Luminal: Compiling fast GPU kernels in Rust

Blog Post
Hi everyone!
I posted about Luminal a few months ago, and since then we've made significant progress in the CUDA and Metal compilers.
For those who don't know, Luminal is a a deep learning library that uses composable compilers to achieve high performance. It's entirely written in Rust and orders of magnitude simpler than comparable ML frameworks.
Recently we've been working on speeding up performance on Nvidia and Apple GPUs, which led us to write compilers capable of generating kernels on the fly. These kernels are compiled once and ran many times, and since the whole compute graph is known at compile time, we can get very aggressive with the kernel optimizations.

We're adding tons more features (autograd landed a few weeks ago, so training is now possible!) so check out the blog post and the repo if you'd like to know more!