JAX

Write ordinary numerical Python against jax.numpy, then wrap it in function transformations: grad builds derivatives (reverse or forward mode, nested to any order, through Python control flow), jit traces the function and hands it to the XLA compiler for fused accelerator code, and vmap turns a per-example function into a batched one without manual reshaping. Composed, they make per-example gradients a one-liner. For scale, programs can lean on compiler-chosen sharding, declare explicit shardings visible in array types, or drop to per-device code with shard_map, and Pallas lets advanced users write custom GPU and TPU kernels. JAX stays deliberately narrow, so layers, optimizers and data loading come from companions like Flax, Equinox, Keras, Optax and Grain. The CPU wheel installs with pip on Linux, macOS and Windows, while NVIDIA CUDA, Google TPU and AMD ROCm builds target Linux, and Apple and Intel GPU support is experimental. Apache-2.0 and free for commercial use, though the team calls it a research project rather than an official Google product. Suits researchers and engineers wanting fine control at any scale.