MarioSieg/magnetron

(WIP) A small but powerful, homemade PyTorch from scratch.

62
/ 100
Established

Implements a native C core with CPUID-based multi-dispatch kernels optimized across Intel, AMD, and ARM microarchitectures, paired with a dynamic reverse-mode autograd engine and explicit view-based tensor system. Supports eager execution with zero-copy memory-mapped model loading via a custom `.mag` format, enabling practical inference of modern LLMs like Qwen3 in bfloat16 while remaining fully inspectable end-to-end. CUDA backend is in active development, with planned support for multi-GPU execution and a custom JIT compiler.

676 stars. Actively maintained with 40 commits in the last 30 days.

No Package No Dependents
Maintenance 23 / 25
Adoption 10 / 25
Maturity 16 / 25
Community 13 / 25

How are scores calculated?

Stars

676

Forks

33

Language

C++

License

Last pushed

Mar 11, 2026

Commits (30d)

40

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/ml-frameworks/MarioSieg/magnetron"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.