JAX

What is JAX?

JAX is an open-source Python library for high-performance numerical computing and machine learning research. It offers a NumPy-style API combined with automatic differentiation and compilation through XLA.

Key capabilities of JAX

• NumPy-style API: jax.numpy mirrors familiar array operations.
• Automatic differentiation: grad computes gradients of Python functions.
• JIT compilation: jit compiles functions with XLA for speed.
• Vectorization: vmap maps a function over batches automatically.
• Parallelism: pmap and sharding scale computation across devices.
• Accelerator support: Runs on CPU, GPU and TPU.

How JAX works

You write array code with jax.numpy, then wrap functions with transformations such as grad, jit and vmap. JAX traces the function, compiles it with XLA and runs it on the available accelerator. Programs must follow a functional style with pure functions, which makes transformations composable.

Who uses JAX?

JAX is used by machine learning researchers and engineers who want custom model and optimizer code with accelerator speed. Neural network libraries such as Flax are built on top of it.

JAX pricing

JAX is free and open source under the Apache 2.0 license. Costs come only from the compute you run it on, such as cloud GPUs or TPUs.

JAX alternatives

Alternatives include PyTorch, which has the largest research community and an imperative style, TensorFlow, which has a mature production serving ecosystem, and MLX, which targets Apple silicon.