mlx_dart 0.1.2

MLX tensor operations and neural network primitives for Dart — the Dart equivalent of mlx-swift, for Apple silicon (macOS / iOS).

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.1.2 - 2026-03-12

Added

• gatedDeltaUpdate — single-step gated delta state update for GatedDeltaNet-style linear attention models. Implements S_new = alpha * S + beta * outer(v − S@k, k) over batched multi-head state [B, H, D, D].

0.1.1 - 2026-03-10

• Updates

0.1.0 - 2026-03-10

Added

MLX Core (MLXContext, MLXArray)

• MLXContext — wraps an MLX device + stream; factory constructors for GPU (Metal) and CPU.
• MLXArray — n-dimensional array backed by the MLX C runtime with full ownership semantics.
• MLXDtype — Dart enum covering all 14 MLX element types (bool_, uint8–uint64, int8–int64, float16, bfloat16, float32, float64, complex64).
• MLXException — typed exception carrying the failed operation name and error code.
• Array construction: zeros, ones, eye, full, linspace, arange, fromFloats, fromInts, bool_, int_, float_.
• Array properties: ndim, shape, size, dtype, nbytes, dim(axis).
• eval() / evalAll() for lazy evaluation.
• Data extraction: toFloat32List, toInt32List, itemFloat, itemInt.
• Arithmetic operators: +, -, *, /.
• Shape ops: reshape, expandDims, squeeze, flatten, transpose, T, swapAxes.
• Slicing & indexing: slice, sliceUpdate, take.
• Reductions: sum, mean, max, min, prod, argmax, argmin, std, variance, softmax, cumsum, sort, argsort.
• Element-wise math: exp, expm1, log, log1p, log2, logSumExp, sqrt, rsqrt, abs, sigmoid, tanh, sin, cos, tan, sinh, cosh, atan, acos, asin, square, erf, erfInverse, floor, ceil, round, clip, pad, sign, negative, degrees, radians, nanToNum.
• Activation functions: relu, relu6, reluSquared, gelu, silu, leakyRelu, elu, celu, logSoftmax, softPlus, softsign, softshrink.
• Comparisons: greaterEqual, less, logicalAnd, notEqual, isNaN, isInf.
• Type casting: astype.
• Fast Metal-accelerated NN primitives: rmsNorm, layerNorm, rope, scaledDotProductAttention.
• Array ops: tril, triu, trace, top, roll, einsum, inner, outer, kron, tensordot, tiled, repeat, split.
• Free functions: conv1d, conv2d, convTransposed1d, convTransposed2d, concatenate, stack, where, evalAll, createCausalMask, bilinearResize.
• Sampling: argmaxSample, categoricalSample.
• Memory management on MLXContext: activeMemory, cacheMemory, peakMemory, memoryLimit, clearCache(), resetPeakMemory(), setMemoryLimit(), setCacheLimit().

MLXRandom

• MLXRandom — static interface for random number generation.
• seed, key, split.
• uniform, normal, multivariateNormal.
• randInt, bernoulli, truncatedNormal, gumbel, laplace.

MLXNN — Layers

• Module base class and ParametersMixin for neural network parameter management.
• WeightLoader for loading checkpoints by dotted-path keys, sanitize.
• UnaryLayer interface for single-input / single-output layers.
• Sequential container.
• Linear layers: Linear, Embedding.
• Normalisation: RMSNorm, LayerNorm, GroupNorm, BatchNorm, InstanceNorm.
• Convolution: Conv1d, Conv2d, ConvTransposed1d, ConvTransposed2d.
• Pooling: MaxPool1d, MaxPool2d, AvgPool1d, AvgPool2d.
• Regularisation: Dropout, Dropout2d.
• Attention: MultiHeadAttention.
• Transformer: TransformerEncoderLayer, TransformerEncoder, TransformerDecoderLayer, TransformerDecoder.
• Positional encoding: SinusoidalPositionalEncoding.
• Recurrent: RNN, GRU, LSTM.
• Upsampling: Upsample (bilinear resize).
• KV cache: KVCache, RotatingKVCache.
• Activation layer classes: ReLU, GELU, SiLU, Sigmoid, Tanh, Softmax, LeakyReLU, ELU, CELU.

MLXNN — Losses

• crossEntropy, binaryCrossEntropy.
• l1Loss, mseLoss.
• nllLoss, klDivLoss.
• smoothL1Loss, huberLoss.
• hingeLoss, tripletLoss.
• logCoshLoss, cosineSimilarityLoss.

MLXOptimizers

• SGD (with momentum and weight decay).
• Adam, AdamW, Adamax.
• RMSprop, AdaGrad, AdaDelta.
• Lion.
• Adafactor.

MLXLinalg

• norm, svd, qr, lu.
• inv, triInv, choleskyInv, cholesky.
• solve, solveTriangular, cross.

MLXFFT

• fft, ifft, fft2, ifft2, fftn, ifftn.
• rfft, irfft, rfft2, irfft2.

IO

• MLXSafetensors — load and save weights in the safetensors format.

Tests

• Pure-Dart unit tests (no native lib required) in test/mlx_dart_test.dart.
• Integration tests tagged @Tags(['integration']) in test/integration/mlx_dart_integration_test.dart, covering all modules above.