Rust backend interest?

[MilesCranmer]

We now have a fairly complete port of SymbolicRegression.jl in Rust: https://github.com/astroautomata/symbolic_regression.rs. I was thinking it could be nice to expose this to PySR. It is much less flexible due to being statically compiled rather than JIT like Julia, but it also means it should offer a faster startup time and smaller install size (we wouldn't need to ship all of Julia for example).

Perhaps a PySR backend could be an pip "extra" such as pysr[rust] or maybe an API config option like PySRRegressor(backend="rust").

Then again, I am not sure the appetite for this, because it would lose some of the key features that make PySR nice, like the high level of configurability like custom operators, custom losses, custom templates. But perhaps for "vanilla" symbolic regression it might be nice.

So far I've also been slightly more successful in building a fast native GPU backend in the Rust port, which could be interesting. But again, this would be nowhere near as flexible as if we got it working from Julia. So I'm not too sure.

Another aside: the Rust backend can compile to WebAssembly so you can literally run symbolic_regression.rs in the browser: https://astroautomata.com/symbolic_regression.rs/ (even works on your phone!).

[gustavorubim]

i would like to contribute to this. a julia free version would be nice and more flexible. i can work on a PR for this.