magnetron 0.1.5

High-performance, lightweight deep-learning library with a PyTorch like API and GPU support.

magnetron

A compact, PyTorch-style machine learning framework written in C and modern Python.
Designed for speed, clarity, and portability - from desktop to embedded.

Documentation »

Qwen3 Example · Report Bug · Request Feature

---

📖 About

Magnetron is a lightweight, research-grade machine learning framework that mirrors the usability of PyTorch - but built entirely from scratch.
Its C99 core, wrapped in a modern Python API, provides dynamic computation graphs, automatic differentiation, and high-performance operators with zero external dependencies.

Originally designed for constrained or experimental environments, Magnetron scales from small embedded systems to full desktop inference and training.
A CUDA backend and mixed-precision support are currently in development.

---

⚡ Highlights

• PyTorch-like API
  Familiar syntax for building and training models - easy to pick up, minimal to extend.

• Dynamic autograd engine
  Eager execution with full gradient tracking on computation graphs.

• Optimized C99 backend
  Custom tensor engine with SIMD acceleration (SSE, AVX2, AVX-512, NEON) and multithreaded execution.

• Minimal dependencies
  No third-party math libraries; only CFFI is required for the Python interface.

• Lightweight neural modules
  Includes Linear, Sequential, ReLU, Tanh, Sigmoid, LayerNorm, Embedding, and more.

• Rich data types with many operators
  Supports bfloat16, float16, float32, int8, uint8, int16, uint16, int32, uint32, int64, uint64, and boolean.

• Custom serialization format
  Fast, portable model saving and loading through Magnetron’s own binary tensor format.

• Clean diagnostics
  Readable validation and error messages for faster debugging and experimentation.

---

🚀 Example Models

Example	Description
Qwen3 Inference	Transformer-based text generation using pretrained Qwen3 weights, loaded from custom .mag file.
GPT-2 Inference	Transformer-based text generation using pretrained GPT-2 weights.
Autoencoder	Image reconstruction using a small dense encoder–decoder network.
Linear Regression	Fits a linear model to noisy synthetic data.
XOR	Trains a small neural network to learn the XOR logical function.

---

📦 Installation

Make sure you are inside a Python virtual environment before installing.

With uv

uv pip install magnetron

With pip

pip install magnetron

🤝 Contributing

Contributions are welcome!
Please open issues for ideas, or submit pull requests for new features.
PRs that only fix typos or minor formatting will not be accepted.

📜 License

(c) 2026 Mario Sieg - [e-mail]
Distributed under the Apache 2 License. See LICENSE for more information.

🧩 Similar Projects

• GGML
• TINYGRAD
• MICROGRAD