A deep learning library for Dart and Flutter. It brings native hardware-accelerated machine learning directly to your device. Train models and run inference locally without Python, cloud APIs, or static binaries.

🟥 Why flutter_ml?

Most machine learning in Flutter relies on external Python servers or read-only TFLite models. This package provides a native alternative. You get full control over tensor math, autograd graphs, and custom model architectures.

• Zero Frame Drops: Standard Dart lists trigger huge garbage collection pauses. This engine pre-allocates memory in native C heap or GPU VRAM. Static tape unrolling reuses buffers. Zero allocations happen during the training loop.
• Compiled Execution Tapes: FFI calls are expensive. Calling native code for every math operation introduces massive overhead. This library records operations into a byte-encoded tape. The entire graph runs on the GPU in a single dispatch.
• On-Device Training: Keep user data private. Train directly on the device.
• Hugging Face Compatibility: Load and export standard .safetensors model weights natively. No Python conversion scripts are needed.
• Modern Architectures: Build Transformers, Multi-Head Attention, RoPE, and spatio-temporal ConvLSTMs entirely in Dart.

🟧 CPU Engine (Eager Execution)

The CPU engine uses an eager computation graph. It is highly transparent. Use it for debugging, testing, or running lightweight models.

🟠 Does not require specified Device architecture

import 'package:flutter_ml/full_library.dart';

void main() {
  List<Layer<dynamic, dynamic>> layers = <Layer<dynamic, dynamic>>[];
  layers.add(DenseLayer(8, activation: ReLU()));
  layers.add(DenseLayer(1, activation: Sigmoid()));

  SNetwork model = SNetwork(layers, name: 'XOR-Model');

  List<Vector> inputs = <Vector>[];
  inputs.add(<double>[0.0, 0.0]);
  inputs.add(<double>[0.0, 1.0]);
  inputs.add(<double>[1.0, 0.0]);
  inputs.add(<double>[1.0, 1.0]);

  List<Vector> targets = <Vector>[];
  targets.add(<double>[0.0]);
  targets.add(<double>[1.0]);
  targets.add(<double>[1.0]);
  targets.add(<double>[0.0]);

  SGD optimizer = SGD(model.parameters, learningRate: 0.1);
  model.compile(configuredOptimizer: optimizer);

  model.fit(inputs, targets, epochs: 1000, debug: true);
}

🟨 GPU Engine (Static Tape Compilation)

The GPU engine builds static execution tapes for extreme performance across multiple operating systems. It requires initialization. It uses TapeLayer structures to orchestrate VRAM.

The engine actively supports and bundles acceleration targets for:

🟡Windows (Target.cuda) via .dll binaries.

🟡Linux (Target.cuda) via .so binaries.

🟡Android (Target.androidArm64, Target.androidX8664) via bundled NDK objects.

For declaring globally to not require gradient allocation the global variable dart GPUTensor.globalGrad=false; can be turned on and off. For local tensors, providing the local value in its instructor can be used.

import 'package:flutter_ml/full_library.dart';

void main() async {
  await GPUEngine.initialize(debug: false, target: Target.cuda); // Select targeted architecture

  List<TapeLayer> layers = <TapeLayer>[];
  layers.add(DenseReluTL(8));
  layers.add(DenseTL(1));

  GPUTensor<Matrix> input = GPUTensor<Matrix>.empty(<int>[4, 2]);
  GPUTensor<Matrix> target = GPUTensor<Matrix>.empty(<int>[4, 1]);

  SeqModel<Matrix, Matrix> model = SeqModel<Matrix, Matrix>(
    layers,
    input,
    target: target,
    lossFunction: mseMatrixGPU,
    optimizerBuilder: (List<GPUTensor> params) {
      return AdamGPU(params, 0.01);
    }
  );

  model.compile();

  List<double> rawX = <double>[0.0, 0.0, 0.0, 1.0, 1.0, 0.0, 1.0, 1.0];
  List<double> rawY = <double>[0.0, 1.0, 1.0, 0.0];

  for (int epoch = 0; epoch < 1000; epoch = epoch + 1) {
    model.runTraining(inputData: rawX, targetData: rawY);
  }

  model.runForward(inputData: rawX);
  model.inferResult!.toCpu();
  print(model.inferResult!.value);

  model.free();
  GPUEngine.dispose();
}

Core Design Decisions

Version 3.0.0 introduces complete framework restructuring. Architecture is focused on memory safety, static compilation, and zero-allocation runtime.

• Direct Pointer Memory: Tensors allocate via native C calloc (CPU) or device VRAM (GPU). Bypasses Dart Garbage Collection entirely.
• Dual Execution Modes: Eager graph evaluation for CPU. Deferred static bytecode compilation for GPU.
• Static Tape Unrolling: Intermediate tensors pre-allocated during build(). Reused across iterations. Zero memory allocations during training loops.
• Hybrid Interop: GPU subgraphs encapsulated within CPU autograd nodes. Native pointer accumulation via GPUEngine.addPointers.
• Safetensors I/O: Native binary decoding/encoding of Hugging Face .safetensors. Zero external dependencies.

---

🟩 CPU Components

CPU Layers (20)

AveragePoolingLayer, BatchNormalizationLayer, Conv2D, ConvLSTMLayer, DenseLayer, DropoutLayer, DualLSTM, EmbeddingLayer, FlattenLayer, GlobalAveragePoolingLayer, LSTMLayer, MaxPoolingLayer, MultiHeadAttentionLayer, MultiLSTMLayer, NormalizationLayer, PositionalEncodingLayer, ReLULayer, RNNLayer, SingleHeadAttentionLayer, TransformerEncodingLayer.

CPU Math Operations (58)

add, add3D, addMatrix, addMatrixAndVector, addScalar, addScalarToMatrix, addVector, avgPool1d, avgPool2d, batchNorm1dMath, batchNorm2dMath, binaryCrossEntropy, concatenate, concatenate3D, concatenateMatricesByColumn, conv2d, dot, dropoutMatrixMath, dropoutVectorMath, elementWiseMultiply, elementWiseMultiply3D, elementWiseMultiplyMatrix, eluMatrix, eluVector, globalAveragePooling, leakyReluMatrix, leakyReluVector, matMul, matVecMul, maxPool1d, maxPool2d, mishMatrix, mishVector, mse, mseMatrix, multiply, padMatrix, relu, reluMatrix, reshapeVectorToMatrix, scaleMatrix, selectRow, sigmoid, sigmoidMatrix, sigmoidScalar, softmaxMatrix, softmaxVector, softplus, stackMatricesTo3D, sum, sumMatrix, swishMatrix, swishVector, tanhMatrix, transpose, vectorExp, vectorLog, vectorTanh.

CPU Activations (8)

ELU, LeakyReLU, Mish, ReLU, Sigmoid, SiLU, Softmax, Tanh.

CPU Optimizers (8)

Adagrad, Adam, AdamW, AMSGrad, NAG, RMSprop, SGD, SGDMomentum.

---

🟦 GPU - Accelerated Components

Version 3.0.0 utilizes CommandBuffer execution tapes. Operations map to 32-bit OpCodes. Compiled graphs dispatch to CUDA 12.1+ devices in a single FFI call or WebGPU depending on device setting.

For full architectural blueprints, see the GPU Tensor Documentation and Tape Layer.

GPU TapeLayers (29)

AveragePooling2DGPU, BatchNorm1DGPU, BatchNorm2DGPU, Conv2DTapeLayer, ConvLSTMTapeLayer, DenseTL, DenseReluTL, DropoutMatrixTapeLayer, DropoutTapeLayer, DualLSTMTapeLayer, EmbeddingMatrixTapeLayer, EmbeddingTapeLayer, FlattenTapeLayer, GeluLayerMatrixTapeLayer, GeluLayerTapeLayer, GlobalAveragePooling1DTapeLayer, GlobalAveragePoolingGPU, LayerNormalizationTapeLayer, LSTMTapeLayer, MaxPooling1DTapeLayer, MaxPooling2DTapeLayer, MultiHeadAttentionTapeLayer, PositionalEncodingTapeLayer, ReLULayerMatrixTapeLayer, ReLULayerTapeLayer, RNNTapeLayer, SigmoidMatrixTapeLayer, SingleHeadAttentionTapeLayer, TransformerEncoderBlockTapeLayer.

GPU Math Operations (81)

absGPU, add3DGPU, addBiasToFeatureMapGPU, addBiasToMatMulOutGPU, addGPU, addMatrixAndVectorGPU, addMatrixGPU, addScalar3DGPU, addScalarMatrixGPU, addScalarVectorGPU, addVectorGPU, applyRopeGPU, argmaxGPU, avgPool2dGPU, batchNorm1dGPU, batchNorm2dGPU, binaryCrossEntropyGPU, broadcastAddVectorToMatrixGPU, buildMarkovTableGPU, causalMaskGPU, clampGPU, concatenate3DGPU, concatenateGPU, concatenateMatricesByColumnGPU, conv2dMultiChannelGPU, conv2dSimpleGPU, cosineSimilarityGPU, crossEntropyLossGPU, divide3DGPU, divideGPU, divideMatrixGPU, divideVectorGPU, dotProductGPU, dropoutGPU, elementWiseMultiply3DGPU, elementWiseMultiplyGPU, elementWiseMultiplyMatrixGPU, embeddingLookupBatchGPU, embeddingLookupGPU, euclideanDistanceGPU, flatten3DToMatrixGPU, geluGPU, geluMatrixGPU, globalAveragePoolingGPU, im2colGPU, l2NormGPU, layerNormMatrixGPU, loadSampleGPU, logGPU, maeLossGPU, markovPredictGPU, matMulBiasReluGPU, matMulGPU, matVecMulGPU, maxPool1dGPU, maxPool2dGPU, mseGPU, mseMatrixGPU, multiplyGPU, multiplyScalarGPU, padMatrixGPU, powGPU, reluGPU, reluMatrixGPU, reshape3DToMatrixGPU, reshapeMatrixTo3DGPU, reshapeVectorToMatrixGPU, rmsNormMatrixGPU, scaleMatrixGPU, scatterHeadsGPU, selectMatrixFrom3DGPU, selectRowGPU, sigmoid3DGPU, sigmoidGPU, sigmoidMatrixGPU, sigmoidScalarGPU, sliceColumnGPU, softmaxMatrixGPU, sqrtGPU, stackMatricesGPU, subtract3DGPU, subtractGPU, subtractMatrixGPU, subtractVectorGPU, sumGPU, sumMatrixGPU, sumReduceColumnsGPU, sumReduceRowsGPU, tanh3DGPU, tanhMatrixGPU, transposeGPU, vectorExpGPU, vectorTanhGPU.

GPU Optimizers (2)

AdamGPU, SGDGPU.

---

Static Graph Compilation Example

The SeqModel object replaces eager execution loops. It handles VRAM lifecycle and compiles reusable execution tapes (trainForward, trainBackward, trainOptimize, inferTape).

import 'package:flutter_ml/full_library.dart';

void main() async {
  // Initialize Native Engine
  await GPUEngine.initialize(debug: false, target: Target.cuda);

  // Data mapping
  List<double> rawX = <double>[0.0, 0.0, 0.0, 1.0, 1.0, 0.0, 1.0, 1.0];
  List<double> rawY = <double>[0.0, 1.0, 1.0, 0.0];

  // Define Architecture
  List<TapeLayer> layers = <TapeLayer>[];
  layers.add(DenseReluTL(8));
  layers.add(DenseTL(1));

  // Define I/O Boundaries
  GPUTensor<Matrix> input = GPUTensor<Matrix>.empty(<int>[4, 2]);
  GPUTensor<Matrix> target = GPUTensor<Matrix>.empty(<int>[4, 1]);

  // Build Sequential Compiler
  SeqModel<Matrix, Matrix> net = SeqModel<Matrix, Matrix>(
    layers,
    input,
    target: target,
    lossFunction: mseMatrixGPU,
    optimizerBuilder: (List<GPUTensor> params) => AdamGPU(params, 0.01)
  );

  // Allocate VRAM and Compile Tapes
  net.compile();
  
  // High-Speed Training Loop
  for (int epoch = 1; epoch <= 10000; epoch = epoch + 1) {
    net.runTraining(inputData: rawX, targetData: rawY);
  }

  // Compile Inference Tape & Predict
  net.predict(input);
  net.runForward(inputData: rawX);
  
  // Synchronize VRAM to CPU
  net.inferResult!.toCpu();
  print(net.inferResult!.value);

  // Cleanup
  net.free();
  GPUEngine.dispose();
}

---

🟪 Benchmark Speed on Consumer Hardware

Test Environment

• Hardware: RTX 3060 12GB
• Backend: CUDA 12.1 (Tensor Core acceleration enabled)
• Workload Size: Vector operations evaluated on 134,217,728 elements (~537 MB); Matrix operations evaluated on 8192x8192 dimensions.
• Methodology: 50 iterations per operation. VRAM is aggressively wiped and reallocated between each run to simulate cold loading and tape compilation times.

Metric Descriptions

• Operation: The executed mathematical or layer function.
• Time (ms): Pure native execution time on the GPU after the tape is compiled.
• Bandwidth (GB/s): Memory throughput achieved during the operation.
• Compute (TFLOPs): Theoretical processing overhead and throughput measured in TeraFLOPS.
• Alloc/Tape (ms): Time required for the CPU to allocate VRAM and compile the execution tape. (Note: In standard training/inference, this is a one-time cost).
• Free (ms): Time required to release VRAM and destroy native pointers.

Operation	Time (ms)	Bandwidth (GB/s)	Compute (TFLOPs)	Alloc/Tape (ms)	Free (ms)
ADD	5.24	307.31	0.0256	2135.94	38.08
MARKOV_TBL	11.98	22.40	0.0056	648.10	16.40
MARKOV_PRD	3.38	213.20	0.0474	1271.30	26.51
SUBTRACT	5.23	308.12	0.0257	2256.46	42.87
MULTIPLY	5.14	313.05	0.0261	2466.38	15.88
DIVIDE	5.39	298.92	0.0249	2129.76	48.59
ABS	3.60	298.63	0.0373	1704.05	33.45
SQRT	3.61	297.78	0.0372	1794.70	35.15
LOG	4.71	227.73	0.0285	2335.28	33.97
POW	3.63	295.84	0.0370	1488.77	30.05
CLAMP	3.66	293.12	0.0366	1673.10	31.99
MATMUL	81.23	9.91	13.5354	1373.70	17.80
TRANSPOSE	2.38	225.77	0.0000	1059.03	24.71
MAT_VEC	0.87	307.53	0.1537	379.82	5.80
ADD_BIAS	1.93	277.69	0.0347	710.38	12.69
SCALE_MAT	1.99	270.11	0.0338	1107.48	17.39
ADD_SCALAR	3.53	303.96	0.0380	1708.55	19.71
RELU	4.19	256.38	0.0320	2446.01	28.16
SIGMOID	4.87	220.38	0.0826	2385.56	12.14
TANH	4.46	240.48	0.0902	2415.69	11.53
GELU	5.33	201.43	0.1259	3790.56	32.06
SOFTMAX	3.12	258.28	0.0646	1023.32	8.64
BCE_LOSS	24.06	44.64	0.0223	1405.21	12.10
MSE_VEC	24.40	44.01	0.0165	1410.52	10.48
MSE_MAT	12.11	44.33	0.0166	711.45	5.07
SUM_VEC	1.82	294.55	0.0736	701.77	17.40
SUM_COLS	0.89	300.85	0.0752	523.13	7.02
SUM_ROWS	7.13	37.65	0.0094	349.99	16.54
EMBED_VEC	15.30	421.33	0.0000	5286.45	84.81
EMBED_MAT	15.26	422.36	0.0000	4544.87	88.81
SLICE_COL	1.00	269.41	0.0000	962.99	14.19
SLICE_ROW	0.11	76.73	0.0000	1174.23	5.80
SLICE_3D	0.08	99.27	0.0000	370.54	3.04
CONCAT_VEC	5.39	199.13	0.0000	2857.50	30.98
STACK_MAT	0.57	235.83	0.0000	224.28	3.67
SCAT_HEADS	0.33	75.67	0.0000	36.79	1.72
PAD_2D	0.47	278.27	0.0000	194.99	0.97
LAYER_NORM	3.47	154.71	0.1547	727.34	5.75
DROPOUT	2.75	293.00	0.0244	1073.25	18.37
DOT_PROD	7.31	293.90	0.0367	2356.80	32.84
L2_NORM	5.66	284.70	0.0475	1797.15	14.86
EUC_DIST	11.90	270.64	0.0338	3751.22	34.63
COS_SIM	18.41	291.55	0.0437	3780.90	31.48
MAE_LOSS	48.99	65.75	0.0082	3310.34	29.07

Future Plans

1. Integration of complete pre-trained LLM pipelines.
2. Additional science and custom kernel mapping.