For research you use Python or whatever. For production you compile your models to whatever their proprietary equivalent to TorchScript and ONNX is, which is surely executed by hyperoptimized C++/Rust/Cuda code.