GPU kernels in Python. Compiled for Apple silicon.
Write GPU kernels in Python; meTile compiles them to Metal. Declare tensors, then choose explicit layouts and schedules where needed. MLX integration is optional.
Read the docs · Get started · First kernel · Benchmarks
From a checkout, on an Apple silicon Mac with Python 3.10+:
python -m pip install -e .For ready-made kernels, install the separate kernels project with
python -m pip install -e ./kernels and import metile_kernels. The compiler
does not require it.
meTile is experimental. The docs include measured results, limitations and commands to reproduce them; performance depends on your workload.
Ideas and inspiration: CuTe, Triton, Horace He's min-cut discussion, and Lap Chi Lau's CS 341 notes. Full references, acknowledgments, and BibTeX.