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).

mlx_dart

A Dart port of mlx-swift — MLX tensor operations and neural network primitives for Dart on Apple silicon.

MLX is an array framework for machine learning on Apple silicon. mlx_dart brings the same API surface to Dart, wrapping the MLX C bindings via FFI.

Installation

Add to your pubspec.yaml:

dependencies:
  mlx_dart: ^0.1.0

The package requires libmlxc.dylib (or libmlx.dylib) to be present on the library search path at runtime. Build it from ml-explore/mlx-c.

Quick Start

import 'package:mlx_dart/mlx_dart.dart';

void main() {
  final ctx = MLXContext.gpu();

  final a = MLXArray.fromFloats(ctx, [1, 2, 3, 4], shape: [2, 2]);
  final b = MLXArray.ones(ctx, [2, 2]);
  final c = a.matmul(b);
  c.eval();

  print(c.toFloat32List()); // [3.0, 3.0, 7.0, 7.0]

  c.dispose();
  b.dispose();
  a.dispose();
  ctx.dispose();
}

API Status

The table below tracks progress against mlx-swift modules.

MLX Core (Source/MLX in mlx-swift)

Feature	mlx-swift	mlx_dart
MLXContext (device + stream)	Device / Stream	MLXContext
MLXDtype	DType	MLXDtype
Array construction	MLXArray.init(...)	MLXArray.zeros/ones/arange/fromFloats/fromInts/bool_/int_/float_
Properties (ndim, shape, size, dtype, nbytes)	yes	yes
eval() / evalAll	yes	yes
Data extraction (toFloat32List, toInt32List, itemFloat, itemInt)	yes	yes
toString()	yes	yes
Arithmetic operators (+, -, *, /)	yes	yes
Shape ops (reshape, expandDims, squeeze, flatten, transpose, T, swapAxes)	yes	yes
Slicing (slice, sliceUpdate, take)	yes	yes
Reductions (sum, mean, max, argmax, softmax, cumsum, sort, argsort)	yes	yes
Element-wise math (exp, log, sqrt, rsqrt, abs, sigmoid, tanh, sin, cos, square, erf, floor, ceil, round, clip, pad)	yes	yes
Comparisons (greaterEqual, less, logicalAnd)	yes	partial
Type casting (astype)	yes	yes
matmul	yes	yes
concatenate, stack	yes	yes
where	yes	yes
conv2d	yes	yes
repeat, split	yes	yes
Fast NN primitives (rmsNorm, layerNorm, rope, scaledDotProductAttention)	MLXFast	yes (on MLXArray)
Causal mask utility	createCausalMask	createCausalMask
Bilinear resize	custom	bilinearResize
Sampling (argmaxSample, categoricalSample)	various	yes
Reductions (min, prod, std, var)	yes	yes
Element-wise math (log1p, log2, logSumExp, maximum, minimum, negative, notEqual, isNaN, isInf, sign, sinh, cosh, tan, atan, acos, asin, expm1, erfInverse)	yes	yes
Array ops (tril, triu, trace, top, roll, einsum, inner, outer, kron, tensordot, tiled, nanToNum, degrees, radians)	yes	yes
eye, full, linspace factory ops	yes	yes
conv1d	yes	yes
conv3d	yes	not yet
convTransposed2d	yes	yes
convTransposed1d	yes	yes
convTransposed3d	yes	not yet
Grad transforms (grad, vjp, jvp, valueAndGrad)	yes	not yet
vmap	yes	not yet
compile	yes	not yet
MLXCustomFunction	yes	not yet
Memory / wired memory	yes	yes (activeMemory, cacheMemory, peakMemory, clearCache, etc.)
State / global stream management	yes	not yet
IO (safetensors)	yes	yes
IO (GGUF)	yes	not yet

MLXRandom (Source/MLXRandom)

Feature	mlx-swift	mlx_dart
seed, key, split	yes	yes
uniform	yes	yes
normal	yes	yes
multivariateNormal	yes	yes
randInt	yes	yes
bernoulli	yes	yes
truncatedNormal	yes	yes
gumbel	yes	yes
categorical	partial (on MLXArray)	categoricalSample
laplace	yes	yes

MLXNN (Source/MLXNN)

Infrastructure

Feature	mlx-swift	mlx_dart
Module base class	yes	Module
Parameter tracking / parameters()	yes	ParametersMixin
loadWeights	yes	yes
sanitize	yes	yes
WeightLoader	—	WeightLoader
Sequential container	yes	yes
UnaryLayer protocol	yes	yes (UnaryLayer interface)
Value-and-grad helpers	ValueAndGrad	not yet

Layers

Layer	mlx-swift	mlx_dart
Linear	yes	yes
Embedding	yes	yes
RMSNorm	yes	yes
LayerNorm	yes	yes
Conv2d	yes	yes
GroupNorm	yes	yes
BatchNorm	yes	yes
InstanceNorm	yes	yes
Conv1d	yes	yes
Conv3d	yes	not yet
ConvTransposed1d	yes	yes
ConvTransposed2d	yes	yes
ConvTransposed3d	yes	not yet
Dropout	yes	yes
Dropout2d	yes	yes
Dropout3d	yes	not yet
MaxPool2d	yes	yes
MaxPool1d	yes	yes
MaxPool3d	yes	not yet
AvgPool2d	yes	yes
AvgPool1d	yes	yes
AvgPool3d	yes	not yet
MultiHeadAttention	yes	yes
Transformer (encoder + decoder)	yes	yes (TransformerEncoder, TransformerDecoder)
SinusoidalPositionalEncoding	yes	yes
RNN, GRU, LSTM	yes	yes
Upsample	yes	yes
Quantized layers	yes	not yet
KV cache (Cache, RotatingKVCache)	yes	yes (KVCache, RotatingKVCache)

Activations

Functional activations available on MLXArray: relu(), gelu(), silu(), sigmoid(), tanh(), softmax(), leakyRelu(), elu(), celu().

Activation	mlx-swift	mlx_dart
relu, gelu, silu, sigmoid, tanh	yes	yes (on MLXArray)
leakyRelu	yes	yes
elu, celu	yes	yes
relu6, reluSquared	yes	yes
logSoftmax	yes	yes
softPlus / softsign	yes	yes
softshrink	yes	yes
Activation layer classes (ReLU, GELU, SiLU, Sigmoid, Tanh, Softmax, LeakyReLU, ELU, CELU)	yes	yes

Losses

Loss	mlx-swift	mlx_dart
crossEntropy	yes	yes
binaryCrossEntropy	yes	yes
l1Loss, mseLoss	yes	yes
nllLoss, klDivLoss	yes	yes
smoothL1Loss, huberLoss	yes	yes
hingeLoss, tripletLoss	yes	yes
logCoshLoss, cosineSimilarityLoss	yes	yes

MLXOptimizers (Source/MLXOptimizers)

Optimizer	mlx-swift	mlx_dart
SGD	yes	yes
Adam, AdamW, Adamax	yes	yes
RMSprop, AdaGrad, AdaDelta	yes	yes
Lion	yes	yes
Adafactor	yes	yes

MLXLinalg (Source/MLXLinalg)

Feature	mlx-swift	mlx_dart
norm	yes	yes
svd, qr	yes	yes
lu	yes	yes
inv	yes	yes
triInv, choleskyInv	yes	yes
cholesky	yes	yes
solve, solveTriangular	yes	yes
cross	yes	yes

MLXFFT (Source/MLXFFT)

Feature	mlx-swift	mlx_dart
fft, ifft, fft2, ifft2, fftn, ifftn	yes	yes
rfft, irfft, rfft2, irfft2	yes	yes

Architecture

mlx_dart
├── lib/src/
│   ├── context.dart   — MLXContext, MLXException
│   ├── array.dart     — MLXArray, MLXDtype, free ops
│   ├── module.dart    — Module, ParametersMixin, WeightLoader
│   ├── layers.dart    — Linear, Embedding, RMSNorm, LayerNorm, Conv2d, Conv1d, Sequential, Dropout, Dropout2d, GroupNorm, InstanceNorm, BatchNorm, MultiHeadAttention, TransformerEncoder, TransformerDecoder, SinusoidalPositionalEncoding, MaxPool1d, MaxPool2d, AvgPool1d, AvgPool2d, Upsample, RNN, GRU, LSTM, activation layer classes
│   ├── losses.dart    — MLX loss functions
│   ├── optimizers.dart — SGD, Adam, AdamW, Adamax, RMSprop, AdaGrad, AdaDelta, Lion, Adafactor
│   ├── kv_cache.dart  — KVCache, RotatingKVCache
│   ├── random.dart    — MLXRandom
│   ├── linalg.dart    — MLXLinalg
│   ├── fft.dart       — MLXFFT
│   └── io.dart        — MLXSafetensors
└── pubspec.yaml

Key design decisions mirroring mlx-swift:

• Explicit context — MLXContext wraps a device + stream. All arrays belong to a context (mlx-swift uses an implicit global default, Dart makes it explicit).
• Manual disposal — MLXArray.dispose() must be called when an array is no longer needed (no GC finaliser). Consider try/finally patterns.
• Functional style — operations return new arrays rather than mutating in place, matching mlx-swift.

Contributing

Contributions expanding coverage of the mlx-swift API surface are welcome. Good first targets:

1. Additional MLXArray ops (missing element-wise math, reductions, factory ops)
2. MLXRandom module
3. Sequential container and remaining MLXNN layers
4. MLXOptimizers
5. Grad / vmap transforms

License

Apache 2.0 — see LICENSE.