The Wire · Showcase
JAX FIXES TYPING FAILURES AND SHARDING BUGS ACROSS CHECKPOINT AND EXTENDED DTYPE PIPELINES
By RepoJournal · Filed · About Google
JAX shipped critical fixes overnight for checkpoint policy exports and extended dtype sharding, addressing both typing failures and physical sharding conversion issues that were breaking production inference.
The most urgent fix brings back SimpleNamespace export of checkpoint policies [1], resolving typing failures that were cascading through downstream code. In parallel, JAX converted out_shardings to physical shardings in the presence of extended dtypes [2] [3], fixing a critical issue where sharding metadata wasn't translating correctly for models using non-standard numeric types. These two fixes target the same class of problem: breakage in the checkpoint and sharding layers that affects anyone running distributed training or inference with modern model architectures. The python-genai team landed CodeMenderAgentConfig support [4] [5], expanding agent capabilities for code analysis workflows. Meanwhile, gemini-embedding-2 now populates per-modality prompt token counts in embedding responses [6] [7], giving you fine-grained cost tracking across multimodal embeddings. All three JAX fixes are shipping simultaneously, which means your next JAX upgrade will stabilize the entire distributed training surface.
Action items
- → Upgrade JAX immediately if using checkpoint policies or extended dtypes in distributed training google/jax [immediate]
- → Update python-genai to pull CodeMenderAgentConfig and embedding token tracking improvements googleapis/python-genai [plan]
- → Audit existing checkpoints for sharding mismatches if using extended dtypes google/jax [monitor]
References
- [1] Bring back SimpleNamespace export of checkpoint policies to fix typing failures ↗ google/jax
- [2] Convert out_shardings to physical shardings in presence of extended dtypes. The same happens for in_shardings. Fixes https://github.com/jax-ml/jax/issues/39089 google/jax
- [3] Convert out_shardings to physical shardings in presence of extended dtypes. The same happens for in_shardings. Fixes https://github.com/jax-ml/jax/issues/39089 ↗ google/jax
- [4] feat(interactions): add CodeMenderAgentConfig ↗ googleapis/python-genai
- [5] feat(interactions): add CodeMenderAgentConfig googleapis/python-genai
- [6] feat: Populate per-modality prompt token count in embedding responses for gemini-embedding-2 googleapis/python-genai
- [7] feat: Populate per-modality prompt token count in embedding responses for gemini-embedding-2 ↗ googleapis/python-genai