Elijah Rippeth

Yeah, I think the heavy lift is supporting the various CUDA toolkits which I don't have experience supporting in conda recipes. I'm happy to be added as a maintainer or...

Tracking triton condafication [here](https://github.com/conda-forge/staged-recipes/pull/17310)

Since you already have xformers as a soft dependency, you should be able to pull [it](https://github.com/facebookresearch/xformers/blob/main/xformers/components/positional_embedding/rotary.py) in directly from there if it's installed (similar to flash attention).

This is a massively helpful PR. I routinely have to re-align my source with my generated output which can be very messy. This is a very easy PR to accept.

The second commit gives us an additional 7% improvement in speed by realizing that ... we only need one sub-process.

Added it in the "header" and fixed the merge conflict. I think this is GTG, @mjpost

I guess for what it's worth: I wanted to submit a PR for this change, but can't quite get the code in `main` to build locally on my WSL rig...

@Vulum that's not true. My reproducer explicitly uses 3.10

Great. There are some caveats about parity between features from legacy torchtext APIs and new ones, but they likely won't be of consequence for this repo. The biggest change is...

:+1: Please consider finishing this!