PyTorch → .NET

Verdict: use what exists, nothing to write. TorchSharp is libtorch (the same C++ engine as PyTorch), with autograd, nn modules and CUDA.

PyTorch need	Recommended .NET
Tensors, autograd, training, GPU	TorchSharp (TorchSharp)
Inference of a pretrained model, without training	ONNX Runtime (Microsoft.ML.OnnxRuntime)
Keras/TF models	TensorFlow.NET

dotnet add package TorchSharp
dotnet add package TorchSharp-cpu     # or libtorch-cuda-* for GPU

using static TorchSharp.torch;

var x = randn(3, 4);
var w = randn(4, 2, requires_grad: true);
var y = x.matmul(w).relu().sum();
y.backward();                 // autograd, like PyTorch

Pitfalls

• Very close API but not identical: torch.xxx → TorchSharp.torch.xxx, PascalCase naming for nn modules. Porting a script is mechanical but not automatic.
• Memory management. Native tensors are freed via using/Dispose or a DisposeScope — there is no GC for native memory. A classic leak pitfall.
• Inference only → ONNX. To serve a Python-trained model, export to ONNX and load with ONNX Runtime: lighter than TorchSharp. That is the path taken by Lodestar.Embeddings.

Guide to be expanded as real needs arise.