tch-rs

Why we included this project

Teams working in Rust who want PyTorch's capabilities without leaving the language get a thin, faithful wrapper around libtorch's C++ API. You get tensors, automatic differentiation, optimizers like SGD and Adam, and the familiar nn module for composing layers, all callable from ordinary Rust code. Because the bindings mirror the upstream C++ surface closely, PyTorch's own docs and examples carry over and you stay aligned with upstream behavior. Installation adapts to your setup: link a system libtorch, point the build at a Python torch install, or let the build script pull a prebuilt CPU or CUDA binary. That flexibility suits both production services needing fast tensor math and small teams prototyping custom architectures entirely in Rust.