dan_the_3rd

Closing this as it's resolved now. @leeruibin this is a different / unrelated issue. Can you open a new one with the entire stacktrace/log of the error? `num_heads=1` should be...

> Just wonder the fix in which release or dev version? Woops I forgot to circle back here. It has been fixed in https://github.com/facebookresearch/xformers/commit/3ea73074f41617857fe001f1e7d46dee2e03517a

Hi, Can you specify your configuration? Which GPU etc...

Hi @apolinario That's a very valid point, and we know it's been a pain point for the users ... Some folks in the community have been willing to contribute to...

Hi, we don't support MacOS anymore, as there is no recent CUDA release available on Mac.

Hi, Windows is supported, but you will have to build xFormers yourself as we don't provide windows binaries at the moment

We provide binaries for cuda 11.3 and 11.6 (but **only** for linux and specific versions of pytorch). If you build from source, we support any version of cuda >= 11.3...

The "filename too long" is a limitation in Windows, that [can be addressed](https://stackoverflow.com/questions/22575662/filename-too-long-in-git-for-windows). You can also download windows builds from there (have to be installed manually): https://github.com/facebookresearch/xformers/actions/runs/3616539007

I think the point is to have a fall-back for CPU (and unsupported devices) that is memory-efficient - even if it's much slower than pytorch

This introduces some performance regression in the fw, and in the bw. For the FW, this should be addressed by using the triton implem: https://github.com/facebookresearch/xformers/pull/479 For the BW, this should...