> For the complete documentation index, see [llms.txt](https://deeplearning4j.konduit.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://deeplearning4j.konduit.ai/nd4j.md).

# ND4J

- [Overview](https://deeplearning4j.konduit.ai/nd4j/overview.md): Architecture, INDArray interface, memory model, data types, and backend system for the ND4J tensor library
- [Quickstart](https://deeplearning4j.konduit.ai/nd4j/quickstart.md): Hands-on quickstart guide for ND4J — creating arrays, basic operations, indexing, and shape manipulation
- [Creating NDArrays](https://deeplearning4j.konduit.ai/nd4j/creating-ndarrays.md): All methods for creating INDArrays — factory methods, from Java arrays, random, combining, and typed creation
- [Indexing and Slicing](https://deeplearning4j.konduit.ai/nd4j/indexing-and-slicing.md): Accessing and modifying elements, rows, columns, and sub-arrays of INDArrays using NDArrayIndex
- [Operations](https://deeplearning4j.konduit.ai/nd4j/operations.md): Scalar, element-wise, transform, reduction, broadcast, comparison, and linear algebra operations on INDArrays
- [Matrix Manipulation](https://deeplearning4j.konduit.ai/nd4j/matrix-manipulation.md): Reshaping, transposing, permuting, concatenating, sorting, and other shape manipulation operations on INDArrays
- [Data Types](https://deeplearning4j.konduit.ai/nd4j/data-types.md): The DataType enum, per-array typing, type casting, mixed precision, and migration from global data type
- [Activations](https://deeplearning4j.konduit.ai/nd4j/activations.md): Activation functions in ND4J — the Activation enum, IActivation interface, mathematical definitions, and usage in layers
- [Updaters](https://deeplearning4j.konduit.ai/nd4j/updaters.md): Optimization algorithms in ND4J — Adam, SGD, AdaGrad, learning rate schedules, and per-layer updater configuration
- [Loss Functions](https://deeplearning4j.konduit.ai/nd4j/loss-functions.md): All loss functions in ND4J — ILossFunction implementations, usage in output layers, weighted loss, and custom loss functions
- [Weight Initialization](https://deeplearning4j.konduit.ai/nd4j/weight-initialization.md): Weight initialization strategies in ND4J — WeightInit enum, IWeightInit interface, and choosing the right initializer
- [Serialization](https://deeplearning4j.konduit.ai/nd4j/serialization.md): Saving and loading INDArrays in binary, text, NumPy, and ByteBuffer formats
- [Workspaces](https://deeplearning4j.konduit.ai/nd4j/workspaces.md): MemoryWorkspace API — configuration, policies, nested workspaces, scope panic, and lifecycle management
- [Random Number Generation](https://deeplearning4j.konduit.ai/nd4j/random.md): Generating random INDArrays, probability distributions, seeding, and reproducibility in ND4J
- [New Operations Reference](https://deeplearning4j.konduit.ai/nd4j/new-operations.md): Complete reference for ~130 new native operations — fused attention, KV cache, PEFT linear layers, normalization, positional encoding, quantization, SSM, MoE, and audio/signal processing
- [Backends](https://deeplearning4j.konduit.ai/nd4j/overview-1.md): How ND4J's backend system works — SPI mechanism, backend selection, and the relationship between nd4j-native and nd4j-cuda
- [CPU (nd4j-native)](https://deeplearning4j.konduit.ai/nd4j/overview-1/cpu.md): Setting up the nd4j-native CPU backend — Maven dependencies, AVX2/AVX512 optimizations, OpenBLAS, MKL, and multi-threading
- [CUDA (nd4j-cuda)](https://deeplearning4j.konduit.ai/nd4j/overview-1/cuda.md): Setting up the nd4j-cuda GPU backend — CUDA versions, cuDNN integration, multi-GPU, and GPU memory management
- [Hardware Backends (1.0.0-rewrite)](https://deeplearning4j.konduit.ai/nd4j/overview-1/hardware-backends.md): GPU, TPU, DSP, and CPU acceleration backends — CUDA, TPU (PJRT), Hexagon (QNN), ZLUDA, ARM ACL, Apple Accelerate, cuDNN, MPS, MLIR, and multi-backend dispatch
- [SameDiff](https://deeplearning4j.konduit.ai/nd4j/overview-2.md): Automatic differentiation framework in ND4J — define-and-run computation graphs, comparison with MultiLayerNetwork and ComputationGraph
- [Variables](https://deeplearning4j.konduit.ai/nd4j/overview-2/variables.md): SDVariable types — VARIABLE, CONSTANT, PLACEHOLDER, ARRAY — data types, naming, and type conversion
- [Operations](https://deeplearning4j.konduit.ai/nd4j/overview-2/operations.md): Op namespaces — sd.math, sd.nn, sd.cnn, sd.rnn, sd.loss, sd.random — and SDVariable methods
- [Training](https://deeplearning4j.konduit.ai/nd4j/overview-2/training.md): Training SameDiff models — TrainingConfig, fit(), listeners, loss curves, and evaluation
- [Execution and Inference](https://deeplearning4j.konduit.ai/nd4j/overview-2/execution.md): Running SameDiff graphs — exec(), output(), batchOutput(), placeholders, and InferenceSession
- [Serialization](https://deeplearning4j.konduit.ai/nd4j/overview-2/serialization.md): Saving and loading SameDiff graphs in FlatBuffers format
- [DSP Execution Engine](https://deeplearning4j.konduit.ai/nd4j/overview-2/dsp.md): Complete guide to the Dynamic Shape Plan execution engine — compiled graph runtime with CUDA graph capture/replay, Triton/NVRTC/PTX JIT, 26-pass optimizer, and multi-backend dispatch