Installing from source

The Homebrew tap (see getting started) is the recommended install. Building from source is two scripts:

git clone https://github.com/nutorch/nutorch
cd nutorch
scripts/bootstrap.sh     # venv + torch 2.11.0 + release build (idempotent)
scripts/install.sh       # → ~/.nutorch (or pass a prefix); add ~/.nutorch/bin to PATH
torch --version

What the scripts do

• bootstrap.sh creates a repo-local Python venv just to obtain the PyTorch 2.11.0 wheel (the strict pairing for tch 0.24.0), symlinks .libtorch at the staged package, and runs the release build. It detects an existing valid venv and skips the download — safe to re-run.
• install.sh copies the binaries (torch, nutorchd, and a nutorch symlink — the two CLI names are the same tool), the four libtorch dylibs they need (libtorch, libtorch_cpu, libc10, libomp), and nutorch.nu into a prefix (default ~/.nutorch).

The installed binaries are relocatable: libtorch’s dylibs resolve through a baked relative rpath — no environment variables, no checkout needed at runtime, and the install keeps working if you delete the source tree.

Verifying

torch --version          # nutorch 0.1.0 (<git sha>)
torch daemon status      # device: mps

--version works on GPU-less machines (it short-circuits before the MPS gate); everything else requires Apple silicon — the daemon refuses to start without MPS.

Requirements

• Apple-silicon Mac (MPS)
• Rust toolchain (for the build)
• Python 3 (only for the bootstrap’s wheel download)