public static void RunGraph()
{
    using var host1 = HostArray.Allocate<float>(4096);
    using var host2 = HostArray.Allocate<float>(4096);
    using var dev = DeviceArray.Allocate<float>(4096);

    // Initialize data
    for (var i = 0; i < 4096; ++i)
        host1[i] = i;

    // Capture operations into graph
    using var stream = new CudaStream();
    stream.BeginCapture();
    host1.CopyTo(dev, stream);
    dev.CopyTo(host2, stream);
    
    using var graph = stream.EndCapture();
    using var graphInstance = graph.CreateInstance();
    
    // Execute graph on new data
    for (var i = 0; i < 4096; ++i)
        host1[i] = i + 10;

    graphInstance.Launch(stream);
    stream.Synchronize();
}

Shape shape = [3, 4, 5];
var reduced = shape.Reduce(axis: 1);
var appended = shape.Append(extent: 1, stride: 0);
var prepended = shape.Prepend(extent: 1, stride: 0);
var transposed = shape.Transpose(^1, ^2);

var rowMajor = Shape.Create([3, 5], Layout.Right);
var colMajor = Shape.Create([3, 5], Layout.Left);

public static void RunToyModel()
{
    const int batchSize = 2048;
    const int inputSize = 28 * 28;
    const int hiddenSize = 512;
    const int outputSize = 10;

    using Sequential model =
    [
        new LinearLayer(inputSize, hiddenSize),
        new LinearLayer(hiddenSize, outputSize),
    ];
    
    // Allocate outputs and prepare kernels
    var inputs = DeviceTensor.Allocate<float>([batchSize, inputSize]);
    var outputs = model.Compose(inputs);
    var graph = model.Compile();

    // Execute graph
    using var stream = new CudaStream();
    using var start = new CudaEvent();
    using var stop = new CudaEvent();
    
    start.Record(stream);
    
    for (var epoch = 0; epoch < 10000; epoch++)
        graph.Launch(stream);
    
    stop.Record(stream);
    stop.Synchronize();
    
    // Get results
    var elapsed = CudaEvent.Elapsed(start, stop);
    var results = HostTensor.Allocate<float>(outputs.Shape);
    outputs.CopyTo(results, stream);
}