Quickstart

Write a small model, check it, and run it in PyTorch. To try it without installing anything, open the quickstart notebook on Colab: it also runs and serves a Hugging Face model.

1. Install

uv venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
uv pip install "linnet-lang[torch]"

This installs the linnet compiler too. Other options are in Installation.

2. Create a package

linnet init hello-model
cd hello-model

Replace the generated src/lib.linnet with:

module hello_model

use std.nn.activations::{relu}
use std.nn.linear::{Linear}

pub block Mlp<In: Dim, Hidden: Dim, Out: Dim, T: Float = f32> {
    sub up: Linear<In, Hidden, T>
    sub down: Linear<Hidden, Out, T>

    pub entry forward<B: Dim>(x: Tensor[B, In; T]) -> Tensor[B, Out; T] {
        return down.forward(relu(up.forward(x)))
    }
}

A block owns parameters and sub-blocks, and an entry is what a backend calls. The compiler checks every shape in terms of In, Hidden, Out and B.

3. Check it

linnet check .
linnet fmt --check .
linnet inspect --parameters src/lib.linnet

hello_model::Mlp<In, Hidden, Out, T>
  param up.weight: Tensor[Hidden, In; T]
  param up.bias: Tensor[Hidden; T]?
  param down.weight: Tensor[Out, Hidden; T]
  param down.bias: Tensor[Out; T]?

These are the tensors a checkpoint must provide; ? marks an optional one. Change the result type to Tensor[B, In; T] and linnet check . reports the mismatch.

4. Run it in PyTorch

import torch
from linnet.torch import load

model = load("hello-model/src/lib.linnet", generics={"In": 4, "Hidden": 8, "Out": 2})
print(model(torch.randn(3, 4)).shape)   # torch.Size([3, 2])

The model now runs as a PyTorch module. Pass weights="weights/" to load SafeTensors named by parameter path; see PyTorch.

Next

• Language tour, or the whole language in one page
• Command line
• Editor setup