Nushell

Bash gets pipelines; Nushell gets pipelines and structure. A generated module wraps every operation so tensors come from and return to native Nushell values — no JSON wrangling.

Setup

If you installed NuTorch with Homebrew, there is nothing to set up. The formula ships a vendor-autoload stub, Nushell sources it at startup, and torch commands are simply there in every new session — no use line, no config.nu edit. (Mechanism: Nushell autoloads every .nu file in $nu.vendor-autoload-dirs, and brew-built Nushell scans $(brew --prefix)/share/nushell/vendor/autoload.)

The manual fallback — for from-source installs, or a Nushell that was not installed by Homebrew (cargo, MacPorts, nightly builds) and so may not scan the brew prefix:

# one-time: autoload in every session
mkdir ~/.config/nushell/autoload
'use "/opt/homebrew/share/nutorch/nutorch.nu" *'
  | save ~/.config/nushell/autoload/nutorch.nu

# or just this session
use /opt/homebrew/share/nutorch/nutorch.nu *

A current copy of nutorch.nu also ships in the repo root; regenerate it any time with:

torch nu-module | save -f nutorch.nu

Structured data in and out

let t = ([[1 2] [3 4]] | torch tensor)
$t | torch mm $t | torch value            # a native table
torch tensors | where bytes > 1_000_000 | get handle | each {|h| torch free $h }

Wrappers honor the dual input pattern — pipe the leftmost tensor in as $in, or pass handles as arguments and the CLI’s grammar fills the missing slots — and every one of the 185 ops is wrapped, plus the registry and daemon verbs. Non-finite values cross the boundary as REAL Nushell NaN/infinity floats; the JSON dialect ("NaN", "Infinity", "-Infinity") is handled for you in both directions.

Use ^torch when you explicitly want the external CLI’s raw text or JSON output instead of the structured Nushell wrapper result. The old nutorch <op> namespace remains as a compatibility alias for existing scripts.

Training, natively

The regression from neural networks, as Nushell:

use nutorch.nu *

torch manual_seed 42 | ignore
let x = ([[0.0] [1.0] [2.0] [3.0]] | torch tensor)
let y = ([[1.0] [3.0] [5.0] [7.0]] | torch tensor)
let model = (torch nn linear 1 1)
let opt = (torch nn sgd $model --lr 0.05)

for i in 1..200 {
  let loss = ($x | torch forward $model | torch mse_loss $y)
  $loss | torch backward
  $opt | torch step
  torch nn zero_grad $opt | ignore
}

The full script (with assertions) is scripts/train-regression.nu in the repo — it converges to the same losses as its zsh twin.

Plain JSON everywhere else

The structured verbs also serve any other tool via --json:

torch tensors --json
torch ops --json
torch daemon status --json
torch nn info $m --json