In AVX512 and Vec512 · Issue #56187 · pytorch/pytorch · GitHub we are considering dropping AVX support. This means to get vectorized CPU kernels, you must have a CPU recent enough to support AVX2, otherwise you will get unvectorized operations. If this would unduly affect you, please let us know on the issue.
Do we have a way to get some statistics about that via PyPi or conda by any chance?
Probably not but there was some steam statistics and it was 82.28% with AVX2 support (and its growing by like ~0.5% month to month) and 94.77% with AVX support. So i would think percentage should be lot higher in ML based builds
While it’s not an issue for researchers/data scientists/cloud services, it can be an issue when you need to use PyTorch inference within softwares meant for consumers.
For instance macOS 10.14 and 10.15 (still supported by Apple) can run on CPUs without AVX2 instructions. Only macOS 11.0 requires AVX2 CPUs.
That being said, it’s not dramatic if PyTorch then fallback to a non-vectorized path. It would be problematic if it couldn’t run at all without AVX2.
Yes, what would happen in those cases is it would just degrade to nonvectorized.