Oops, didn't know, thanks! I just added a section on how to compile and try it without a system installation to avoid any name clash :)

Hi everyone,

I’ve recently been digging into the PyTorch FSDP implementation. It’s powerful and highly optimized, which naturally means the codebase is extensive and isn't always straightforward to navigate. In the process, I decided to write a minimal implementation based on my findings, mainly to emphasize and show the different states and pre/post forward/backward of FSDP, all in a single place!

Hope this helps others!

You made me remember about this project: https://hackernewsbooks.com/ Despite it is not related to clustering I think it is a smart idea that offers a complete different view of HN