Home

simplegrad

A lightweight deep learning framework with automatic differentiation. Runs on NumPy (CPU) or CuPy (GPU/NVIDIA CUDA).

Installation

The package is not yet on PyPI — install from source:

# CPU only (NumPy backend)
pip install git+https://github.com/simplegrad/simplegrad.git

# GPU support (CuPy — requires CUDA 12.x)
pip install "simplegrad[gpu] @ git+https://github.com/simplegrad/simplegrad.git"

The SimpleBoard frontend is compiled from npm sources during installation. If Node.js is not available, skip the build with:

SIMPLEGRAD_NO_BUILD_WEB=1 pip install git+https://github.com/simplegrad/simplegrad.git

Optional extras

Extra	Installs	Use when
`gpu`	`cupy-cuda12x`	Running on an NVIDIA GPU (CUDA 12.x)
`bench`	`torch`	Running the benchmarks in `benchmarks/`
`dev`	`pytest`, `black`, `mypy`, `ipykernel`	Contributing / running tests
`docs`	`mkdocs`, `mkdocs-material`, `mkdocstrings`	Building the documentation locally

# Install multiple extras at once
pip install "simplegrad[gpu,dev] @ git+https://github.com/simplegrad/simplegrad.git"

Quick start

import simplegrad as sg

x = sg.Tensor([[1.0, 2.0], [3.0, 4.0]], label="x")
w = sg.Tensor([[0.5], [-0.5]], label="w")

y = sg.mean(x @ w)
y.backward()

print(x.grad)  # d(mean(x @ w)) / dx
print(w.grad)  # d(mean(x @ w)) / dw

See the Introduction for a full walkthrough of the framework architecture, a training example, and guides to lazy mode and experiment tracking.