ml_transformer is a lightweight, modular, and extensible machine learning and neural network library written in Dart.

• Tensor primitives and basic mathematical ops
• Common neural network layers (Embedding, Linear, LayerNorm, Attention, TransformerBlock)
• Mini autograd and AdamW optimizer
• Quantization helpers (int8, int4, int2, fp16, bf16)
• Basic ONNX import/export bridge (JSON demo; protobuf ONNX planned)
• Simple dataset loader and batching API
• Model save/load utilities
• Training callbacks and tracking (EarlyStopping, History)
• Unit tests covering core components

git clone <repo-url>
cd ml_transformer
dart pub get

import 'lib/dart_transformers.dart';

void main() {
 final cfg = GPTConfig(
  vocabSize: 256,
  maxSeqLen: 128,
  dModel: 64,
  nHeads: 8,
  nLayers: 4,
 );
 final model = GPTModel(cfg);
 final logits = model.forward([1, 2, 3, 4]);
 print('Logits: ${logits.rows} x ${logits.cols}');
}

There is also an example that demonstrates a tiny training loop, quantization and the ONNX import stub:

dart run example/demo_train_quant_onnx.dart

Run unit tests with:

dart test

• The ONNX bridge is a JSON-based demo; full ONNX (protobuf) support requires additional dependencies and parsing.
• Some layers (RNN / LSTM / GRU) are currently basic examples and may need optimization for production.
• Performance for very large models is limited by pure-Dart execution; native acceleration (FFI/GPU) is planned.
• Quantization utilities are intended as helpers and need calibration/validation for production models.

1. Full ONNX protobuf import/export and weight mapping
2. Complete and optimize RNN/LSTM/GRU implementations with tests
3. Optional native acceleration (FFI / GPU backends)
4. More quantization modes and calibration utilities
5. High-level Model API (fit/predict/evaluate) with checkpoints and schedulers

Contributions are welcome. Please open issues or pull requests, and include tests for new functionality.

MIT