smem_p, which broadens what workloads can now run in this debugging mode. These are targeted fixes addressing real memory and compatibility pain points from the last cycle.
Action items
- → Sync JAX to latest if you hit triangular_solve OOMs google/jax [immediate]
- → Test GPU Interpret Mode with while_p workloads if debugging GPU kernels google/jax [plan]
References
- [1] restore triangular_solve batching rule, improve sharding rule ↗ google/jax
- [2] Support while_p in GPU Interpret Mode. ↗ google/jax
FAQ
- What changed in Google on May 31, 2026?
- A partial revert of yesterday's batching rule is rolling back OOMs in triangular_solve while keeping the performance gains.
- What should Google teams do about it?
- Sync JAX to latest if you hit triangular_solve OOMs • Test GPU Interpret Mode with while_p workloads if debugging GPU kernels
- Which Google repositories shipped on May 31, 2026?
- google/jax