{"author":"JAX team","author_email":"jax-dev@google.com","bugtrack_url":null,"classifiers":["Programming Language :: Python :: 3.10","Programming Language :: Python :: 3.11","Programming Language :: Python :: 3.12","Programming Language :: Python :: 3.13"],"description":"

\n $\"logo\"$ \n

\n\n# Transformable numerical computing at scale\n\n[![Continuous integration](https://github.com/jax-ml/jax/actions/workflows/ci-build.yaml/badge.svg)](https://github.com/jax-ml/jax/actions/workflows/ci-build.yaml)\n[![PyPI version](https://img.shields.io/pypi/v/jax)](https://pypi.org/project/jax/)\n\n[**Transformations**](#transformations)\n| [**Scaling**](#scaling)\n| [**Install guide**](#installation)\n| [**Change logs**](https://docs.jax.dev/en/latest/changelog.html)\n| [**Reference docs**](https://docs.jax.dev/en/latest/)\n\n\n## What is JAX?\n\nJAX is a Python library for accelerator-oriented array computation and program transformation,\ndesigned for high-performance numerical computing and large-scale machine learning.\n\nJAX can automatically differentiate native\nPython and NumPy functions. It can differentiate through loops, branches,\nrecursion, and closures, and it can take derivatives of derivatives of\nderivatives. It supports reverse-mode differentiation (a.k.a. backpropagation)\nvia [`jax.grad`](#automatic-differentiation-with-grad) as well as forward-mode differentiation,\nand the two can be composed arbitrarily to any order.\n\nJAX uses [XLA](https://www.tensorflow.org/xla)\nto compile and scale your NumPy programs on TPUs, GPUs, and other hardware accelerators.\nYou can compile your own pure functions with [`jax.jit`](#compilation-with-jit).\nCompilation and automatic differentiation can be composed arbitrarily.\n\nDig a little deeper, and you'll see that JAX is really an extensible system for\n[composable function transformations](#transformations) at [scale](#scaling).\n\nThis is a research project, not an official Google product. Expect\n[sharp edges](https://docs.jax.dev/en/latest/notebooks/Common_Gotchas_in_JAX.html).\nPlease help by trying it out, [reporting bugs](https://github.com/jax-ml/jax/issues),\nand letting us know what you think!\n\n```python\nimport jax\nimport jax.numpy as jnp\n\ndef predict(params, inputs):\n for W, b in params:\n outputs = jnp.dot(inputs, W) + b\n inputs = jnp.tanh(outputs) # inputs to the next layer\n return outputs # no activation on last layer\n\ndef loss(params, inputs, targets):\n preds = predict(params, inputs)\n return jnp.sum((preds - targets)**2)\n\ngrad_loss = jax.jit(jax.grad(loss)) # compiled gradient evaluation function\nperex_grads = jax.jit(jax.vmap(grad_loss, in_axes=(None, 0, 0))) # fast per-example grads\n```\n\n### Contents\n* [Transformations](#transformations)\n* [Scaling](#scaling)\n* [Current gotchas](#current-gotchas)\n* [Installation](#installation)\n* [Neural net libraries](#neural-network-libraries)\n* [Citing JAX](#citing-jax)\n* [Reference documentation](#reference-documentation)\n\n## Transformations\n\nAt its core, JAX is an extensible system for transforming numerical functions.\nHere are three: `jax.grad`, `jax.jit`, and `jax.vmap`.\n\n### Automatic differentiation with `grad`\n\nUse [`jax.grad`](https://docs.jax.dev/en/latest/jax.html#jax.grad)\nto efficiently compute reverse-mode gradients:\n\n```python\nimport jax\nimport jax.numpy as jnp\n\ndef tanh(x):\n y = jnp.exp(-2.0 * x)\n return (1.0 - y) / (1.0 + y)\n\ngrad_tanh = jax.grad(tanh)\nprint(grad_tanh(1.0))\n# prints 0.4199743\n```\n\nYou can differentiate to any order with `grad`:\n\n```python\nprint(jax.grad(jax.grad(jax.grad(tanh)))(1.0))\n# prints 0.62162673\n```\n\nYou're free to use differentiation with Python control flow:\n\n```python\ndef abs_val(x):\n if x > 0:\n return x\n else:\n return -x\n\nabs_val_grad = jax.grad(abs_val)\nprint(abs_val_grad(1.0)) # prints 1.0\nprint(abs_val_grad(-1.0)) # prints -1.0 (abs_val is re-evaluated)\n```\n\nSee the [JAX Autodiff\nCookbook](https://docs.jax.dev/en/latest/notebooks/autodiff_cookbook.html)\nand the [reference docs on automatic\ndifferentiation](https://docs.jax.dev/en/latest/jax.html#automatic-differentiation)\nfor more.\n\n### Compilation with `jit`\n\nUse XLA to compile your functions end-to-end with\n[`jit`](https://docs.jax.dev/en/latest/jax.html#just-in-time-compilation-jit),\nused either as an `@jit` decorator or as a higher-order function.\n\n```python\nimport jax\nimport jax.numpy as jnp\n\ndef slow_f(x):\n # Element-wise ops see a large benefit from fusion\n return x * x + x * 2.0\n\nx = jnp.ones((5000, 5000))\nfast_f = jax.jit(slow_f)\n%timeit -n10 -r3 fast_f(x)\n%timeit -n10 -r3 slow_f(x)\n```\n\nUsing `jax.jit` constrains the kind of Python control flow\nthe function can use; see\nthe tutorial on [Control Flow and Logical Operators with JIT](https://docs.jax.dev/en/latest/control-flow.html)\nfor more.\n\n### Auto-vectorization with `vmap`\n\n[`vmap`](https://docs.jax.dev/en/latest/jax.html#vectorization-vmap) maps\na function along array axes.\nBut instead of just looping over function applications, it pushes the loop down\nonto the function’s primitive operations, e.g. turning matrix-vector multiplies into\nmatrix-matrix multiplies for better performance.\n\nUsing `vmap` can save you from having to carry around batch dimensions in your\ncode:\n\n```python\nimport jax\nimport jax.numpy as jnp\n\ndef l1_distance(x, y):\n assert x.ndim == y.ndim == 1 # only works on 1D inputs\n return jnp.sum(jnp.abs(x - y))\n\ndef pairwise_distances(dist1D, xs):\n return jax.vmap(jax.vmap(dist1D, (0, None)), (None, 0))(xs, xs)\n\nxs = jax.random.normal(jax.random.key(0), (100, 3))\ndists = pairwise_distances(l1_distance, xs)\ndists.shape # (100, 100)\n```\n\nBy composing `jax.vmap` with `jax.grad` and `jax.jit`, we can get efficient\nJacobian matrices, or per-example gradients:\n\n```python\nper_example_grads = jax.jit(jax.vmap(jax.grad(loss), in_axes=(None, 0, 0)))\n```\n\n## Scaling\n\nTo scale your computations across thousands of devices, you can use any\ncomposition of these:\n* [**Compiler-based automatic parallelization**](https://docs.jax.dev/en/latest/notebooks/Distributed_arrays_and_automatic_parallelization.html)\nwhere you program as if using a single global machine, and the compiler chooses\nhow to shard data and partition computation (with some user-provided constraints);\n* [**Explicit sharding and automatic partitioning**](https://docs.jax.dev/en/latest/notebooks/explicit-sharding.html)\nwhere you still have a global view but data shardings are\nexplicit in JAX types, inspectable using `jax.typeof`;\n* [**Manual per-device programming**](https://docs.jax.dev/en/latest/notebooks/shard_map.html)\nwhere you have a per-device view of data\nand computation, and can communicate with explicit collectives.\n\n| Mode | View? | Explicit sharding? | Explicit Collectives? |\n|---|---|---|---|\n| Auto | Global | ❌ | ❌ |\n| Explicit | Global | ✅ | ❌ |\n| Manual | Per-device | ✅ | ✅ |\n\n```python\nfrom jax.sharding import set_mesh, AxisType, PartitionSpec as P\nmesh = jax.make_mesh((8,), ('data',), axis_types=(AxisType.Explicit,))\nset_mesh(mesh)\n\n# parameters are sharded for FSDP:\nfor W, b in params:\n print(f'{jax.typeof(W)}') # f32[512@data,512]\n print(f'{jax.typeof(b)}') # f32[512]\n\n# shard data for batch parallelism:\ninputs, targets = jax.device_put((inputs, targets), P('data'))\n\n# evaluate gradients, automatically parallelized!\ngradfun = jax.jit(jax.grad(loss))\nparam_grads = gradfun(params, (inputs, targets))\n```\n\nSee the [tutorial](https://docs.jax.dev/en/latest/sharded-computation.html) and\n[advanced guides](https://docs.jax.dev/en/latest/advanced_guide.html) for more.\n\n## Gotchas and sharp bits\n\nSee the [Gotchas\nNotebook](https://docs.jax.dev/en/latest/notebooks/Common_Gotchas_in_JAX.html).\n\n## Installation\n\n### Supported platforms\n\n| | Linux x86_64 | Linux aarch64 | Mac x86_64 | Mac aarch64 | Windows x86_64 | Windows WSL2 x86_64 |\n|------------|--------------|---------------|--------------|--------------|----------------|---------------------|\n| CPU | yes | yes | yes | yes | yes | yes |\n| NVIDIA GPU | yes | yes | no | n/a | no | experimental |\n| Google TPU | yes | n/a | n/a | n/a | n/a | n/a |\n| AMD GPU | yes | no | experimental | n/a | no | no |\n| Apple GPU | n/a | no | n/a | experimental | n/a | n/a |\n| Intel GPU | experimental | n/a | n/a | n/a | no | no |\n\n\n### Instructions\n\n| Platform | Instructions |\n|-----------------|-----------------------------------------------------------------------------------------------------------------|\n| CPU | `pip install -U jax` |\n| NVIDIA GPU | `pip install -U \"jax[cuda12]\"` |\n| Google TPU | `pip install -U \"jax[tpu]\"` |\n| AMD GPU (Linux) | Follow [AMD's instructions](https://github.com/jax-ml/jax/blob/main/build/rocm/README.md). |\n| Mac GPU | Follow [Apple's instructions](https://developer.apple.com/metal/jax/). |\n| Intel GPU | Follow [Intel's instructions](https://github.com/intel/intel-extension-for-openxla/blob/main/docs/acc_jax.md). |\n\nSee [the documentation](https://docs.jax.dev/en/latest/installation.html)\nfor information on alternative installation strategies. These include compiling\nfrom source, installing with Docker, using other versions of CUDA, a\ncommunity-supported conda build, and answers to some frequently-asked questions.\n\n## Citing JAX\n\nTo cite this repository:\n\n```\n@software{jax2018github,\n author = {James Bradbury and Roy Frostig and Peter Hawkins and Matthew James Johnson and Chris Leary and Dougal Maclaurin and George Necula and Adam Paszke and Jake Vander{P}las and Skye Wanderman-{M}ilne and Qiao Zhang},\n title = {{JAX}: composable transformations of {P}ython+{N}um{P}y programs},\n url = {http://github.com/jax-ml/jax},\n version = {0.3.13},\n year = {2018},\n}\n```\n\nIn the above bibtex entry, names are in alphabetical order, the version number\nis intended to be that from [jax/version.py](../main/jax/version.py), and\nthe year corresponds to the project's open-source release.\n\nA nascent version of JAX, supporting only automatic differentiation and\ncompilation to XLA, was described in a [paper that appeared at SysML\n2018](https://mlsys.org/Conferences/2019/doc/2018/146.pdf). We're currently working on\ncovering JAX's ideas and capabilities in a more comprehensive and up-to-date\npaper.\n\n## Reference documentation\n\nFor details about the JAX API, see the\n[reference documentation](https://docs.jax.dev/).\n\nFor getting started as a JAX developer, see the\n[developer documentation](https://docs.jax.dev/en/latest/developer.html).\n","description_content_type":"text/markdown","docs_url":null,"download_url":null,"downloads":{"last_day":-1,"last_month":-1,"last_week":-1},"dynamic":["Author","Author-Email","Classifier","Description","Description-Content-Type","Home-Page","License","Provides-Extra","Requires-Dist","Requires-Python","Summary"],"home_page":"https://github.com/jax-ml/jax","keywords":null,"license":"Apache-2.0","license_expression":null,"license_files":null,"maintainer":null,"maintainer_email":null,"name":"jax","package_url":"https://pypi.org/project/jax/","platform":null,"project_url":"https://pypi.org/project/jax/","project_urls":{"Homepage":"https://github.com/jax-ml/jax"},"provides_extra":["minimum-jaxlib","cpu","ci","tpu","cuda","cuda12","cuda12-local","rocm","k8s"],"release_url":"https://pypi.org/project/jax/0.6.1/","requires_dist":["jaxlib<=0.6.1,>=0.6.1","ml_dtypes>=0.5.0","numpy>=1.25","numpy>=1.26.0; python_version >= \"3.12\"","opt_einsum","scipy>=1.11.1","jaxlib==0.6.1; extra == \"minimum-jaxlib\"","jaxlib==0.6.0; extra == \"ci\"","jaxlib<=0.6.1,>=0.6.1; extra == \"tpu\"","libtpu==0.0.15.*; extra == \"tpu\"","requests; extra == \"tpu\"","jaxlib<=0.6.1,>=0.6.1; extra == \"cuda\"","jax-cuda12-plugin[with-cuda]<=0.6.1,>=0.6.1; extra == \"cuda\"","jaxlib<=0.6.1,>=0.6.1; extra == \"cuda12\"","jax-cuda12-plugin[with-cuda]<=0.6.1,>=0.6.1; extra == \"cuda12\"","jaxlib<=0.6.1,>=0.6.1; extra == \"cuda12-local\"","jax-cuda12-plugin<=0.6.1,>=0.6.1; extra == \"cuda12-local\"","jaxlib<=0.6.1,>=0.6.1; extra == \"rocm\"","jax-rocm60-plugin<=0.6.1,>=0.6.1; extra == \"rocm\"","kubernetes; extra == \"k8s\""],"requires_python":">=3.10","summary":"Differentiate, compile, and transform Numpy code.","version":"0.6.1","yanked":false,"yanked_reason":null}

Dependency	Optional
jaxlib	false
ml_dtypes	false
numpy	false
opt_einsum	false
scipy	false
jax-cuda12-plugin	true
jax-rocm60-plugin	true
kubernetes	true
libtpu	true
requests	true

Package search

jax

project_urls

Dependencies

All package metadata

Releases