Tobias van der Werff

I have not tested tokens/s generation @yiliu30, but I can test this if you want.

Thanks for bringing this up @yiliu30 -- I tested this and it looks like the custom Hadamard transform kernel indeed breaks torch.compile. I'll investigate this and get back to you.

@yiliu30 FYI I fixed the issue with torch.compile -- you can see the benchmark results [here](https://github.com/pytorch/ao/pull/1069#issuecomment-2413275597).