JAX 0.4.1 made the unified jax.Array implementation the default. The same array type could now represent both single-device and sharded values.
JAX
Also known as: Google JAX · JAX and XLA
Plain English
JAX is a Python library for numerical computing and machine learning. It can differentiate a function, apply it across batches, and compile it to run on accelerators.
Technical Overview
JAX exposes composable transformations such as grad, vmap, and jit over NumPy-style operations. JIT compilation uses XLA for CPU, GPU, and TPU targets. Sharding associates array dimensions with axes in a device mesh and lets the compiler or programmer specify how work and communication are divided.
Timeline (4)
2022
Release2025
ReleaseJAX 0.7 began using Shardy by default for compiler-driven sharding. The compiler propagates sharding choices across a program to decide how to partition operations and where to communicate between devices.
- Release
JAX 0.8 reimplemented pmap on top of jit and shard_map by default. The maintainers encouraged new code to use shard_map directly for explicit per-device behavior.
2026
ReleaseJAX 0.9 added support for exporting computations with explicit sharding. A serialized module could now retain its logical mesh and partition specification, which the caller must match.