torch_dart 0.0.1

Dart bindings for PyTorch with a replaceable backend.

torch_dart

A small Dart API for PyTorch. The first backend uses py_embed and the official Python torch package. Public APIs do not expose Python objects, so a LibTorch or other native backend can be added later.

Setup

Create and activate a Python environment, then install PyTorch:

uv venv --seed
pip install torch

To select a particular Python executable, call Python.configure from package:py_embed/py_embed.dart before the first torch operation.

Usage

import 'package:py_embed/py_embed.dart';
import 'package:torch_dart/torch.dart';

void main() {
  final a = tensor([
    [1.0, 2.0],
    [3.0, 4.0],
  ]);
  final b = ones([2, 2]);
  final result = (a + b).matmul(b);

  try {
    print(result.shape);
    print(result.toList());
  } finally {
    result.dispose();
    b.dispose();
    a.dispose();
    Python.shutdown();
  }
}

The MVP includes tensor creation, shape/dtype/device metadata, element-wise arithmetic, matrix multiplication, reshape, transpose, reductions, ReLU, conversion to Dart lists, and deterministic random seeds.

Inference mode and TorchScript modules use a namespaced API:

final module = torch.jit.load('model.pt', mapLocation: 'cpu')..eval();
final input = randn([1, 3, 224, 224]);
final output = torch.inferenceMode(() => module(input));

try {
  print(output.shape);
} finally {
  output.dispose();
  input.dispose();
  module.dispose();
  Python.shutdown();
}

Each operation creates a new tensor. Dispose tensors explicitly when they are no longer needed. Call Python.shutdown() once at application shutdown, after disposing every object from every Python-backed package. The embedded Python runtime cannot be restarted in the same process afterward.

Architecture

Tensor / top-level factories
            |
       TorchBackend
            |
    PythonTorchBackend