jax-ml/jax

Features

• Automatic Differentiation Engines - Rewrites mathematical functions into their derivative equivalents by applying symbolic transformation rules to the underlying computational graph.
• Automatic Differentiation Frameworks - Computes exact gradients and higher-order derivatives of complex functions using both forward and reverse modes.
• Gradient Computation - Calculates function gradients to support repeated differentiation for complex models.
• Compiler Infrastructure - Transforms high-level code into optimized intermediate representations by tracing execution paths to enable just-in-time compilation.
• Runtime Environments - Executes identical code across multiple hardware devices by mapping operations over partitioned data shards to achieve massive scale numerical computation.
• Numerical Libraries - Transforms standard code into optimized, hardware-accelerated operations for large-scale scientific and machine learning workloads.
• Neural Network Parallelism - Distributes neural network training and inference across multiple devices using manual parallelism strategies.
• Distributed Computing Platforms - Manages data sharding and collective communication across multiple hardware devices and multi-dimensional meshes.
• Language Features - Ensures consistent function outputs by passing all data through parameters and avoiding side effects to maintain compatibility with automated transformations.
• Reverse-Mode Differentiation - Evaluates vector-Jacobian products to enable efficient reverse-mode automatic differentiation.
• Distributed Function Mapping - Executes functions across multiple devices while maintaining explicit control over data partitioning.
• Distributed Computing Topologies - Organizes physical processing units into logical grids to define communication topologies and control data distribution across hardware memory.
• Distributed Parallelism APIs - Distributes computations across multiple hardware units by mapping functions over data shards and managing device meshes.
• GPU Acceleration Libraries - Simplifies setup for running intensive computations on compatible graphics processing units.
• High-Performance and Parallel Computing - Accelerates complex mathematical operations and scientific simulations by leveraging specialized hardware for maximum execution speed.
• Deep Learning - A library for high-performance numerical computing with automatic differentiation.
• Machine Learning - Composable transformations for high-performance numerical computing.
• Machine Learning Frameworks - High-performance library for numerical computing and composable transformations.
• Machine Learning Libraries - Composable transformations for high-performance numerical computing.
• Model Training - Library for high-performance computing and automatic differentiation.
• Model Training Frameworks - High-performance numerical computing and automatic differentiation library.
• Computation and Optimization - Library for composable transformations of Python and NumPy programs.
• Array Partitioning Specifications - Defines how array values are stored across physical device memories to control data layout.
• Array Vectorization Utilities - Maps operations over array axes automatically to enable efficient batch processing and broadcasting across multidimensional data structures.
• Data Sharding Strategies - Controls the physical layout of multidimensional arrays across distributed memory by applying explicit constraints to guide data movement.
• Distributed Array Processing - Scales large-scale data computations across multiple hardware devices by automatically partitioning arrays and managing communication.
• Immutable Array Updates - Modifies array values using immutable indexing operations to ensure compatibility with functional programming patterns.
• Device Mesh Topologies - Organizes hardware into multi-dimensional grids to control data and computation distribution.
• Forward-Mode Differentiation - Evaluates Jacobian-vector products to enable efficient forward-mode automatic differentiation.
• Parallel Matrix Operations - Implements custom parallel matrix multiplication by sharding input tensors across devices.
• Containerized Computing - Supports deployment of pre-configured software images for isolated execution on specialized hardware.
• Functional Programming Interfaces - Enforces immutable data transformations and pure function composition to enable reliable optimization and automated code analysis.
• Diagnostic Tools - Detects and locates the source of invalid numerical values like NaN or Inf during function execution and gradient calculation.
• Custom Differentiation Rules - Overrides standard differentiation behavior by specifying custom Jacobian-vector or vector-Jacobian products.
• Higher-Order Differentiation - Calculates full Jacobian or dense Hessian matrices using advanced differentiation modes.
• Collective Communication Operations - Performs explicit collective operations to manage communication between devices.
• Distributed Synchronization - Executes communication between function instances to synchronize data across hardware.
• Logical Array Operations - Executes element-wise logical operations on arrays that remain compatible with compilation and avoid reliance on standard language-specific short-circuiting.
• Dependency Managers - Automates dependency and driver management using community-maintained package managers.
• Cloud Infrastructure Support - Provides pre-built packages for configuring software environments on cloud-based tensor processing units.
• Hardware Acceleration Drivers - Provides manual installation procedures for hardware drivers and software packages on host systems.
• Functional - Enforces pure functional programming patterns by treating array updates as new allocations to ensure compatibility with automated graph optimization.
• Automatic Differentiation - Calculates derivatives for complex-valued functions using real-valued pair identification for robust mathematical modeling.

Star history