jax

name	jax
description	Comprehensive reference documentation and skill for JAX - Google's library for high-performance numerical computing and machine learning research. Covers JAX core (transformations, tracing, jaxprs, pytrees), jax.numpy, jax.lax, jax.nn, jax.random, automatic differentiation (grad, custom_jvp, custom_vjp, checkpoint/remat), automatic vectorization (vmap), parallel computing (pjit, shard_map, Mesh, Sharding), jax.scipy, jax.image, Pallas GPU/TPU kernels, export, FFI, debugging, profiling, distributed computing, experimental features, and JAX Enhancement Proposals (JEPs).
version	0.6.1

JAX - High-Performance Numerical Computing and Machine Learning Research

Overview

JAX is Google's library for high-performance numerical computing and machine learning research. It provides a familiar NumPy-style API with composable function transformations that enable just-in-time compilation to CPU, GPU, and TPU backends via XLA (Accelerated Linear Algebra).

JAX extends the NumPy API with four key transformations that compose freely:

1. jax.jit - Just-in-time compilation for accelerating JAX programs on accelerators
2. jax.grad - Automatic differentiation for computing gradients of arbitrary functions
3. jax.vmap - Automatic vectorization for batching without rewriting functions
4. jax.pjit - Parallel computation across multiple devices with data sharding

Supported Hardware: CPU, NVIDIA GPUs (CUDA 12+), AMD GPUs (ROCm), Google TPUs, Apple Metal (experimental), Intel GPUs (experimental)

Supported Platforms: Linux, macOS, Windows (CPU only)

JAX Version: 0.6.1 | jaxlib Version: 0.6.1 | XLA Version: OpenXLA

Key Architecture Concepts

• Transformations: Composable function transforms (jit, grad, vmap, pjit) that work on pure functions
• Tracing: Mechanism by which JAX inspects Python functions to build intermediate representations
• Jaxpr: JAX's intermediate language (JAX expression) representing traced computations
• Pytrees: Tree-like structures of nested containers (dicts, lists, tuples) that JAX can operate on
• XLA: Accelerated Linear Algebra compiler that generates optimized hardware code
• Sharding: Describing how data is distributed across devices (GSPMD, shard_map)
• Pallas: Low-level kernel language for writing custom GPU/TPU kernels within JAX
• Export: Ahead-of-time lowering and export of JAX programs to stable artifacts

Architecture Overview

+--------------------------------------------------------------+
|                    Python Frontend                            |
|  jax.numpy  |  jax.lax  |  jax.nn  |  jax.random  |  ...    |
+--------------------------------------------------------------+
|                 Function Transformations                       |
|  jit  |  grad  |  vmap  |  pjit  |  checkpoint  |  ...       |
+--------------------------------------------------------------+
|                    Tracing Engine                             |
|  jax.core.Trace  |  jax.core.Tracer  |  jax.core.Jaxpr      |
+--------------------------------------------------------------+
|                    XLA Compiler                               |
|  HLO -> MHLO -> Linalg -> Hardware-specific optimizations    |
+--------------------------------------------------------------+
|                    Runtime Backends                           |
|  CPU (LLVM)  |  NVIDIA GPU (PTX)  |  AMD GPU  |  TPU        |
+--------------------------------------------------------------+

Quick Reference

Installation

# CPU only
pip install jax

# NVIDIA GPU (CUDA 12)
pip install jax[cuda12]

# NVIDIA GPU (CUDA 12 + latest jaxlib nightly)
pip install -U jax[cuda12] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html

# TPU (Google Cloud)
pip install jax[tpu] -f https://storage.googleapis.com/jax-releases/libtpu_releases.html

Minimal Example

import jax
import jax.numpy as jnp

# JIT-compiled function
@jax.jit
def selu(x, alpha=1.67, lmbda=1.05):
    return lmbda * jnp.where(x > 0, x, alpha * jnp.exp(x) - alpha)

# Automatic differentiation
@jax.grad
def loss_fn(params, x, y):
    pred = params["w"] @ x + params["b"]
    return jnp.mean((pred - y) ** 2)

# Automatic vectorization
batched_selu = jax.vmap(selu)

# Random numbers (explicit PRNG state)
key = jax.random.PRNGKey(42)
x = jax.random.normal(key, (3, 3))

# Compute gradients
params = {"w": jnp.ones((2, 3)), "b": jnp.zeros(2)}
grads = loss_fn(params, jnp.ones((3,)), jnp.ones((2,)))

Basic Array Operations

import jax.numpy as jnp

# Array creation
x = jnp.array([1.0, 2.0, 3.0])
z = jnp.zeros((3, 4))
o = jnp.ones((2, 3))
r = jnp.arange(0, 10, 2)       # [0, 2, 4, 6, 8]
l = jnp.linspace(0, 1, 5)      # [0.0, 0.25, 0.5, 0.75, 1.0]
e = jnp.eye(3)                  # Identity matrix

# Math operations
a = jnp.dot(x, x)               # Dot product
b = jnp.matmul(z, z.T)          # Matrix multiply
c = jnp.sum(x)                   # Reduction
d = jnp.mean(x, axis=0)          # Mean along axis

# Indexing (same as NumPy)
m = jnp.arange(12).reshape(3, 4)
row = m[1, :]                    # Second row
col = m[:, 2]                    # Third column
sub = m[0:2, 1:3]                # Sub-matrix

# IMPORTANT: JAX arrays are immutable
# m[0, 0] = 99  # ERROR! Use functional update instead:
m = m.at[0, 0].set(99)

Composable Transformations

import jax
import jax.numpy as jnp

def f(x):
    return jnp.sum(x ** 2)

# Compose transformations freely
f_jit = jax.jit(f)                    # Compiled
f_grad = jax.grad(f)                   # Gradient
f_vmap = jax.vmap(f)                   # Vectorized
f_jit_grad = jax.jit(jax.grad(f))     # Compiled gradient
f_grad_vmap = jax.vmap(jax.grad(f))   # Per-sample gradients
f_jit_grad_vmap = jax.jit(jax.vmap(jax.grad(f)))  # All three

Training Loop

import jax
import jax.numpy as jnp
import optax  # Optimizer library

# Model as a pure function
def init_params(key, layer_sizes):
    params = []
    for i in range(len(layer_sizes) - 1):
        key, k1, k2 = jax.random.split(key, 3)
        w = jax.random.normal(k1, (layer_sizes[i], layer_sizes[i + 1])) * 0.01
        b = jax.random.normal(k2, (layer_sizes[i + 1],)) * 0.01
        params.append({"w": w, "b": b})
    return params

def predict(params, x):
    for layer in params[:-1]:
        x = jnp.dot(x, layer["w"]) + layer["b"]
        x = jax.nn.relu(x)
    x = jnp.dot(x, params[-1]["w"]) + params[-1]["b"]
    return x

def loss_fn(params, x, y):
    pred = predict(params, x)
    return jnp.mean(optax.l2_loss(pred, y).sum())

@jax.jit
def train_step(params, opt_state, x, y):
    loss, grads = jax.value_and_grad(loss_fn)(params, x, y)
    updates, opt_state = optimizer.update(grads, opt_state, params)
    params = optax.apply_updates(params, updates)
    return params, opt_state, loss


key = jax.random.PRNGKey()
params = init_params(key, [, , ])
optimizer = optax.adam()
opt_state = optimizer.init(params)


 epoch  (num_epochs):
     batch_x, batch_y  dataloader:
        params, opt_state, loss = train_step(params, opt_state, batch_x, batch_y)

Distributed Computing

import jax
import jax.numpy as jnp
from jax.sharding import Mesh, PartitionSpec as P

# Define device mesh
devices = jax.devices()
mesh = Mesh(devices.reshape((len(devices), 1)), ("data", "model"))

# Shard data across devices
from jax.sharding import NamedSharding
sharding = NamedSharding(mesh, P("data", None))
x = jax.random.normal(jax.random.PRNGKey(0), (1024, 784))
x_sharded = jax.device_put(x, sharding)

# Use pjit for distributed computation
@jax.jit
def distributed_train_step(params, x_sharded):
    ...

Pytrees

import jax

# Pytrees are tree-like structures of containers
params = {
    "layer1": {"w": jnp.ones((3, 4)), "b": jnp.zeros(4)},
    "layer2": {"w": jnp.ones((4, 2)), "b": jnp.zeros(2)},
}

# Flatten and unflatten
flat_values, tree_def = jax.tree.flatten(params)
params_restored = jax.tree.unflatten(tree_def, flat_values)

# Apply a function to every leaf
scaled = jax.tree.map(lambda x: x * 2, params)

# Tree utilities
struct = jax.tree.map(lambda x: jnp.zeros_like(x), params)  # Same structure, zero values

Reference Chapters

Core Concepts

1. Overview and Architecture - What is JAX, design philosophy, architecture layers, package structure, and ecosystem
2. Installation and Setup - CPU/GPU/TPU installation, Docker, nightly builds, verification, environment configuration
3. Key Concepts - Transformations, tracing, jaxprs, pytrees, functional programming model, pure functions
4. JIT Compilation - jax.jit, compilation caching, tracing mechanics, static arguments, control flow restrictions
5. Automatic Differentiation - jax.grad, forward-mode (jvp), reverse-mode (vjp), higher-order derivatives, stop_gradient
6. Advanced Automatic Differentiation - Custom derivative rules, perturbations, tangents, cotangents, complex number differentiation
7. Automatic Vectorization (vmap) - jax.vmap, batching semantics, spmd_axis_name, in_axes/out_axes, nested vmap
8. Control Flow - jax.lax.cond, jax.lax.while_loop, jax.lax.fori_loop, jax.lax.scan, jax.lax.switch, Python control flow in traced code
9. Pytrees - Tree flattening/unflattening, custom pytree nodes, tree_map, tree utilities, common patterns
10. Random Numbers - PRNG design, explicit state, jax.random module, Threefry/philox PRNG, key splitting

Numerical Libraries

11. jax.lax - Linear Algebra Extensions - Low-level operators, convolution, reduce_window, scatter/gather, clamp, sorting, platform-specific ops
12. jax.numpy - NumPy API - Complete NumPy-compatible API, differences from NumPy, indexing, broadcasting, ufuncs
13. jax.nn - Neural Network Functions - Activations (relu, gelu, silu, softmax, etc.), initializers, one_hot, standardize
14. jax.scipy - SciPy API - Linear algebra, optimization, special functions, signal processing, stats functions
15. jax.image - Image Processing - Resize, affine transforms, scaling functions, interpolation methods
16. jax.random - Random Number Generation - PRNGKey, split, normal, uniform, categorical, permutation, bernoulli, beta, gamma, distributions

Core Internals

17. Core Types and Abstract Values - ShapedArray, DShapedArray, abstract evaluation, shaped abstract values, bond dimensions
18. Tracing and Jaxpr Internals - Trace/Tracer objects, jaxpr language, eqns, vars, constants, evaluation, pretty-printing
19. Custom Derivatives (custom_jvp, custom_vjp) - Defining custom differentiation rules, custom_jvp, custom_vjp, nondiff_argnums, bypassing AD
20. Gradient Checkpointing and Rematerialization - jax.checkpoint (remat), policy-based rematerialization, memory-compute tradeoffs, scan checkpointing

Parallel and Distributed Computing

21. Sharding and Distributed Computing - NamedSharding, PositionalSharding, GSPMD, device_put, Mesh, multi-host, distributed arrays
22. shard_map - Per-device computation, collective operations, explicit communication, SPMD vs manual sharding
23. Pallas - GPU and TPU Kernels - Low-level kernel language, grid programming model, memory spaces, Pallas primitives
24. Pallas GPU Programming - GPU-specific Pallas features, WMMA, VMEM, SMEM, HBM, async copies, barriers, TMA
25. Pallas TPU Programming - TPU-specific Pallas features, VMEM, SEM, dot primitives, matmul fusion, systolic array

Debugging and Performance

26. Debugging and Error Handling - jax.debug.print, jax.debug.breakpoint, checkify, common errors, NaN debugging, shape errors
27. Profiling and Performance - jax.profiler, TensorBoard integration, memory profiling, timeline analysis, performance best practices
28. GPU Performance Tips - Kernel fusion, memory layout, async dispatch, data transfer optimization, kernel launch overhead

Compilation and Export

29. Ahead-of-Time Compilation (AOT) - jax.jit with static args, lowering, compilation, AOT lowering and compilation APIs
30. Export and jax2tf - jax.export, StableHLO export, jax2tf conversion, TensorFlow SavedModel, cross-platform deployment
31. Foreign Function Interface (FFI) - Calling C/C++ code from JAX, jax.extend.ffi, XLA custom calls, registration, lowering rules
32. External Callbacks - jax.debug.print, jax.debug.breakpoint, jax.pure_callback, jax.io_callback, host callbacks

Extension and Advanced Topics

33. Building on JAX (Extensions) - Creating JAX extensions, custom interpreters, custom traverse rules, library development patterns
34. Type Promotion - Type promotion rules, NumPy compatibility, weak types, dtype behavior in transformations
35. Data Types (dtypes) - float16, bfloat16, float32, float64, int8, int32, uint32, complex64, extended precision, custom dtypes
36. Stateful Computations and Effects - Handling state in a functional framework, jax.extend, effect systems, side-effect management
37. Checkify - Functional Error Checking - jax.checkify, functional error handling, assert-enable transforms, error propagation in JIT
38. JAX Enhancement Proposals (JEPs) - Design documents for major features, JEP process, historical context, evolution of JAX
39. Configuration and Environment Variables - jax.config, JAX_PLATFORMS, JAX_TRACEBACK_FILTERING, XLA_FLAGS, all config options
40. Common Patterns and Gotchas - Stateful patterns, in-place updates, array mutation, device placement, async dispatch, common pitfalls

Extension API and Experimental Features

41. jax.extend - Extension API - Public extension API, custom JAXpr interpreters, interop with JAX internals, stable extension points
42. Experimental Features - jax.experimental modules, Array, sharding extensions, custom mesh utils, work in progress features
43. Sparse Operations - jax.experimental.sparse, BCOO format, sparse matrix operations, sparse autodiff, sparse Pallas
44. Training Cookbook - End-to-end training recipes, multi-GPU training, mixed precision, checkpointing, data loading patterns
45. Benchmarking - Benchmarking JAX code, BlockUntilReady, timing best practices, throughput measurement, comparison techniques

Key APIs Quick Reference

Transformations

Transform	Purpose	Example
jax.jit	Just-in-time compilation	jax.jit(f)(x) or @jax.jit
jax.grad	Reverse-mode differentiation	jax.grad(f)(x) returns df/dx
jax.jacfwd	Forward-mode Jacobian	jax.jacfwd(f)(x)
jax.jacrev	Reverse-mode Jacobian	jax.jacrev(f)(x)
jax.hessian	Hessian matrix	jax.hessian(f)(x)
jax.vmap	Automatic vectorization	jax.vmap(f)(batch_x)
jax.pmap	Parallel map (legacy)	jax.pmap(f)(sharded_x)
jax.value_and_grad	Value + gradient	loss, grads = jax.value_and_grad(f)(x)
jax.checkpoint	Gradient checkpointing	jax.checkpoint(f)(x) or @jax.checkpoint
jax.custom_jvp	Custom forward-mode rule	Decorator for custom AD
jax.custom_vjp	Custom reverse-mode rule	Decorator for custom AD
jax.linearize	Forward-mode AD	y, f_jvp = jax.linearize(f, x)
jax.vjp	Reverse-mode AD	y, f_vjp = jax.vjp(f, x)

Array Creation

import jax.numpy as jnp

jnp.array([1, 2, 3])           # From Python list
jnp.zeros((3, 4))              # Zeros
jnp.ones((2, 3))               # Ones
jnp.eye(3)                     # Identity matrix
jnp.arange(0, 10, 2)           # Range: [0, 2, 4, 6, 8]
jnp.linspace(0, 1, 5)          # 5 evenly-spaced values
jnp.full((2, 3), 7.0)          # Filled with value
jnp.zeros_like(x)              # Same shape/dtype, zeros
jnp.ones_like(x)               # Same shape/dtype, ones
jnp.empty((2, 3))              # Uninitialized (do not rely on values)

Common Neural Network Operations

import jax.nn as nn
import jax.numpy as jnp

# Activations
nn.relu(x)                     # ReLU
nn.gelu(x)                     # GELU
nn.silu(x)                     # SiLU (Swish)
nn.softmax(x, axis=-1)         # Softmax
nn.log_softmax(x, axis=-1)     # Log-softmax (numerically stable)
nn.sigmoid(x)                  # Sigmoid
nn.tanh(x)                     # Tanh
nn.elu(x)                      # ELU
nn.leaky_relu(x)               # Leaky ReLU
nn.one_hot(x, 10)              # One-hot encoding

# Initializers
import jax.nn.initializers as init
key = jax.random.PRNGKey(0)
w = init.kaiming_normal()(key, (256, 128))
b = init.zeros(key, (128,))

Device Management

import jax

# Query devices
jax.devices()                   # All available devices
jax.devices("cpu")              # CPU devices
jax.devices("gpu")              # GPU devices
jax.devices("tpu")              # TPU devices
jax.local_devices()             # Local devices only
jax.device_count()              # Number of devices
jax.local_device_count()        # Number of local devices

# Data placement
x = jax.device_put(x, jax.devices()[0])   # Move to specific device
x = jax.device_put(x, sharding)            # Shard across devices
y = jax.device_get(x)                      # Copy to host (numpy)

# Async dispatch
x = jnp.ones(1000)
y = jnp.dot(x, x)
y.block_until_ready()           # Wait for computation to finish

Important Differences from NumPy

1. Immutable arrays: JAX arrays are immutable; use .at[].set() for functional updates
2. PRNG state: JAX uses explicit PRNG state (no global random state); pass key everywhere
3. 64-bit disabled by default: Use jax.config.update("jax_enable_x64", True) to enable float64
4. Pure functions: Transformations require pure functions (no side effects)
5. Control flow: Use jax.lax.cond, jax.lax.while_loop, jax.lax.scan inside JIT
6. Async dispatch: Operations are asynchronous; use .block_until_ready() for timing

JAX Ecosystem

Library	Description
Flax	Neural network library with Linen module system
Optax	Gradient processing and optimization library
Haiku	Sonnet-style neural network library
Diffrax	Differential equation solvers in JAX
Equinox	Neural networks via elegant Pytree manipulations
Chex	Testing utilities and dataclass extensions
DM-Haiku	DeepMind's neural network library
JAXopt	Hardware-accelerated optimization library
Orbax	Checkpointing and export utilities
T5X	T5 model implementation in JAX/Flax
MaxText	Large-scale Transformer training in pure JAX
Levanter	Scalable LLM training with JAX
Penzai	JAX model visualization and manipulation
Keras 3	Multi-backend Keras with JAX support
NumPyro	Probabilistic programming with JAX
JAX-Cosmo	Cosmological computations
Oryx	Probabilistic programming and bijectors

Version and Compatibility

• JAX: 0.6.1
• Minimum Python: 3.10+
• NumPy: 1.24+
• Supported CUDA: 12.x
• Supported ROCm: 6.x
• XLA: OpenXLA (StableHLO)