dan_the_3rd

Hi, This is a known issue, and it can be mitigated by limiting the number of parallel tasks for build. Just set `MAX_JOBS=4` before building for instance

We're working to make the compilation way faster in the future, but this might take some weeks/months and require a new PT release. In the meantime, you should be able...

Hi, We plan to add prebuilt wheels for py312 when PyTorch 2.4.0 is released (expected [end of july](https://dev-discuss.pytorch.org/t/pytorch-release-2-4-0-call-for-features/2051))

Yes we plan to add 3.12 for both :)

The PT 2.4.0 is actually a requirement to have smaller build sizes, which we need to be able to host more versions (eg py312 etc...), so I don't expect us...

We now have wheels for py312. Closing

Hi, Thanks for opening this issue. I'm not sure I get what you want tho. Do you want us to move the functionality from xFormers to PyTorch? Or do you...

Oh I see. It depends which components you want to move. I assume you have calls to `memory_efficient_attention` that you want to replace with PyTorch's SDPA. If you have specific...

I don't know why this happens only with Ninja, but would not recommend to modify pytorch internals..

Hi @conceptofmind There is an overhead to pay to sparsify a GEMM input, but then the sparse GEMM is faster. So the best scenario is when the operand that you...