[Mlir-commits] [mlir] Added free-threading CPython mode support in MLIR Python bindings (PR #107103)

[Stella Laurenzo]

stellaraccident wrote:

> > Off topic, but one of these days, I'm going to have to create a stripped down, build only (probably nanobind based) fast version of the API. Not at the top of the list right now, but yeah... slow. On the torch side, we only use the non-ODS API so that we have more flexibility here in the future.
> 
> Yeah, at some point soon I'm either going to end up trying to send PRs to speed up the existing bindings (switch to nanobind?) or I'll migrate off them and do my own thing.

I haven't benchmarked it in detail, but there are quite a bit of "productivity features" in there. I don't think for fastest-path builder use, the general IR manipulation APIs are going to be super easy to upgrade -- and they do get use. If the whole thing was ported to nanobind, I'd feel better about contributing a build-only fast-path subset library that could live alongside. In the current state, that would go out of tree because if going to the trouble to write something fast, I'm not going to base it on pybind.

https://github.com/llvm/llvm-project/pull/107103