Dan Fu

The 3x3 wouldn't be a great use case right now - we specialize for really long filters (size on the order of the input image).

That's correct, we don't have a Conv2D implemented right now.

Hello! Hopefully it helps to work through the backprop calculation by hand. First, I'll note that `iFFT(FFT(x)) = x`, so if `out = iFFT(FFT(x))`, then the gradient `dx = dout...

Great question! If you look at section 5.1, we use the Monarchs to implement long convolutions in conjunction with gating for a lot of the backbones. (also see this [image](https://hazyresearch.stanford.edu/static/posts/2023-07-25-m2-bert/m2-arch.png)...

Thanks for the detailed bug report! I believe the issues on non-A100 are related to #6. We’ll have to take a closer look at the others. It may be a...

Ah this is currently not supported - see the requirements: https://github.com/HazyResearch/flash-fft-conv?tab=readme-ov-file#input-requirements-and-notes We’ll add a more obvious error message!

Great question! Every convolution is an SSM so that’s what we mean by SSM model. The dimension mixer is orthogonal. On Wed, May 29, 2024 at 12:41 AM 41924076 ***@***.***>...

This seems like it could be a problem with windows, I haven’t tried compiling kernels on windows before. On Tue, Sep 24, 2024 at 6:10 AM Chao He ***@***.***> wrote:...

Thanks so much for this PR, it looks really helpful. Two requests: 1. Do you know if there's a way to test the release build before merging? 2. Can you...

I don't think I can squash it, since it's a PR from your branch. I'm not an expert in github's online interfaces though :) For the tests - I mean...