Operations

NuTorch’s operation surface is a single declarative table shared by the daemon and the CLI: 185 ops spanning creation, pointwise math, linear algebra, reductions, shape manipulation, indexing, losses, and more.

PyTorch fidelity

Command names, argument order, defaults, and semantics match PyTorch wherever possible. If you know torch.add(a, b, alpha=2), you know:

torch add $a $b --alpha 2

torch add $a $b --alpha 2

Broadcasting follows PyTorch’s rules. Non-broadcastable shapes error with both shapes named — validation happens in Rust before any GPU call, so error messages talk about your tensors, not C++ internals.

Discoverability

torch ops                # every op: name, category, one-line summary
torch ops --json         # the same as JSON (name, category, summary)
torch mean --help        # usage, parameters, defaults for any op

torch ops                             # every op: a native table
torch ops | where category == "loss"  # filter natively — no JSON needed
torch mean --help                       # usage, parameters, defaults for any op

Every op supports the dual input pattern — pipe the leftmost tensor in or pass it as an argument (see tensors).

Reference

The full per-op reference is generated from the same table the binaries use, so it cannot drift: creation, pointwise, comparison, reduction, linear algebra, shape, loss, autograd, and utility — or jump in from the Reference section in the sidebar.