Tri Dao

This is great! Our forward pass implementations are very similar. The backward looks different (https://github.com/HazyResearch/flash-attention) but let's figure that out during our chat!

I agree that this could be quite misleading for the default to diverge from sklearn (which I assume is most ppl are used to).

Thanks for trying out the code. As mentioned in the README, you can see `learning_transforms/run_exp.sh` for examples of how to run the experiments. For example, to learn the fast algorithm...

To get the dense matrix from a butterfly you can just multiply the butterfly with an identity matrix: ``` import torch import torch.nn.functional as F from torch_butterfly import Butterfly b...

Are you compiling the code from the latest commit on master? Which OS and gcc version are you using? Was Pytorch installed with conda or pip?

One possibility is incorrect Pytorch version. Old Pytorch versions don't have `TORCH_CHECK` and that might cause the error. When I compile with Pytorch 1.6 it defaults to `-std=c++14` (https://github.com/pytorch/pytorch/blob/v1.6.0/torch/utils/cpp_extension.py#L360) but...

I've pushed a commit to change `TORCH_ERROR` to support non-CUDA compilation. Can you try that out?

> Yes. I am able to successfully compile this. I see some warnings here and there but no errors. That's great. Thanks for trying that out.

No, it was just an error on my part. Thanks for reporting it.

Hi, thanks for trying our code. Which OS and gcc version are you using?