I told you all about FlashAttention 2 last July. Well, it's just been announced that FlashAttention 2 is has been integrated into PyTorch, the leading tools for producing deep learning neural networks.
So you see, when I tell you about new techniques, they sometimes do get incorporated into the industry toolchain and transform the way AI systems are built.
Also mentioned is that they are deprecating macOS x86 support, and will soon drop all support for macOS x64. I was surprised by that as it didn't seem that long since Apple switched to M1 processors? I looked it up. November 2020. So, a little over 3 years. So, I guess AI people in the industry upgrade to new hardware quickly.
This update of PyTorch also has something called AOTInductor, which looks like it's designed to let you develop models in Python and export them to C++ (and other "non-Python environments).
